src/vol/ihandle.c

   1 /*
   2  * Copyright 2000, International Business Machines Corporation and others.
   3  * All Rights Reserved.
   4  *
   5  * This software has been released under the terms of the IBM Public
   6  * License.  For details, see the LICENSE file in the top-level source
   7  * directory or online at http://www.openafs.org/dl/license10.html
   8  */
   9
  10 /*  ihandle.c   - file descriptor cacheing for Inode handles.           */
  11 /*                                                                      */
  12 /************************************************************************/
  13
  14 #include <afsconfig.h>
  15 #include <afs/param.h>
  16
  17
  18 #include <stdio.h>
  19 #include <sys/types.h>
  20 #include <errno.h>
  21 #include <string.h>
  22 #ifdef AFS_NT40_ENV
  23 #include <fcntl.h>
  24 #else
  25 #include <sys/file.h>
  26 #include <unistd.h>
  27 #include <sys/stat.h>
  28 #if defined(AFS_SUN5_ENV) || defined(AFS_NBSD_ENV)
  29 #include <sys/fcntl.h>
  30 #include <sys/resource.h>
  31 #endif
  32 #endif
  33
  34 #include <rx/xdr.h>
  35 #include <afs/afsint.h>
  36 #include <errno.h>
  37 #include <afs/afssyscalls.h>
  38 #include "nfs.h"
  39 #include "ihandle.h"
  40 #include "viceinode.h"
  41 #include "afs/afs_assert.h"
  42 #include <limits.h>
  43
  44 #ifndef AFS_NT40_ENV
  45 #ifdef O_LARGEFILE
  46 #define afs_stat        stat64
  47 #define afs_fstat       fstat64
  48 #else /* !O_LARGEFILE */
  49 #define afs_stat        stat
  50 #define afs_fstat       fstat
  51 #endif /* !O_LARGEFILE */
  52 #endif /* AFS_NT40_ENV */
  53
  54 #ifdef AFS_PTHREAD_ENV
  55 pthread_once_t ih_glock_once = PTHREAD_ONCE_INIT;
  56 pthread_mutex_t ih_glock_mutex;
  57 #endif /* AFS_PTHREAD_ENV */
  58
  59 /* Linked list of available inode handles */
  60 IHandle_t *ihAvailHead;
  61 IHandle_t *ihAvailTail;
  62
  63 /* Linked list of available file descriptor handles */
  64 FdHandle_t *fdAvailHead;
  65 FdHandle_t *fdAvailTail;
  66
  67 /* Linked list of available stream descriptor handles */
  68 StreamHandle_t *streamAvailHead;
  69 StreamHandle_t *streamAvailTail;
  70
  71 /* LRU list for file descriptor handles */
  72 FdHandle_t *fdLruHead;
  73 FdHandle_t *fdLruTail;
  74
  75 int ih_Inited = 0;
  76 int ih_PkgDefaultsSet = 0;
  77
  78 /* Most of the servers use fopen/fdopen. Since the FILE structure
  79  * only has eight bits for the file descriptor, the cache size
  80  * has to be less than 256. The cache can be made larger as long
  81  * as you are sure you don't need fopen/fdopen. */
  82
  83 /* As noted in ihandle.h, the fileno member of FILE on most platforms
  84  * in 2008 is a 16- or 32-bit signed int. -Matt
  85  */
  86 int fdMaxCacheSize = 0;
  87 int fdCacheSize = 0;
  88
  89 /* Number of in use file descriptors */
  90 int fdInUseCount = 0;
  91
  92 /* Hash table for inode handles */
  93 IHashBucket_t ihashTable[I_HANDLE_HASH_SIZE];
  94
  95 void *ih_sync_thread(void *);
  96
  97 /* start-time configurable I/O limits */
  98 ih_init_params vol_io_params;
  99
 100 void ih_PkgDefaults(void)
 101 {
 102     /* once */
 103     ih_PkgDefaultsSet = 1;
 104
 105     /* default to well-known values */
 106     vol_io_params.fd_handle_setaside = FD_HANDLE_SETASIDE;
 107
 108     /* initial fd cachesize.  the only one that will be used if
 109      * the application does not call ih_UseLargeCache().  set this
 110      * to a value representable in fileno member of the system's
 111      * FILE structure (or equivalent). */
 112     vol_io_params.fd_initial_cachesize = FD_DEFAULT_CACHESIZE;
 113
 114     /* fd cache size that will be used if/when ih_UseLargeCache()
 115      * is called */
 116     vol_io_params.fd_max_cachesize = FD_MAX_CACHESIZE;
 117
 118     vol_io_params.sync_behavior = IH_SYNC_ONCLOSE;
 119 }
 120
 121 int
 122 ih_SetSyncBehavior(const char *behavior)
 123 {
 124     int val;
 125
 126     if (strcmp(behavior, "always") == 0) {
 127         val = IH_SYNC_ALWAYS;
 128
 129     } else if (strcmp(behavior, "delayed") == 0) {
 130         val = IH_SYNC_DELAYED;
 131
 132     } else if (strcmp(behavior, "onclose") == 0) {
 133         val = IH_SYNC_ONCLOSE;
 134
 135     } else if (strcmp(behavior, "never") == 0) {
 136         val = IH_SYNC_NEVER;
 137
 138     } else {
 139         /* invalid behavior name */
 140         return -1;
 141     }
 142
 143     vol_io_params.sync_behavior = val;
 144     return 0;
 145 }
 146
 147 #ifdef AFS_PTHREAD_ENV
 148 /* Initialize the global ihandle mutex */
 149 void
 150 ih_glock_init(void)
 151 {
 152     MUTEX_INIT(&ih_glock_mutex, "ih glock", MUTEX_DEFAULT, 0);
 153 }
 154 #endif /* AFS_PTHREAD_ENV */
 155
 156 /* Initialize the file descriptor cache */
 157 void
 158 ih_Initialize(void)
 159 {
 160     int i;
 161     osi_Assert(!ih_Inited);
 162     ih_Inited = 1;
 163     DLL_INIT_LIST(ihAvailHead, ihAvailTail);
 164     DLL_INIT_LIST(fdAvailHead, fdAvailTail);
 165     DLL_INIT_LIST(fdLruHead, fdLruTail);
 166     for (i = 0; i < I_HANDLE_HASH_SIZE; i++) {
 167         DLL_INIT_LIST(ihashTable[i].ihash_head, ihashTable[i].ihash_tail);
 168     }
 169 #if defined(AFS_NT40_ENV)
 170     fdMaxCacheSize = vol_io_params.fd_max_cachesize;
 171 #elif defined(AFS_SUN5_ENV) || defined(AFS_NBSD_ENV)
 172     {
 173         struct rlimit rlim;
 174         osi_Assert(getrlimit(RLIMIT_NOFILE, &rlim) == 0);
 175         rlim.rlim_cur = rlim.rlim_max;
 176         osi_Assert(setrlimit(RLIMIT_NOFILE, &rlim) == 0);
 177         fdMaxCacheSize = rlim.rlim_cur - vol_io_params.fd_handle_setaside;
 178 #ifdef AFS_NBSD_ENV
 179         /* XXX this is to avoid using up all system fd netbsd is
 180          * somewhat broken and have set maximum fd for a root process
 181          * to the same as system fd that is avaible, so if the
 182          * fileserver uses all up process fds, all system fd will be
 183          * used up too !
 184          *
 185          * Check for this better
 186          */
 187         fdMaxCacheSize /= 4;
 188 #endif
 189         fdMaxCacheSize = MIN(fdMaxCacheSize, vol_io_params.fd_max_cachesize);
 190         osi_Assert(fdMaxCacheSize > 0);
 191     }
 192 #elif defined(AFS_HPUX_ENV)
 193     /* Avoid problems with "UFSOpen: igetinode failed" panics on HPUX 11.0 */
 194     fdMaxCacheSize = 0;
 195 #else
 196     {
 197         long fdMax = MAX(sysconf(_SC_OPEN_MAX) - vol_io_params.fd_handle_setaside,
 198                                          0);
 199         fdMaxCacheSize = (int)MIN(fdMax, vol_io_params.fd_max_cachesize);
 200     }
 201 #endif
 202     fdCacheSize = MIN(fdMaxCacheSize, vol_io_params.fd_initial_cachesize);
 203
 204     if (vol_io_params.sync_behavior == IH_SYNC_DELAYED) {
 205 #ifdef AFS_PTHREAD_ENV
 206         pthread_t syncer;
 207         pthread_attr_t tattr;
 208
 209         pthread_attr_init(&tattr);
 210         pthread_attr_setdetachstate(&tattr, PTHREAD_CREATE_DETACHED);
 211
 212         pthread_create(&syncer, &tattr, ih_sync_thread, NULL);
 213 #else /* AFS_PTHREAD_ENV */
 214         PROCESS syncer;
 215         LWP_CreateProcess(ih_sync_thread, 16*1024, LWP_MAX_PRIORITY - 2,
 216             NULL, "ih_syncer", &syncer);
 217 #endif /* AFS_PTHREAD_ENV */
 218     }
 219
 220 }
 221
 222 /* Make the file descriptor cache as big as possible. Don't this call
 223  * if the program uses fopen or fdopen, if fd_max_cachesize cannot be
 224  * represented in the fileno member of the system FILE structure (or
 225  * equivalent).
 226  */
 227 void
 228 ih_UseLargeCache(void)
 229 {
 230     IH_LOCK;
 231
 232     if (!ih_PkgDefaultsSet) {
 233         ih_PkgDefaults();
 234     }
 235
 236     if (!ih_Inited) {
 237         ih_Initialize();
 238     }
 239
 240     fdCacheSize = fdMaxCacheSize;
 241
 242     IH_UNLOCK;
 243 }
 244
 245 /* Allocate a chunk of inode handles */
 246 void
 247 iHandleAllocateChunk(void)
 248 {
 249     int i;
 250     IHandle_t *ihP;
 251
 252     osi_Assert(ihAvailHead == NULL);
 253     ihP = (IHandle_t *) malloc(I_HANDLE_MALLOCSIZE * sizeof(IHandle_t));
 254     osi_Assert(ihP != NULL);
 255     for (i = 0; i < I_HANDLE_MALLOCSIZE; i++) {
 256         ihP[i].ih_refcnt = 0;
 257         DLL_INSERT_TAIL(&ihP[i], ihAvailHead, ihAvailTail, ih_next, ih_prev);
 258     }
 259 }
 260
 261 /* Initialize an inode handle */
 262 IHandle_t *
 263 ih_init(int dev, int vid, Inode ino)
 264 {
 265     int ihash = IH_HASH(dev, vid, ino);
 266     IHandle_t *ihP;
 267
 268     if (!ih_PkgDefaultsSet) {
 269         ih_PkgDefaults();
 270     }
 271
 272     IH_LOCK;
 273     if (!ih_Inited) {
 274         ih_Initialize();
 275     }
 276
 277     /* Do we already have a handle for this Inode? */
 278     for (ihP = ihashTable[ihash].ihash_head; ihP; ihP = ihP->ih_next) {
 279         if (ihP->ih_ino == ino && ihP->ih_vid == vid && ihP->ih_dev == dev) {
 280             ihP->ih_refcnt++;
 281             IH_UNLOCK;
 282             return ihP;
 283         }
 284     }
 285
 286     /* Allocate and initialize a new Inode handle */
 287     if (ihAvailHead == NULL) {
 288         iHandleAllocateChunk();
 289     }
 290     ihP = ihAvailHead;
 291     osi_Assert(ihP->ih_refcnt == 0);
 292     DLL_DELETE(ihP, ihAvailHead, ihAvailTail, ih_next, ih_prev);
 293     ihP->ih_dev = dev;
 294     ihP->ih_vid = vid;
 295     ihP->ih_ino = ino;
 296     ihP->ih_flags = 0;
 297     ihP->ih_synced = 0;
 298     ihP->ih_refcnt = 1;
 299     DLL_INIT_LIST(ihP->ih_fdhead, ihP->ih_fdtail);
 300     DLL_INSERT_TAIL(ihP, ihashTable[ihash].ihash_head,
 301                     ihashTable[ihash].ihash_tail, ih_next, ih_prev);
 302     IH_UNLOCK;
 303     return ihP;
 304 }
 305
 306 /* Copy an inode handle */
 307 IHandle_t *
 308 ih_copy(IHandle_t * ihP)
 309 {
 310     IH_LOCK;
 311     osi_Assert(ih_Inited);
 312     osi_Assert(ihP->ih_refcnt > 0);
 313     ihP->ih_refcnt++;
 314     IH_UNLOCK;
 315     return ihP;
 316 }
 317
 318 /* Allocate a chunk of file descriptor handles */
 319 void
 320 fdHandleAllocateChunk(void)
 321 {
 322     int i;
 323     FdHandle_t *fdP;
 324
 325     osi_Assert(fdAvailHead == NULL);
 326     fdP = (FdHandle_t *) malloc(FD_HANDLE_MALLOCSIZE * sizeof(FdHandle_t));
 327     osi_Assert(fdP != NULL);
 328     for (i = 0; i < FD_HANDLE_MALLOCSIZE; i++) {
 329         fdP[i].fd_status = FD_HANDLE_AVAIL;
 330         fdP[i].fd_refcnt = 0;
 331         fdP[i].fd_ih = NULL;
 332         fdP[i].fd_fd = INVALID_FD;
 333         fdP[i].fd_ihnext = NULL;
 334         fdP[i].fd_ihprev = NULL;
 335         DLL_INSERT_TAIL(&fdP[i], fdAvailHead, fdAvailTail, fd_next, fd_prev);
 336     }
 337 }
 338
 339 /* Allocate a chunk of stream handles */
 340 void
 341 streamHandleAllocateChunk(void)
 342 {
 343     int i;
 344     StreamHandle_t *streamP;
 345
 346     osi_Assert(streamAvailHead == NULL);
 347     streamP = (StreamHandle_t *)
 348         malloc(STREAM_HANDLE_MALLOCSIZE * sizeof(StreamHandle_t));
 349     osi_Assert(streamP != NULL);
 350     for (i = 0; i < STREAM_HANDLE_MALLOCSIZE; i++) {
 351         streamP[i].str_fd = INVALID_FD;
 352         DLL_INSERT_TAIL(&streamP[i], streamAvailHead, streamAvailTail,
 353                         str_next, str_prev);
 354     }
 355 }
 356
 357 /*
 358  * Get a file descriptor handle given an Inode handle
 359  * Takes the given file descriptor, and creates a new FdHandle_t for it,
 360  * attached to the given IHandle_t. fd can be INVALID_FD, indicating that the
 361  * caller failed to open the relevant file because we had too many FDs open;
 362  * ih_attachfd_r will then just evict/close an existing fd in the cache, and
 363  * return NULL.
 364  */
 365 static FdHandle_t *
 366 ih_attachfd_r(IHandle_t *ihP, FD_t fd)
 367 {
 368     FD_t closeFd;
 369     FdHandle_t *fdP;
 370
 371     /* fdCacheSize limits the size of the descriptor cache, but
 372      * we permit the number of open files to exceed fdCacheSize.
 373      * We only recycle open file descriptors when the number
 374      * of open files reaches the size of the cache */
 375     if ((fdInUseCount > fdCacheSize || fd == INVALID_FD)  && fdLruHead != NULL) {
 376         fdP = fdLruHead;
 377         osi_Assert(fdP->fd_status == FD_HANDLE_OPEN);
 378         DLL_DELETE(fdP, fdLruHead, fdLruTail, fd_next, fd_prev);
 379         DLL_DELETE(fdP, fdP->fd_ih->ih_fdhead, fdP->fd_ih->ih_fdtail,
 380                    fd_ihnext, fd_ihprev);
 381         closeFd = fdP->fd_fd;
 382         if (fd == INVALID_FD) {
 383             fdCacheSize--;          /* reduce in order to not run into here too often */
 384             DLL_INSERT_TAIL(fdP, fdAvailHead, fdAvailTail, fd_next, fd_prev);
 385             fdP->fd_status = FD_HANDLE_AVAIL;
 386             fdP->fd_ih = NULL;
 387             fdP->fd_fd = INVALID_FD;
 388             IH_UNLOCK;
 389             OS_CLOSE(closeFd);
 390             IH_LOCK;
 391             fdInUseCount -= 1;
 392             return NULL;
 393         }
 394     } else {
 395         if (fdAvailHead == NULL) {
 396             fdHandleAllocateChunk();
 397         }
 398         fdP = fdAvailHead;
 399         osi_Assert(fdP->fd_status == FD_HANDLE_AVAIL);
 400         DLL_DELETE(fdP, fdAvailHead, fdAvailTail, fd_next, fd_prev);
 401         closeFd = INVALID_FD;
 402     }
 403
 404     fdP->fd_status = FD_HANDLE_INUSE;
 405     fdP->fd_fd = fd;
 406     fdP->fd_ih = ihP;
 407     fdP->fd_refcnt++;
 408
 409     ihP->ih_refcnt++;
 410
 411     /* Add this handle to the Inode's list of open descriptors */
 412     DLL_INSERT_TAIL(fdP, ihP->ih_fdhead, ihP->ih_fdtail, fd_ihnext,
 413                     fd_ihprev);
 414
 415     if (closeFd != INVALID_FD) {
 416         IH_UNLOCK;
 417         OS_CLOSE(closeFd);
 418         IH_LOCK;
 419         fdInUseCount -= 1;
 420     }
 421
 422     return fdP;
 423 }
 424
 425 FdHandle_t *
 426 ih_attachfd(IHandle_t *ihP, FD_t fd)
 427 {
 428     FdHandle_t *fdP;
 429
 430     IH_LOCK;
 431
 432     fdInUseCount += 1;
 433
 434     fdP = ih_attachfd_r(ihP, fd);
 435     if (!fdP) {
 436         fdInUseCount -= 1;
 437     }
 438
 439     IH_UNLOCK;
 440
 441     return fdP;
 442 }
 443
 444 /*
 445  * Get a file descriptor handle given an Inode handle
 446  */
 447 FdHandle_t *
 448 ih_open(IHandle_t * ihP)
 449 {
 450     FdHandle_t *fdP;
 451     FD_t fd;
 452
 453     if (!ihP)                   /* XXX should log here in the fileserver */
 454         return NULL;
 455
 456     IH_LOCK;
 457
 458     /* Do we already have an open file handle for this Inode? */
 459     for (fdP = ihP->ih_fdtail; fdP != NULL; fdP = fdP->fd_ihprev) {
 460         if (fdP->fd_status == FD_HANDLE_CLOSING) {
 461             /* The handle was open when an IH_REALLYCLOSE was issued, so we
 462              * cannot reuse it; it will be closed soon. */
 463             continue;
 464         }
 465 #ifndef HAVE_PIO
 466         /*
 467          * If we don't have positional i/o, don't try to share fds, since
 468          * we can't do so in a threadsafe way.
 469          */
 470         if (fdP->fd_status == FD_HANDLE_INUSE) {
 471             continue;
 472         }
 473         osi_Assert(fdP->fd_status == FD_HANDLE_OPEN);
 474 #else /* HAVE_PIO */
 475         osi_Assert(fdP->fd_status != FD_HANDLE_AVAIL);
 476 #endif /* HAVE_PIO */
 477
 478         fdP->fd_refcnt++;
 479         if (fdP->fd_status == FD_HANDLE_OPEN) {
 480             fdP->fd_status = FD_HANDLE_INUSE;
 481             DLL_DELETE(fdP, fdLruHead, fdLruTail, fd_next, fd_prev);
 482         }
 483         ihP->ih_refcnt++;
 484         IH_UNLOCK;
 485         return fdP;
 486     }
 487
 488     /*
 489      * Try to open the Inode, return NULL on error.
 490      */
 491     fdInUseCount += 1;
 492     IH_UNLOCK;
 493 ih_open_retry:
 494     fd = OS_IOPEN(ihP);
 495     IH_LOCK;
 496     if (fd == INVALID_FD && (errno != EMFILE || fdLruHead == NULL) ) {
 497         fdInUseCount -= 1;
 498         IH_UNLOCK;
 499         return NULL;
 500     }
 501
 502     fdP = ih_attachfd_r(ihP, fd);
 503     if (!fdP) {
 504         osi_Assert(fd == INVALID_FD);
 505         IH_UNLOCK;
 506         goto ih_open_retry;
 507     }
 508
 509     IH_UNLOCK;
 510
 511     return fdP;
 512 }
 513
 514 /*
 515  * Return a file descriptor handle to the cache
 516  */
 517 int
 518 fd_close(FdHandle_t * fdP)
 519 {
 520     IHandle_t *ihP;
 521
 522     if (!fdP)
 523         return 0;
 524
 525     IH_LOCK;
 526     osi_Assert(ih_Inited);
 527     osi_Assert(fdInUseCount > 0);
 528     osi_Assert(fdP->fd_status == FD_HANDLE_INUSE ||
 529                fdP->fd_status == FD_HANDLE_CLOSING);
 530
 531     ihP = fdP->fd_ih;
 532
 533     /* Call fd_reallyclose to really close the unused file handles if
 534      * the previous attempt to close (ih_reallyclose()) all file handles
 535      * failed (this is determined by checking the ihandle for the flag
 536      * IH_REALLY_CLOSED) or we have too many open files.
 537      */
 538     if (fdP->fd_status == FD_HANDLE_CLOSING ||
 539         ihP->ih_flags & IH_REALLY_CLOSED || fdInUseCount > fdCacheSize) {
 540         IH_UNLOCK;
 541         return fd_reallyclose(fdP);
 542     }
 543
 544     fdP->fd_refcnt--;
 545     if (fdP->fd_refcnt == 0) {
 546         /* Put this descriptor back into the cache */
 547         fdP->fd_status = FD_HANDLE_OPEN;
 548         DLL_INSERT_TAIL(fdP, fdLruHead, fdLruTail, fd_next, fd_prev);
 549     }
 550
 551     /* If this is not the only reference to the Inode then we can decrement
 552      * the reference count, otherwise we need to call ih_release.
 553      */
 554     if (ihP->ih_refcnt > 1) {
 555         ihP->ih_refcnt--;
 556         IH_UNLOCK;
 557     } else {
 558         IH_UNLOCK;
 559         ih_release(ihP);
 560     }
 561
 562     return 0;
 563 }
 564
 565 /*
 566  * Actually close the file descriptor handle and return it to
 567  * the free list.
 568  */
 569 int
 570 fd_reallyclose(FdHandle_t * fdP)
 571 {
 572     FD_t closeFd;
 573     IHandle_t *ihP;
 574
 575     if (!fdP)
 576         return 0;
 577
 578     IH_LOCK;
 579     osi_Assert(ih_Inited);
 580     osi_Assert(fdInUseCount > 0);
 581     osi_Assert(fdP->fd_status == FD_HANDLE_INUSE ||
 582                fdP->fd_status == FD_HANDLE_CLOSING);
 583
 584     ihP = fdP->fd_ih;
 585     closeFd = fdP->fd_fd;
 586     fdP->fd_refcnt--;
 587
 588     if (fdP->fd_refcnt == 0) {
 589         DLL_DELETE(fdP, ihP->ih_fdhead, ihP->ih_fdtail, fd_ihnext, fd_ihprev);
 590         DLL_INSERT_TAIL(fdP, fdAvailHead, fdAvailTail, fd_next, fd_prev);
 591
 592         fdP->fd_status = FD_HANDLE_AVAIL;
 593         fdP->fd_refcnt = 0;
 594         fdP->fd_ih = NULL;
 595         fdP->fd_fd = INVALID_FD;
 596     }
 597
 598     /* All the file descriptor handles have been closed; reset
 599      * the IH_REALLY_CLOSED flag indicating that ih_reallyclose
 600      * has completed its job.
 601      */
 602     if (!ihP->ih_fdhead) {
 603         ihP->ih_flags &= ~IH_REALLY_CLOSED;
 604     } else {
 605         FdHandle_t *lfdP, *next;
 606         int clear = 1;
 607         for (lfdP = ihP->ih_fdhead; lfdP != NULL; lfdP = next) {
 608             next = lfdP->fd_ihnext;
 609             osi_Assert(lfdP->fd_ih == ihP);
 610             if (lfdP->fd_status != FD_HANDLE_CLOSING) {
 611                 clear = 0;
 612                 break;
 613             }
 614         }
 615         /* no *future* fd should be subjected to this */
 616         if (clear)
 617             ihP->ih_flags &= ~IH_REALLY_CLOSED;
 618     }
 619
 620     if (fdP->fd_refcnt == 0) {
 621         IH_UNLOCK;
 622         OS_CLOSE(closeFd);
 623         IH_LOCK;
 624         fdInUseCount -= 1;
 625     }
 626
 627     /* If this is not the only reference to the Inode then we can decrement
 628      * the reference count, otherwise we need to call ih_release. */
 629     if (ihP->ih_refcnt > 1) {
 630         ihP->ih_refcnt--;
 631         IH_UNLOCK;
 632     } else {
 633         IH_UNLOCK;
 634         ih_release(ihP);
 635     }
 636
 637     return 0;
 638 }
 639
 640 /* Enable buffered I/O on a file descriptor */
 641 StreamHandle_t *
 642 stream_fdopen(FD_t fd)
 643 {
 644     StreamHandle_t *streamP;
 645
 646     IH_LOCK;
 647     if (streamAvailHead == NULL) {
 648         streamHandleAllocateChunk();
 649     }
 650     streamP = streamAvailHead;
 651     DLL_DELETE(streamP, streamAvailHead, streamAvailTail, str_next, str_prev);
 652     IH_UNLOCK;
 653     streamP->str_fd = fd;
 654     streamP->str_buflen = 0;
 655     streamP->str_bufoff = 0;
 656     streamP->str_fdoff = 0;
 657     streamP->str_error = 0;
 658     streamP->str_eof = 0;
 659     streamP->str_direction = STREAM_DIRECTION_NONE;
 660     return streamP;
 661 }
 662
 663 /* Open a file for buffered I/O */
 664 StreamHandle_t *
 665 stream_open(const char *filename, const char *mode)
 666 {
 667     FD_t fd = INVALID_FD;
 668
 669     if (strcmp(mode, "r") == 0) {
 670         fd = OS_OPEN(filename, O_RDONLY, 0);
 671     } else if (strcmp(mode, "r+") == 0) {
 672         fd = OS_OPEN(filename, O_RDWR, 0);
 673     } else if (strcmp(mode, "w") == 0) {
 674         fd = OS_OPEN(filename, O_WRONLY | O_TRUNC | O_CREAT, 0);
 675     } else if (strcmp(mode, "w+") == 0) {
 676         fd = OS_OPEN(filename, O_RDWR | O_TRUNC | O_CREAT, 0);
 677     } else if (strcmp(mode, "a") == 0) {
 678         fd = OS_OPEN(filename, O_WRONLY | O_APPEND | O_CREAT, 0);
 679     } else if (strcmp(mode, "a+") == 0) {
 680         fd = OS_OPEN(filename, O_RDWR | O_APPEND | O_CREAT, 0);
 681     } else {
 682         osi_Assert(FALSE);              /* not implemented */
 683     }
 684
 685     if (fd == INVALID_FD) {
 686         return NULL;
 687     }
 688     return stream_fdopen(fd);
 689 }
 690
 691 /* fread for buffered I/O handles */
 692 afs_sfsize_t
 693 stream_read(void *ptr, afs_fsize_t size, afs_fsize_t nitems,
 694             StreamHandle_t * streamP)
 695 {
 696     afs_fsize_t nbytes, bytesRead, bytesToRead;
 697     char *p;
 698
 699     /* Need to seek before changing direction */
 700     if (streamP->str_direction == STREAM_DIRECTION_NONE) {
 701         streamP->str_direction = STREAM_DIRECTION_READ;
 702         streamP->str_bufoff = 0;
 703         streamP->str_buflen = 0;
 704     } else {
 705         osi_Assert(streamP->str_direction == STREAM_DIRECTION_READ);
 706     }
 707
 708     bytesRead = 0;
 709     nbytes = size * nitems;
 710     p = (char *)ptr;
 711     while (nbytes > 0 && !streamP->str_eof) {
 712         if (streamP->str_buflen == 0) {
 713             streamP->str_bufoff = 0;
 714             streamP->str_buflen =
 715                 OS_PREAD(streamP->str_fd, streamP->str_buffer,
 716                         STREAM_HANDLE_BUFSIZE, streamP->str_fdoff);
 717             if (streamP->str_buflen < 0) {
 718                 streamP->str_error = errno;
 719                 streamP->str_buflen = 0;
 720                 bytesRead = 0;
 721                 break;
 722             } else if (streamP->str_buflen == 0) {
 723                 streamP->str_eof = 1;
 724                 break;
 725             }
 726             streamP->str_fdoff += streamP->str_buflen;
 727         }
 728
 729         bytesToRead = nbytes;
 730         if (bytesToRead > streamP->str_buflen) {
 731             bytesToRead = streamP->str_buflen;
 732         }
 733         memcpy(p, streamP->str_buffer + streamP->str_bufoff, bytesToRead);
 734         p += bytesToRead;
 735         streamP->str_bufoff += bytesToRead;
 736         streamP->str_buflen -= bytesToRead;
 737         bytesRead += bytesToRead;
 738         nbytes -= bytesToRead;
 739     }
 740
 741     return (bytesRead / size);
 742 }
 743
 744 /* fwrite for buffered I/O handles */
 745 afs_sfsize_t
 746 stream_write(void *ptr, afs_fsize_t size, afs_fsize_t nitems,
 747              StreamHandle_t * streamP)
 748 {
 749     char *p;
 750     afs_sfsize_t rc;
 751     afs_fsize_t nbytes, bytesWritten, bytesToWrite;
 752
 753     /* Need to seek before changing direction */
 754     if (streamP->str_direction == STREAM_DIRECTION_NONE) {
 755         streamP->str_direction = STREAM_DIRECTION_WRITE;
 756         streamP->str_bufoff = 0;
 757         streamP->str_buflen = STREAM_HANDLE_BUFSIZE;
 758     } else {
 759         osi_Assert(streamP->str_direction == STREAM_DIRECTION_WRITE);
 760     }
 761
 762     nbytes = size * nitems;
 763     bytesWritten = 0;
 764     p = (char *)ptr;
 765     while (nbytes > 0) {
 766         if (streamP->str_buflen == 0) {
 767             rc = OS_PWRITE(streamP->str_fd, streamP->str_buffer,
 768                           STREAM_HANDLE_BUFSIZE, streamP->str_fdoff);
 769             if (rc < 0) {
 770                 streamP->str_error = errno;
 771                 bytesWritten = 0;
 772                 break;
 773             }
 774             streamP->str_fdoff += rc;
 775             streamP->str_bufoff = 0;
 776             streamP->str_buflen = STREAM_HANDLE_BUFSIZE;
 777         }
 778
 779         bytesToWrite = nbytes;
 780         if (bytesToWrite > streamP->str_buflen) {
 781             bytesToWrite = streamP->str_buflen;
 782         }
 783         memcpy(streamP->str_buffer + streamP->str_bufoff, p, bytesToWrite);
 784         p += bytesToWrite;
 785         streamP->str_bufoff += bytesToWrite;
 786         streamP->str_buflen -= bytesToWrite;
 787         bytesWritten += bytesToWrite;
 788         nbytes -= bytesToWrite;
 789     }
 790
 791     return (bytesWritten / size);
 792 }
 793
 794 /* fseek for buffered I/O handles */
 795 int
 796 stream_aseek(StreamHandle_t * streamP, afs_foff_t offset)
 797 {
 798     ssize_t rc;
 799     int retval = 0;
 800
 801     if (streamP->str_direction == STREAM_DIRECTION_WRITE
 802         && streamP->str_bufoff > 0) {
 803         rc = OS_PWRITE(streamP->str_fd, streamP->str_buffer,
 804                       streamP->str_bufoff, streamP->str_fdoff);
 805         if (rc < 0) {
 806             streamP->str_error = errno;
 807             retval = -1;
 808         }
 809     }
 810     streamP->str_fdoff = offset;
 811     streamP->str_bufoff = 0;
 812     streamP->str_buflen = 0;
 813     streamP->str_eof = 0;
 814     streamP->str_direction = STREAM_DIRECTION_NONE;
 815     return retval;
 816 }
 817
 818 /* fflush for buffered I/O handles */
 819 int
 820 stream_flush(StreamHandle_t * streamP)
 821 {
 822     ssize_t rc;
 823     int retval = 0;
 824
 825     if (streamP->str_direction == STREAM_DIRECTION_WRITE
 826         && streamP->str_bufoff > 0) {
 827         rc = OS_PWRITE(streamP->str_fd, streamP->str_buffer,
 828                       streamP->str_bufoff, streamP->str_fdoff);
 829         if (rc < 0) {
 830             streamP->str_error = errno;
 831             retval = -1;
 832         } else {
 833             streamP->str_fdoff += rc;
 834         }
 835         streamP->str_bufoff = 0;
 836         streamP->str_buflen = STREAM_HANDLE_BUFSIZE;
 837     }
 838
 839     return retval;
 840 }
 841
 842 /* Free a buffered I/O handle */
 843 int
 844 stream_close(StreamHandle_t * streamP, int reallyClose)
 845 {
 846     ssize_t rc;
 847     int retval = 0;
 848
 849     osi_Assert(streamP != NULL);
 850     if (streamP->str_direction == STREAM_DIRECTION_WRITE
 851         && streamP->str_bufoff > 0) {
 852         rc = OS_PWRITE(streamP->str_fd, streamP->str_buffer,
 853                       streamP->str_bufoff, streamP->str_fdoff);
 854         if (rc < 0) {
 855             retval = -1;
 856         } else {
 857             streamP->str_fdoff += rc;
 858         }
 859     }
 860     if (reallyClose) {
 861         rc = OS_CLOSE(streamP->str_fd);
 862         if (rc < 0) {
 863             retval = -1;
 864         }
 865     }
 866     streamP->str_fd = INVALID_FD;
 867
 868     IH_LOCK;
 869     DLL_INSERT_TAIL(streamP, streamAvailHead, streamAvailTail,
 870                     str_next, str_prev);
 871     IH_UNLOCK;
 872     return retval;
 873 }
 874
 875 /* Close all unused file descriptors associated with the inode
 876  * handle. Called with IH_LOCK held. May drop and reacquire
 877  * IH_LOCK. Sets the IH_REALLY_CLOSED flag in the inode handle
 878  * if it fails to close all file handles.
 879  */
 880 static int
 881 ih_fdclose(IHandle_t * ihP)
 882 {
 883     int closeCount, closedAll;
 884     FdHandle_t *fdP, *head, *tail, *next;
 885
 886     osi_Assert(ihP->ih_refcnt > 0);
 887
 888     closedAll = 1;
 889     DLL_INIT_LIST(head, tail);
 890     ihP->ih_flags &= ~IH_REALLY_CLOSED;
 891
 892     /*
 893      * Remove the file descriptors for this Inode from the LRU queue
 894      * and the IHandle queue and put them on a temporary queue so we
 895      * can drop the lock before we close the files.
 896      */
 897     for (fdP = ihP->ih_fdhead; fdP != NULL; fdP = next) {
 898         next = fdP->fd_ihnext;
 899         osi_Assert(fdP->fd_ih == ihP);
 900         osi_Assert(fdP->fd_status == FD_HANDLE_OPEN
 901                || fdP->fd_status == FD_HANDLE_INUSE
 902                || fdP->fd_status == FD_HANDLE_CLOSING);
 903         if (fdP->fd_status == FD_HANDLE_OPEN) {
 904             /* Note that FdHandle_t's do not count against the parent
 905              * IHandle_t ref count when they are FD_HANDLE_OPEN. So, we don't
 906              * need to dec the parent IHandle_t ref count for each one we pull
 907              * off here. */
 908             DLL_DELETE(fdP, ihP->ih_fdhead, ihP->ih_fdtail, fd_ihnext,
 909                        fd_ihprev);
 910             DLL_DELETE(fdP, fdLruHead, fdLruTail, fd_next, fd_prev);
 911             DLL_INSERT_TAIL(fdP, head, tail, fd_next, fd_prev);
 912         } else {
 913             closedAll = 0;
 914             fdP->fd_status = FD_HANDLE_CLOSING;
 915             ihP->ih_flags |= IH_REALLY_CLOSED;
 916         }
 917     }
 918
 919     /* If the ihandle reference count is 1, we should have
 920      * closed all file descriptors.
 921      */
 922     if (ihP->ih_refcnt == 1 || closedAll) {
 923         osi_Assert(closedAll);
 924         osi_Assert(!ihP->ih_fdhead);
 925         osi_Assert(!ihP->ih_fdtail);
 926     }
 927
 928     if (head == NULL) {
 929         return 0;               /* No file descriptors closed */
 930     }
 931
 932     IH_UNLOCK;
 933     /*
 934      * Close the file descriptors
 935      */
 936     closeCount = 0;
 937     for (fdP = head; fdP != NULL; fdP = fdP->fd_next) {
 938         OS_CLOSE(fdP->fd_fd);
 939         fdP->fd_status = FD_HANDLE_AVAIL;
 940         fdP->fd_refcnt = 0;
 941         fdP->fd_fd = INVALID_FD;
 942         fdP->fd_ih = NULL;
 943         closeCount++;
 944     }
 945
 946     IH_LOCK;
 947     osi_Assert(fdInUseCount >= closeCount);
 948     fdInUseCount -= closeCount;
 949
 950     /*
 951      * Append the temporary queue to the list of available descriptors
 952      */
 953     if (fdAvailHead == NULL) {
 954         fdAvailHead = head;
 955         fdAvailTail = tail;
 956     } else {
 957         fdAvailTail->fd_next = head;
 958         head->fd_prev = fdAvailTail;
 959         fdAvailTail = tail;
 960     }
 961
 962     return 0;
 963 }
 964
 965 /* Close all cached file descriptors for this inode. */
 966 int
 967 ih_reallyclose(IHandle_t * ihP)
 968 {
 969     if (!ihP)
 970         return 0;
 971
 972     IH_LOCK;
 973     ihP->ih_refcnt++;   /* must not disappear over unlock */
 974     if (ihP->ih_synced) {
 975         FdHandle_t *fdP;
 976         osi_Assert(vol_io_params.sync_behavior != IH_SYNC_ALWAYS);
 977         osi_Assert(vol_io_params.sync_behavior != IH_SYNC_NEVER);
 978         ihP->ih_synced = 0;
 979         IH_UNLOCK;
 980
 981         fdP = IH_OPEN(ihP);
 982         if (fdP) {
 983             OS_SYNC(fdP->fd_fd);
 984             FDH_CLOSE(fdP);
 985         }
 986
 987         IH_LOCK;
 988     }
 989
 990     osi_Assert(ihP->ih_refcnt > 0);
 991
 992     ih_fdclose(ihP);
 993
 994     if (ihP->ih_refcnt > 1) {
 995         ihP->ih_refcnt--;
 996         IH_UNLOCK;
 997     } else {
 998         IH_UNLOCK;
 999         ih_release(ihP);
1000     }
1001     return 0;
1002 }
1003
1004 /* Release an Inode handle. All cached file descriptors for this
1005  * inode are closed when the last reference to this handle is released
1006  */
1007 int
1008 ih_release(IHandle_t * ihP)
1009 {
1010     int ihash;
1011
1012     if (!ihP)
1013         return 0;
1014
1015     IH_LOCK;
1016     osi_Assert(ihP->ih_refcnt > 0);
1017
1018     if (ihP->ih_refcnt > 1) {
1019         ihP->ih_refcnt--;
1020         IH_UNLOCK;
1021         return 0;
1022     }
1023
1024     ihash = IH_HASH(ihP->ih_dev, ihP->ih_vid, ihP->ih_ino);
1025     DLL_DELETE(ihP, ihashTable[ihash].ihash_head,
1026                ihashTable[ihash].ihash_tail, ih_next, ih_prev);
1027
1028     ih_fdclose(ihP);
1029
1030     ihP->ih_refcnt--;
1031
1032     DLL_INSERT_TAIL(ihP, ihAvailHead, ihAvailTail, ih_next, ih_prev);
1033
1034     IH_UNLOCK;
1035     return 0;
1036 }
1037
1038 /* Sync an inode to disk if its handle isn't NULL */
1039 int
1040 ih_condsync(IHandle_t * ihP)
1041 {
1042     int code;
1043     FdHandle_t *fdP;
1044
1045     if (!ihP)
1046         return 0;
1047
1048     fdP = IH_OPEN(ihP);
1049     if (fdP == NULL)
1050         return -1;
1051
1052     code = FDH_SYNC(fdP);
1053     FDH_CLOSE(fdP);
1054
1055     return code;
1056 }
1057
1058 void
1059 ih_sync_all(void) {
1060
1061     int ihash;
1062
1063     IH_LOCK;
1064     for (ihash = 0; ihash < I_HANDLE_HASH_SIZE; ihash++) {
1065         IHandle_t *ihP, *ihPnext;
1066
1067         ihP = ihashTable[ihash].ihash_head;
1068         if (ihP)
1069             ihP->ih_refcnt++;   /* must not disappear over unlock */
1070         for (; ihP; ihP = ihPnext) {
1071
1072             if (ihP->ih_synced) {
1073                 FD_t fd;
1074
1075                 ihP->ih_synced = 0;
1076                 IH_UNLOCK;
1077
1078                 fd = OS_IOPEN(ihP);
1079                 if (fd != INVALID_FD) {
1080                     OS_SYNC(fd);
1081                     OS_CLOSE(fd);
1082                 }
1083
1084                 IH_LOCK;
1085             }
1086
1087             /* when decrementing the refcount, the ihandle might disappear
1088                and we might not even be able to proceed to the next one.
1089                Hence the gymnastics putting a hold on the next one already */
1090             ihPnext = ihP->ih_next;
1091             if (ihPnext) ihPnext->ih_refcnt++;
1092
1093             if (ihP->ih_refcnt > 1) {
1094                 ihP->ih_refcnt--;
1095             } else {
1096                 IH_UNLOCK;
1097                 ih_release(ihP);
1098                 IH_LOCK;
1099             }
1100
1101         }
1102     }
1103     IH_UNLOCK;
1104 }
1105
1106 void *
1107 ih_sync_thread(void *dummy) {
1108     while(1) {
1109
1110 #ifdef AFS_PTHREAD_ENV
1111         sleep(10);
1112 #else /* AFS_PTHREAD_ENV */
1113         IOMGR_Sleep(60);
1114 #endif /* AFS_PTHREAD_ENV */
1115
1116         ih_sync_all();
1117     }
1118     return NULL;
1119 }
1120
1121
1122 /*************************************************************************
1123  * OS specific support routines.
1124  *************************************************************************/
1125 #ifndef AFS_NAMEI_ENV
1126 Inode
1127 ih_icreate(IHandle_t * ih, int dev, char *part, Inode nI, int p1, int p2,
1128            int p3, int p4)
1129 {
1130     Inode ino;
1131 #ifdef  AFS_3DISPARES
1132     /* See viceinode.h */
1133     if (p2 == INODESPECIAL) {
1134         int tp = p3;
1135         p3 = p4;
1136         p4 = tp;
1137     }
1138 #endif
1139     ino = ICREATE(dev, part, nI, p1, p2, p3, p4);
1140     return ino;
1141 }
1142 #endif /* AFS_NAMEI_ENV */
1143
1144 #if defined(AFS_NT40_ENV) || !defined(AFS_NAMEI_ENV)
1145 /* Unix namei implements its own more efficient IH_CREATE_INIT; this wrapper
1146  * is for everyone else */
1147 IHandle_t *
1148 ih_icreate_init(IHandle_t *lh, int dev, char *part, Inode nearInode,
1149                 afs_uint32 p1, afs_uint32 p2, afs_uint32 p3, afs_uint32 p4)
1150 {
1151     IHandle_t *ihP;
1152     Inode ino = IH_CREATE(lh, dev, part, nearInode, p1, p2, p3, p4);
1153     if (!VALID_INO(ino)) {
1154         return NULL;
1155     }
1156     IH_INIT(ihP, dev, p1, ino);
1157     return ihP;
1158 }
1159 #endif
1160
1161 afs_sfsize_t
1162 ih_size(FD_t fd)
1163 {
1164 #ifdef AFS_NT40_ENV
1165     LARGE_INTEGER size;
1166     if (!GetFileSizeEx(fd, &size))
1167         return -1;
1168     return size.QuadPart;
1169 #else
1170     struct afs_stat status;
1171     if (afs_fstat(fd, &status) < 0)
1172         return -1;
1173     return status.st_size;
1174 #endif
1175 }
1176
1177 #ifndef HAVE_PIO
1178 ssize_t
1179 ih_pread(int fd, void * buf, size_t count, afs_foff_t offset)
1180 {
1181         afs_foff_t code;
1182         code = OS_SEEK(fd, offset, 0);
1183         if (code < 0)
1184             return code;
1185         return OS_READ(fd, buf, count);
1186 }
1187
1188 ssize_t
1189 ih_pwrite(int fd, const void * buf, size_t count, afs_foff_t offset)
1190 {
1191         afs_foff_t code;
1192         code = OS_SEEK(fd, offset, 0);
1193         if (code < 0)
1194             return code;
1195         return OS_WRITE(fd, buf, count);
1196 }
1197 #endif /* !HAVE_PIO */
1198
1199 #ifndef AFS_NT40_ENV
1200 int
1201 ih_isunlinked(int fd)
1202 {
1203     struct afs_stat status;
1204     if (afs_fstat(fd, &status) < 0) {
1205         return -1;
1206     }
1207     if (status.st_nlink < 1) {
1208         return 1;
1209     }
1210     return 0;
1211 }
1212 #endif /* !AFS_NT40_ENV */
1213
1214 int
1215 ih_fdsync(FdHandle_t *fdP)
1216 {
1217     switch (vol_io_params.sync_behavior) {
1218     case IH_SYNC_ALWAYS:
1219         return OS_SYNC(fdP->fd_fd);
1220     case IH_SYNC_DELAYED:
1221     case IH_SYNC_ONCLOSE:
1222         if (fdP->fd_ih) {
1223             fdP->fd_ih->ih_synced = 1;
1224             return 0;
1225         }
1226         return 1;
1227     case IH_SYNC_NEVER:
1228         return 0;
1229     default:
1230         osi_Assert(0);
1231     }
1232 }