src/WINNT/afsd/cm_buf.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 /* Copyright (C) 1994 Cazamar Systems, Inc. */
  11
  12 #include <afs/param.h>
  13 #include <afs/stds.h>
  14
  15 #ifndef DJGPP
  16 #include <windows.h>
  17 #endif
  18 #include <osi.h>
  19 #include <malloc.h>
  20 #include <stdio.h>
  21 #include <assert.h>
  22
  23 #include "afsd.h"
  24
  25 extern void afsi_log(char *pattern, ...);
  26
  27 /* This module implements the buffer package used by the local transaction
  28  * system (cm).  It is initialized by calling cm_Init, which calls buf_Init;
  29  * it must be initalized before any of its main routines are called.
  30  *
  31  * Each buffer is hashed into a hash table by file ID and offset, and if its
  32  * reference count is zero, it is also in a free list.
  33  *
  34  * There are two locks involved in buffer processing.  The global lock
  35  * buf_globalLock protects all of the global variables defined in this module,
  36  * the reference counts and hash pointers in the actual cm_buf_t structures,
  37  * and the LRU queue pointers in the buffer structures.
  38  *
  39  * The mutexes in the buffer structures protect the remaining fields in the
  40  * buffers, as well the data itself.
  41  *
  42  * The locking hierarchy here is this:
  43  *
  44  * - resv multiple simul. buffers reservation
  45  * - lock buffer I/O flags
  46  * - lock buffer's mutex
  47  * - lock buf_globalLock
  48  *
  49  */
  50
  51 /* global debugging log */
  52 osi_log_t *buf_logp = NULL;
  53
  54 /* Global lock protecting hash tables and free lists */
  55 osi_rwlock_t buf_globalLock;
  56
  57 /* ptr to head of the free list (most recently used) and the
  58  * tail (the guy to remove first).  We use osi_Q* functions
  59  * to put stuff in buf_freeListp, and maintain the end
  60  * pointer manually
  61  */
  62 cm_buf_t *buf_freeListp;
  63 cm_buf_t *buf_freeListEndp;
  64
  65 /* a pointer to a list of all buffers, just so that we can find them
  66  * easily for debugging, and for the incr syncer.  Locked under
  67  * the global lock.
  68  */
  69 cm_buf_t *buf_allp;
  70
  71 /* defaults setup; these variables may be manually assigned into
  72  * before calling cm_Init, as a way of changing these defaults.
  73  */
  74 long buf_nbuffers = CM_BUF_BUFFERS;
  75 long buf_nOrigBuffers;
  76 long buf_bufferSize = CM_BUF_SIZE;
  77 long buf_hashSize = CM_BUF_HASHSIZE;
  78
  79 #ifndef DJGPP
  80 static
  81 HANDLE CacheHandle;
  82
  83 static
  84 SYSTEM_INFO sysInfo;
  85 #endif /* !DJGPP */
  86
  87 /* buffer reservation variables */
  88 long buf_reservedBufs;
  89 long buf_maxReservedBufs;
  90 int buf_reserveWaiting;
  91
  92 /* callouts for reading and writing data, etc */
  93 cm_buf_ops_t *cm_buf_opsp;
  94
  95 /* pointer to hash table; size computed dynamically */
  96 cm_buf_t **buf_hashTablepp;
  97
  98 /* another hash table */
  99 cm_buf_t **buf_fileHashTablepp;
 100
 101 #ifdef DISKCACHE95
 102 /* for experimental disk caching support in Win95 client */
 103 cm_buf_t *buf_diskFreeListp;
 104 cm_buf_t *buf_diskFreeListEndp;
 105 cm_buf_t *buf_diskAllp;
 106 extern int cm_diskCacheEnabled;
 107 #endif /* DISKCACHE95 */
 108
 109 /* hold a reference to an already held buffer */
 110 void buf_Hold(cm_buf_t *bp)
 111 {
 112         lock_ObtainWrite(&buf_globalLock);
 113         bp->refCount++;
 114         lock_ReleaseWrite(&buf_globalLock);
 115 }
 116
 117 /* incremental sync daemon.  Writes 1/10th of all the buffers every 5000 ms */
 118 void buf_IncrSyncer(long parm)
 119 {
 120         cm_buf_t *bp;                   /* buffer we're hacking on; held */
 121         long i;                         /* counter */
 122         long nAtOnce;                   /* how many to do at once */
 123         cm_req_t req;
 124
 125         lock_ObtainWrite(&buf_globalLock);
 126         bp = buf_allp;
 127         bp->refCount++;
 128         lock_ReleaseWrite(&buf_globalLock);
 129         nAtOnce = buf_nbuffers / 10;
 130         while (1) {
 131 #ifndef DJGPP
 132                 i = SleepEx(5000, 1);
 133                 if (i != 0) continue;
 134 #else
 135                 thrd_Sleep(5000);
 136 #endif /* DJGPP */
 137
 138                 /* now go through our percentage of the buffers */
 139                 for(i=0; i<nAtOnce; i++) {
 140                         /* don't want its identity changing while we're
 141                          * messing with it, so must do all of this with
 142                          * bp held.
 143                          */
 144
 145                         /* start cleaning the buffer; don't touch log pages since
 146                          * the log code counts on knowing exactly who is writing
 147                          * a log page at any given instant.
 148                          */
 149                         cm_InitReq(&req);
 150                         req.flags |= CM_REQ_NORETRY;
 151                         buf_CleanAsync(bp, &req);
 152
 153                         /* now advance to the next buffer; the allp chain never changes,
 154                          * and so can be followed even when holding no locks.
 155                          */
 156                         lock_ObtainWrite(&buf_globalLock);
 157                         buf_LockedRelease(bp);
 158                         bp = bp->allp;
 159                         if (!bp) bp = buf_allp;
 160                         bp->refCount++;
 161                         lock_ReleaseWrite(&buf_globalLock);
 162                 }       /* for loop over a bunch of buffers */
 163         }               /* whole daemon's while loop */
 164 }
 165
 166 #ifndef DJGPP
 167 /* Create a security attribute structure suitable for use when the cache file
 168  * is created.  What we mainly want is that only the administrator should be
 169  * able to do anything with the file.  We create an ACL with only one entry,
 170  * an entry that grants all rights to the administrator.
 171  */
 172 PSECURITY_ATTRIBUTES CreateCacheFileSA()
 173 {
 174         PSECURITY_ATTRIBUTES psa;
 175         PSECURITY_DESCRIPTOR psd;
 176         SID_IDENTIFIER_AUTHORITY authority = SECURITY_NT_AUTHORITY;
 177         PSID AdminSID;
 178         DWORD AdminSIDlength;
 179         PACL AdminOnlyACL;
 180         DWORD ACLlength;
 181
 182         /* Get Administrator SID */
 183         AllocateAndInitializeSid(&authority, 2,
 184                                  SECURITY_BUILTIN_DOMAIN_RID,
 185                                  DOMAIN_ALIAS_RID_ADMINS,
 186                                  0, 0, 0, 0, 0, 0,
 187                                  &AdminSID);
 188
 189         /* Create Administrator-only ACL */
 190         AdminSIDlength = GetLengthSid(AdminSID);
 191         ACLlength = sizeof(ACL) + sizeof(ACCESS_ALLOWED_ACE)
 192                         + AdminSIDlength - sizeof(DWORD);
 193         AdminOnlyACL = GlobalAlloc(GMEM_FIXED, ACLlength);
 194         InitializeAcl(AdminOnlyACL, ACLlength, ACL_REVISION);
 195         AddAccessAllowedAce(AdminOnlyACL, ACL_REVISION,
 196                             STANDARD_RIGHTS_ALL | SPECIFIC_RIGHTS_ALL,
 197                             AdminSID);
 198
 199         /* Create security descriptor */
 200         psd = GlobalAlloc(GMEM_FIXED, sizeof(SECURITY_DESCRIPTOR));
 201         InitializeSecurityDescriptor(psd, SECURITY_DESCRIPTOR_REVISION);
 202         SetSecurityDescriptorDacl(psd, TRUE, AdminOnlyACL, FALSE);
 203
 204         /* Create security attributes structure */
 205         psa = GlobalAlloc(GMEM_FIXED, sizeof(SECURITY_ATTRIBUTES));
 206         psa->nLength = sizeof(SECURITY_ATTRIBUTES);
 207         psa->lpSecurityDescriptor = psd;
 208         psa->bInheritHandle = TRUE;
 209
 210         return psa;
 211 }
 212 #endif /* !DJGPP */
 213
 214 #ifndef DJGPP
 215 /* Free a security attribute structure created by CreateCacheFileSA() */
 216 VOID FreeCacheFileSA(PSECURITY_ATTRIBUTES psa)
 217 {
 218         BOOL b1, b2;
 219         PACL pAcl;
 220
 221         GetSecurityDescriptorDacl(psa->lpSecurityDescriptor, &b1, &pAcl, &b2);
 222         GlobalFree(pAcl);
 223         GlobalFree(psa->lpSecurityDescriptor);
 224         GlobalFree(psa);
 225 }
 226 #endif /* !DJGPP */
 227
 228 /* initialize the buffer package; called with no locks
 229  * held during the initialization phase.
 230  */
 231 long buf_Init(cm_buf_ops_t *opsp)
 232 {
 233         static osi_once_t once;
 234         cm_buf_t *bp;
 235         long sectorSize;
 236         thread_t phandle;
 237 #ifndef DJGPP
 238         HANDLE hf, hm;
 239         PSECURITY_ATTRIBUTES psa;
 240 #endif /* !DJGPP */
 241         long i;
 242         unsigned long pid;
 243         char *data;
 244         long cs;
 245
 246 #ifndef DJGPP
 247         /* Get system info; all we really want is the allocation granularity */
 248         GetSystemInfo(&sysInfo);
 249 #endif /* !DJGPP */
 250
 251         /* Have to be able to reserve a whole chunk */
 252         if (((buf_nbuffers - 3) * buf_bufferSize) < cm_chunkSize)
 253                 return CM_ERROR_TOOFEWBUFS;
 254
 255         /* recall for callouts */
 256         cm_buf_opsp = opsp;
 257
 258         if (osi_Once(&once)) {
 259                 /* initialize global locks */
 260                 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock");
 261
 262 #ifndef DJGPP
 263                 /*
 264                  * Cache file mapping constrained by
 265                  * system allocation granularity;
 266                  * round up, assuming granularity is a power of two
 267                  */
 268                 cs = buf_nbuffers * buf_bufferSize;
 269                 cs = (cs + (sysInfo.dwAllocationGranularity - 1))
 270                         & ~(sysInfo.dwAllocationGranularity - 1);
 271                 if (cs != buf_nbuffers * buf_bufferSize) {
 272                         buf_nbuffers = cs / buf_bufferSize;
 273                         afsi_log("Cache size rounded up to %d buffers",
 274                                  buf_nbuffers);
 275                 }
 276 #endif /* !DJGPP */
 277
 278                 /* remember this for those who want to reset it */
 279                 buf_nOrigBuffers = buf_nbuffers;
 280
 281                 /* lower hash size to a prime number */
 282                 buf_hashSize = osi_PrimeLessThan(buf_hashSize);
 283
 284                 /* create hash table */
 285                 buf_hashTablepp = malloc(buf_hashSize * sizeof(cm_buf_t *));
 286                 memset((void *)buf_hashTablepp, 0,
 287                         buf_hashSize * sizeof(cm_buf_t *));
 288
 289                 /* another hash table */
 290                 buf_fileHashTablepp = malloc(buf_hashSize * sizeof(cm_buf_t *));
 291                 memset((void *)buf_fileHashTablepp, 0,
 292                         buf_hashSize * sizeof(cm_buf_t *));
 293
 294                 /* min value for which this works */
 295                 sectorSize = 1;
 296
 297 #ifndef DJGPP
 298                 /* Reserve buffer space by mapping cache file */
 299                 psa = CreateCacheFileSA();
 300                 hf = CreateFile(cm_CachePath,
 301                         GENERIC_READ | GENERIC_WRITE,
 302                         FILE_SHARE_READ | FILE_SHARE_WRITE,
 303                         psa,
 304                         OPEN_ALWAYS,
 305                         FILE_ATTRIBUTE_NORMAL,
 306                         NULL);
 307                 if (hf == INVALID_HANDLE_VALUE) {
 308                         afsi_log("create file error %d", GetLastError());
 309                         return CM_ERROR_INVAL;
 310                 }
 311                 FreeCacheFileSA(psa);
 312                 CacheHandle = hf;
 313                 hm = CreateFileMapping(hf,
 314                         NULL,
 315                         PAGE_READWRITE,
 316                         0, buf_nbuffers * buf_bufferSize,
 317                         NULL);
 318                 if (hm == NULL) {
 319                         if (GetLastError() == ERROR_DISK_FULL) {
 320                                 afsi_log("Error creating cache file mapping: disk full");
 321                                 return CM_ERROR_TOOMANYBUFS;
 322                         }
 323                         return CM_ERROR_INVAL;
 324                 }
 325                 data = MapViewOfFile(hm,
 326                         FILE_MAP_ALL_ACCESS,
 327                         0, 0,
 328                         buf_nbuffers * buf_bufferSize);
 329                 if (data == NULL) {
 330                         CloseHandle(hf);
 331                         CloseHandle(hm);
 332                         return CM_ERROR_INVAL;
 333                 }
 334                 CloseHandle(hm);
 335 #else
 336                 /* djgpp doesn't support memory mapped files */
 337                 data = malloc(buf_nbuffers * buf_bufferSize);
 338 #endif /* !DJGPP */
 339
 340                 /* create buffer headers and put in free list */
 341                 bp = malloc(buf_nbuffers * sizeof(cm_buf_t));
 342                 buf_allp = NULL;
 343                 for(i=0; i<buf_nbuffers; i++) {
 344                         /* allocate and zero some storage */
 345                         memset(bp, 0, sizeof(cm_buf_t));
 346
 347                         /* thread on list of all buffers */
 348                         bp->allp = buf_allp;
 349                         buf_allp = bp;
 350
 351                         osi_QAdd((osi_queue_t **)&buf_freeListp, &bp->q);
 352                         bp->flags |= CM_BUF_INLRU;
 353                         lock_InitializeMutex(&bp->mx, "Buffer mutex");
 354
 355                         /* grab appropriate number of bytes from aligned zone */
 356                         bp->datap = data;
 357
 358                         /* setup last buffer pointer */
 359                         if (i == 0)
 360                                 buf_freeListEndp = bp;
 361
 362                         /* next */
 363                         bp++;
 364                         data += buf_bufferSize;
 365                 }
 366
 367                 /* none reserved at first */
 368                 buf_reservedBufs = 0;
 369
 370                 /* just for safety's sake */
 371                 buf_maxReservedBufs = buf_nbuffers - 3;
 372
 373                 /* init the buffer trace log */
 374                 buf_logp = osi_LogCreate("buffer", 10);
 375
 376                 osi_EndOnce(&once);
 377
 378                 /* and create the incr-syncer */
 379                 phandle = thrd_Create(0, 0,
 380                                       (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
 381                                       "buf_IncrSyncer");
 382
 383                 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
 384 #ifndef DJGPP
 385                 CloseHandle(phandle);
 386 #endif /* !DJGPP */
 387         }
 388
 389         return 0;
 390 }
 391
 392 /* add nbuffers to the buffer pool, if possible.
 393  * Called with no locks held.
 394  */
 395 long buf_AddBuffers(long nbuffers)
 396 {
 397         cm_buf_t *bp;
 398         int i;
 399         char *data;
 400 #ifndef DJGPP
 401         HANDLE hm;
 402         long cs;
 403
 404     afsi_log("%d buffers being added to the existing cache of size %d",
 405               nbuffers, buf_nbuffers);
 406
 407         /*
 408          * Cache file mapping constrained by
 409          * system allocation granularity;
 410          * round up, assuming granularity is a power of two;
 411          * assume existing cache size is already rounded
 412          */
 413         cs = nbuffers * buf_bufferSize;
 414         cs = (cs + (sysInfo.dwAllocationGranularity - 1))
 415                 & ~(sysInfo.dwAllocationGranularity - 1);
 416         if (cs != nbuffers * buf_bufferSize) {
 417                 nbuffers = cs / buf_bufferSize;
 418         }
 419
 420         /* Reserve additional buffer space by remapping cache file */
 421         hm = CreateFileMapping(CacheHandle,
 422                 NULL,
 423                 PAGE_READWRITE,
 424                 0, (buf_nbuffers + nbuffers) * buf_bufferSize,
 425                 NULL);
 426         if (hm == NULL) {
 427                 if (GetLastError() == ERROR_DISK_FULL)
 428                         return CM_ERROR_TOOMANYBUFS;
 429                 else
 430                         return CM_ERROR_INVAL;
 431         }
 432         data = MapViewOfFile(hm,
 433                 FILE_MAP_ALL_ACCESS,
 434                 0, buf_nbuffers * buf_bufferSize,
 435                 nbuffers * buf_bufferSize);
 436         if (data == NULL) {
 437                 CloseHandle(hm);
 438                 return CM_ERROR_INVAL;
 439         }
 440         CloseHandle(hm);
 441 #else
 442         data = malloc(buf_nbuffers * buf_bufferSize);
 443 #endif /* DJGPP */
 444
 445         /* Create buffer headers and put in free list */
 446         bp = malloc(nbuffers * sizeof(*bp));
 447
 448         for(i=0; i<nbuffers; i++) {
 449                 memset(bp, 0, sizeof(*bp));
 450
 451                 lock_InitializeMutex(&bp->mx, "cm_buf_t");
 452
 453                 /* grab appropriate number of bytes from aligned zone */
 454                 bp->datap = data;
 455
 456                 bp->flags |= CM_BUF_INLRU;
 457
 458                 lock_ObtainWrite(&buf_globalLock);
 459                 /* note that buf_allp chain is covered by buf_globalLock now */
 460                 bp->allp = buf_allp;
 461                 buf_allp = bp;
 462                 osi_QAdd((osi_queue_t **) &buf_freeListp, &bp->q);
 463                 if (!buf_freeListEndp) buf_freeListEndp = bp;
 464                 buf_nbuffers++;
 465                 lock_ReleaseWrite(&buf_globalLock);
 466
 467                 bp++;
 468                 data += buf_bufferSize;
 469
 470         }        /* for loop over all buffers */
 471
 472         return 0;
 473 }
 474
 475 /* interface to set the number of buffers to an exact figure.
 476  * Called with no locks held.
 477  */
 478 long buf_SetNBuffers(long nbuffers)
 479 {
 480         if (nbuffers < 10) return CM_ERROR_INVAL;
 481         if (nbuffers == buf_nbuffers) return 0;
 482         else if (nbuffers > buf_nbuffers)
 483                 return buf_AddBuffers(nbuffers - buf_nbuffers);
 484         else return CM_ERROR_INVAL;
 485 }
 486
 487 /* release a buffer.  Buffer must be referenced, but unlocked. */
 488 void buf_Release(cm_buf_t *bp)
 489 {
 490         lock_ObtainWrite(&buf_globalLock);
 491         buf_LockedRelease(bp);
 492         lock_ReleaseWrite(&buf_globalLock);
 493 }
 494
 495 /* wait for reading or writing to clear; called with write-locked
 496  * buffer, and returns with locked buffer.
 497  */
 498 void buf_WaitIO(cm_buf_t *bp)
 499 {
 500         while (1) {
 501                 /* if no IO is happening, we're done */
 502                 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
 503                         break;
 504
 505                 /* otherwise I/O is happening, but some other thread is waiting for
 506                  * the I/O already.  Wait for that guy to figure out what happened,
 507                  * and then check again.
 508                  */
 509                 bp->flags |= CM_BUF_WAITING;
 510                 osi_SleepM((long) bp, &bp->mx);
 511                 lock_ObtainMutex(&bp->mx);
 512                 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%x", bp);
 513         }
 514
 515         /* if we get here, the IO is done, but we may have to wakeup people waiting for
 516          * the I/O to complete.  Do so.
 517          */
 518         if (bp->flags & CM_BUF_WAITING) {
 519                 bp->flags &= ~CM_BUF_WAITING;
 520                 osi_Wakeup((long) bp);
 521         }
 522         osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%x", (long) bp);
 523 }
 524
 525 /* code to drop reference count while holding buf_globalLock */
 526 void buf_LockedRelease(cm_buf_t *bp)
 527 {
 528         /* ensure that we're in the LRU queue if our ref count is 0 */
 529         osi_assert(bp->refCount > 0);
 530         if (--bp->refCount == 0) {
 531                 if (!(bp->flags & CM_BUF_INLRU)) {
 532                         osi_QAdd((osi_queue_t **) &buf_freeListp, &bp->q);
 533
 534                         /* watch for transition from empty to one element */
 535                         if (!buf_freeListEndp)
 536                                 buf_freeListEndp = buf_freeListp;
 537                         bp->flags |= CM_BUF_INLRU;
 538                 }
 539         }
 540 }
 541
 542 /* find a buffer, if any, for a particular file ID and offset.  Assumes
 543  * that buf_globalLock is write locked when called.
 544  */
 545 cm_buf_t *buf_LockedFind(struct cm_scache *scp, osi_hyper_t *offsetp)
 546 {
 547         long i;
 548         cm_buf_t *bp;
 549
 550         i = BUF_HASH(&scp->fid, offsetp);
 551         for(bp = buf_hashTablepp[i]; bp; bp=bp->hashp) {
 552                 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
 553                         && offsetp->LowPart == bp->offset.LowPart
 554                         && offsetp->HighPart == bp->offset.HighPart) {
 555                         bp->refCount++;
 556                         break;
 557                 }
 558         }
 559
 560         /* return whatever we found, if anything */
 561         return bp;
 562 }
 563
 564 /* find a buffer with offset *offsetp for vnode *scp.  Called
 565  * with no locks held.
 566  */
 567 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
 568 {
 569         cm_buf_t *bp;
 570
 571         lock_ObtainWrite(&buf_globalLock);
 572         bp = buf_LockedFind(scp, offsetp);
 573         lock_ReleaseWrite(&buf_globalLock);
 574
 575         return bp;
 576 }
 577
 578 /* start cleaning I/O on this buffer.  Buffer must be write locked, and is returned
 579  * write-locked.
 580  *
 581  * Makes sure that there's only one person writing this block
 582  * at any given time, and also ensures that the log is forced sufficiently far,
 583  * if this buffer contains logged data.
 584  */
 585 void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
 586 {
 587         long code;
 588
 589         code = 0;
 590         while ((bp->flags & (CM_BUF_WRITING | CM_BUF_DIRTY)) == CM_BUF_DIRTY) {
 591                 lock_ReleaseMutex(&bp->mx);
 592
 593                 code = (*cm_buf_opsp->Writep)(&bp->fid, &bp->offset,
 594                                                 buf_bufferSize, 0, bp->userp,
 595                                                 reqp);
 596
 597                 lock_ObtainMutex(&bp->mx);
 598                 if (code) break;
 599
 600 #ifdef DISKCACHE95
 601                 /* Disk cache support */
 602                 /* write buffer to disk cache (synchronous for now) */
 603                 diskcache_Update(bp->dcp, bp->datap, buf_bufferSize, bp->dataVersion);
 604 #endif /* DISKCACHE95 */
 605         };
 606
 607         /* do logging after call to GetLastError, or else */
 608         osi_Log2(buf_logp, "buf_CleanAsync starts I/O on 0x%x, done=%d", bp, code);
 609
 610         /* if someone was waiting for the I/O that just completed or failed,
 611          * wake them up.
 612          */
 613         if (bp->flags & CM_BUF_WAITING) {
 614                 /* turn off flags and wakeup users */
 615                 bp->flags &= ~CM_BUF_WAITING;
 616                 osi_Wakeup((long) bp);
 617         }
 618 }
 619
 620 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
 621  * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
 622  * The buffer must already be clean, and no I/O should be happening to it.
 623  */
 624 void buf_Recycle(cm_buf_t *bp)
 625 {
 626         int i;
 627         cm_buf_t **lbpp;
 628         cm_buf_t *tbp;
 629         cm_buf_t *prevBp, *nextBp;
 630
 631         /* if we get here, we know that the buffer still has a 0 ref count,
 632          * and that it is clean and has no currently pending I/O.  This is
 633          * the dude to return.
 634          * Remember that as long as the ref count is 0, we know that we won't
 635          * have any lock conflicts, so we can grab the buffer lock out of
 636          * order in the locking hierarchy.
 637          */
 638         osi_Log2(buf_logp,
 639                 "buf_Recycle recycles 0x%x, off 0x%x",
 640                 bp, bp->offset.LowPart);
 641
 642         osi_assert(bp->refCount == 0);
 643         osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)));
 644         lock_AssertWrite(&buf_globalLock);
 645
 646         if (bp->flags & CM_BUF_INHASH) {
 647                 /* Remove from hash */
 648
 649                 i = BUF_HASH(&bp->fid, &bp->offset);
 650                 lbpp = &(buf_hashTablepp[i]);
 651                 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
 652                         if (tbp == bp) break;
 653                 }
 654
 655                 /* we better find it */
 656                 osi_assertx(tbp != NULL, "buf_GetNewLocked: hash table screwup");
 657
 658                 *lbpp = bp->hashp;      /* hash out */
 659
 660                 /* Remove from file hash */
 661
 662                 i = BUF_FILEHASH(&bp->fid);
 663                 prevBp = bp->fileHashBackp;
 664                 nextBp = bp->fileHashp;
 665                 if (prevBp)
 666                         prevBp->fileHashp = nextBp;
 667                 else
 668                         buf_fileHashTablepp[i] = nextBp;
 669                 if (nextBp)
 670                         nextBp->fileHashBackp = prevBp;
 671
 672                 bp->flags &= ~CM_BUF_INHASH;
 673         }
 674
 675         /* bump the soft reference counter now, to invalidate softRefs; no
 676          * wakeup is required since people don't sleep waiting for this
 677          * counter to change.
 678          */
 679         bp->idCounter++;
 680
 681         /* make the fid unrecognizable */
 682         memset(&bp->fid, 0, sizeof(bp->fid));
 683 }
 684
 685 /* recycle a buffer, removing it from the free list, hashing in its new identity
 686  * and returning it write-locked so that no one can use it.  Called without
 687  * any locks held, and can return an error if it loses the race condition and
 688  * finds that someone else created the desired buffer.
 689  *
 690  * If success is returned, the buffer is returned write-locked.
 691  *
 692  * May be called with null scp and offsetp, if we're just trying to reclaim some
 693  * space from the buffer pool.  In that case, the buffer will be returned
 694  * without being hashed into the hash table.
 695  */
 696 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 697 {
 698         cm_buf_t *bp;           /* buffer we're dealing with */
 699         cm_buf_t *nextBp;       /* next buffer in file hash chain */
 700         long i;                 /* temp */
 701         cm_req_t req;
 702
 703         cm_InitReq(&req);       /* just in case */
 704
 705         while(1) {
 706 retry:
 707                 lock_ObtainWrite(&buf_globalLock);
 708                 /* check to see if we lost the race */
 709                 if (scp) {
 710                         if (bp = buf_LockedFind(scp, offsetp)) {
 711                                 bp->refCount--;
 712                                 lock_ReleaseWrite(&buf_globalLock);
 713                                 return CM_BUF_EXISTS;
 714                         }
 715                 }
 716
 717                 /* for debugging, assert free list isn't empty, although we
 718                  * really should try waiting for a running tranasction to finish
 719                  * instead of this; or better, we should have a transaction
 720                  * throttler prevent us from entering this situation.
 721                  */
 722                 osi_assertx(buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
 723
 724                 /* look at all buffers in free list, some of which may temp.
 725                  * have high refcounts and which then should be skipped,
 726                  * starting cleaning I/O for those which are dirty.  If we find
 727                  * a clean buffer, we rehash it, lock it and return it.
 728                  */
 729                 for(bp = buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
 730                         /* check to see if it really has zero ref count.  This
 731                          * code can bump refcounts, at least, so it may not be
 732                          * zero.
 733                          */
 734                         if (bp->refCount > 0) continue;
 735
 736                         /* we don't have to lock buffer itself, since the ref
 737                          * count is 0 and we know it will stay zero as long as
 738                          * we hold the global lock.
 739                          */
 740
 741                         /* don't recycle someone in our own chunk */
 742                         if (!cm_FidCmp(&bp->fid, &scp->fid)
 743                             && (bp->offset.LowPart & (-cm_chunkSize))
 744                                   == (offsetp->LowPart & (-cm_chunkSize)))
 745                                 continue;
 746
 747                         /* if this page is being filled (!) or cleaned, see if
 748                          * the I/O has completed.  If not, skip it, otherwise
 749                          * do the final processing for the I/O.
 750                          */
 751                         if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
 752                                 /* probably shouldn't do this much work while
 753                                  * holding the big lock?  Watch for contention
 754                                  * here.
 755                                  */
 756                                 continue;
 757                         }
 758
 759                         if (bp->flags & CM_BUF_DIRTY) {
 760                                 /* if the buffer is dirty, start cleaning it and
 761                                  * move on to the next buffer.  We do this with
 762                                  * just the lock required to minimize contention
 763                                  * on the big lock.
 764                                  */
 765                                 bp->refCount++;
 766                                 lock_ReleaseWrite(&buf_globalLock);
 767
 768                                 /* grab required lock and clean; this only
 769                                  * starts the I/O.  By the time we're back,
 770                                  * it'll still be marked dirty, but it will also
 771                                  * have the WRITING flag set, so we won't get
 772                                  * back here.
 773                                  */
 774                                 buf_CleanAsync(bp, &req);
 775
 776                                 /* now put it back and go around again */
 777                                 buf_Release(bp);
 778                                 goto retry;
 779                         }
 780
 781                         /* if we get here, we know that the buffer still has a 0
 782                          * ref count, and that it is clean and has no currently
 783                          * pending I/O.  This is the dude to return.
 784                          * Remember that as long as the ref count is 0, we know
 785                          * that we won't have any lock conflicts, so we can grab
 786                          * the buffer lock out of order in the locking hierarchy.
 787                          */
 788                         buf_Recycle(bp);
 789
 790                         /* clean up junk flags */
 791                         bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
 792                         bp->dataVersion = -1;   /* unknown so far */
 793
 794                         /* now hash in as our new buffer, and give it the
 795                          * appropriate label, if requested.
 796                          */
 797                         if (scp) {
 798                                 bp->flags |= CM_BUF_INHASH;
 799                                 bp->fid = scp->fid;
 800                                 bp->offset = *offsetp;
 801                                 i = BUF_HASH(&scp->fid, offsetp);
 802                                 bp->hashp = buf_hashTablepp[i];
 803                                 buf_hashTablepp[i] = bp;
 804                                 i = BUF_FILEHASH(&scp->fid);
 805                                 nextBp = buf_fileHashTablepp[i];
 806                                 bp->fileHashp = nextBp;
 807                                 bp->fileHashBackp = NULL;
 808                                 if (nextBp)
 809                                         nextBp->fileHashBackp = bp;
 810                                 buf_fileHashTablepp[i] = bp;
 811                         }
 812
 813                         /* prepare to return it.  Start by giving it a good
 814                          * refcount */
 815                         bp->refCount = 1;
 816
 817                         /* and since it has a non-zero ref count, we should move
 818                          * it from the lru queue.  It better be still there,
 819                          * since we've held the global (big) lock since we found
 820                          * it there.
 821                          */
 822                         osi_assertx(bp->flags & CM_BUF_INLRU,
 823                                     "buf_GetNewLocked: LRU screwup");
 824                         if (buf_freeListEndp == bp) {
 825                                 /* we're the last guy in this queue, so maintain it */
 826                                 buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
 827                         }
 828                         osi_QRemove((osi_queue_t **) &buf_freeListp, &bp->q);
 829                         bp->flags &= ~CM_BUF_INLRU;
 830
 831                         /* finally, grab the mutex so that people don't use it
 832                          * before the caller fills it with data.  Again, no one
 833                          * should have been able to get to this dude to lock it.
 834                          */
 835                         osi_assertx(lock_TryMutex(&bp->mx),
 836                                     "buf_GetNewLocked: TryMutex failed");
 837
 838                         lock_ReleaseWrite(&buf_globalLock);
 839                         *bufpp = bp;
 840                         return 0;
 841                 } /* for all buffers in lru queue */
 842                 lock_ReleaseWrite(&buf_globalLock);
 843         }       /* while loop over everything */
 844         /* not reached */
 845 } /* the proc */
 846
 847 /* get a page, returning it held but unlocked.  Doesn't fill in the page
 848  * with I/O, since we're going to write the whole thing new.
 849  */
 850 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 851 {
 852         cm_buf_t *bp;
 853         long code;
 854         osi_hyper_t pageOffset;
 855         int created;
 856
 857         created = 0;
 858         pageOffset.HighPart = offsetp->HighPart;
 859         pageOffset.LowPart = offsetp->LowPart & ~(buf_bufferSize-1);
 860         while (1) {
 861                 lock_ObtainWrite(&buf_globalLock);
 862                 bp = buf_LockedFind(scp, &pageOffset);
 863                 lock_ReleaseWrite(&buf_globalLock);
 864                 if (bp) {
 865                         /* lock it and break out */
 866                         lock_ObtainMutex(&bp->mx);
 867                         break;
 868                 }
 869
 870                 /* otherwise, we have to create a page */
 871                 code = buf_GetNewLocked(scp, &pageOffset, &bp);
 872
 873                 /* check if the buffer was created in a race condition branch.
 874                  * If so, go around so we can hold a reference to it.
 875                  */
 876                 if (code == CM_BUF_EXISTS) continue;
 877
 878                 /* something else went wrong */
 879                 if (code != 0) return code;
 880
 881                 /* otherwise, we have a locked buffer that we just created */
 882                 created = 1;
 883                 break;
 884         } /* big while loop */
 885
 886         /* wait for reads */
 887         if (bp->flags & CM_BUF_READING)
 888                 buf_WaitIO(bp);
 889
 890         /* once it has been read once, we can unlock it and return it, still
 891          * with its refcount held.
 892          */
 893         lock_ReleaseMutex(&bp->mx);
 894         *bufpp = bp;
 895         osi_Log3(buf_logp, "buf_GetNew returning bp 0x%x for file 0x%x, offset 0x%x",
 896                 bp, (long) scp, offsetp->LowPart);
 897         return 0;
 898 }
 899
 900 /* get a page, returning it held but unlocked.  Make sure it is complete */
 901 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 902 {
 903         cm_buf_t *bp;
 904         long code;
 905         osi_hyper_t pageOffset;
 906         unsigned long tcount;
 907         int created;
 908 #ifdef DISKCACHE95
 909         cm_diskcache_t *dcp;
 910 #endif /* DISKCACHE95 */
 911
 912         created = 0;
 913         pageOffset.HighPart = offsetp->HighPart;
 914         pageOffset.LowPart = offsetp->LowPart & ~(buf_bufferSize-1);
 915         while (1) {
 916                 lock_ObtainWrite(&buf_globalLock);
 917                 bp = buf_LockedFind(scp, &pageOffset);
 918                 lock_ReleaseWrite(&buf_globalLock);
 919                 if (bp) {
 920                         /* lock it and break out */
 921                         lock_ObtainMutex(&bp->mx);
 922                         break;
 923
 924 #ifdef DISKCACHE95
 925                         /* touch disk chunk to update LRU info */
 926                         diskcache_Touch(bp->dcp);
 927 #endif /* DISKCACHE95 */
 928                 }
 929
 930                 /* otherwise, we have to create a page */
 931                 code = buf_GetNewLocked(scp, &pageOffset, &bp);
 932
 933                 /* check if the buffer was created in a race condition branch.
 934                  * If so, go around so we can hold a reference to it.
 935                  */
 936                 if (code == CM_BUF_EXISTS) continue;
 937
 938                 /* something else went wrong */
 939                 if (code != 0) return code;
 940
 941                 /* otherwise, we have a locked buffer that we just created */
 942                 created = 1;
 943                 break;
 944         } /* big while loop */
 945
 946         /* if we get here, we have a locked buffer that may have just been
 947          * created, in which case it needs to be filled with data.
 948          */
 949         if (created) {
 950                 /* load the page; freshly created pages should be idle */
 951                 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)));
 952
 953                 /* setup offset, event */
 954 #ifndef DJGPP  /* doesn't seem to be used */
 955                 bp->over.Offset = bp->offset.LowPart;
 956                 bp->over.OffsetHigh = bp->offset.HighPart;
 957 #endif /* !DJGPP */
 958
 959                 /* start the I/O; may drop lock */
 960                 bp->flags |= CM_BUF_READING;
 961                 code = (*cm_buf_opsp->Readp)(bp, buf_bufferSize, &tcount, NULL);
 962
 963 #ifdef DISKCACHE95
 964                 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, buf_bufferSize, &bp->dataVersion, &tcount, &dcp);
 965                 bp->dcp = dcp;    /* pointer to disk cache struct. */
 966 #endif /* DISKCACHE95 */
 967
 968                 if (code != 0) {
 969                         /* failure or queued */
 970 #ifndef DJGPP   /* cm_bufRead always returns 0 */
 971                         if (code != ERROR_IO_PENDING) {
 972 #endif
 973                                 bp->error = code;
 974                                 bp->flags |= CM_BUF_ERROR;
 975                                 bp->flags &= ~CM_BUF_READING;
 976                                 if (bp->flags & CM_BUF_WAITING) {
 977                                         bp->flags &= ~CM_BUF_WAITING;
 978                                         osi_Wakeup((long) bp);
 979                                 }
 980                                 lock_ReleaseMutex(&bp->mx);
 981                                 buf_Release(bp);
 982                                 return code;
 983 #ifndef DJGPP
 984                         }
 985 #endif
 986                 } else {
 987                         /* otherwise, I/O completed instantly and we're done, except
 988                          * for padding the xfr out with 0s and checking for EOF
 989                          */
 990                         if (tcount < (unsigned long) buf_bufferSize) {
 991                                 memset(bp->datap+tcount, 0, buf_bufferSize - tcount);
 992                                 if (tcount == 0)
 993                                         bp->flags |= CM_BUF_EOF;
 994                         }
 995                         bp->flags &= ~CM_BUF_READING;
 996                         if (bp->flags & CM_BUF_WAITING) {
 997                                 bp->flags &= ~CM_BUF_WAITING;
 998                                 osi_Wakeup((long) bp);
 999                         }
1000                 }
1001
1002         } /* if created */
1003
1004         /* wait for reads, either that which we started above, or that someone
1005          * else started.  We don't care if we return a buffer being cleaned.
1006          */
1007         if (bp->flags & CM_BUF_READING)
1008                 buf_WaitIO(bp);
1009
1010         /* once it has been read once, we can unlock it and return it, still
1011          * with its refcount held.
1012          */
1013         lock_ReleaseMutex(&bp->mx);
1014         *bufpp = bp;
1015
1016         /* now remove from queue; will be put in at the head (farthest from
1017          * being recycled) when we're done in buf_Release.
1018          */
1019         lock_ObtainWrite(&buf_globalLock);
1020         if (bp->flags & CM_BUF_INLRU) {
1021                 if (buf_freeListEndp == bp)
1022                         buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1023                 osi_QRemove((osi_queue_t **) &buf_freeListp, &bp->q);
1024                 bp->flags &= ~CM_BUF_INLRU;
1025         }
1026         lock_ReleaseWrite(&buf_globalLock);
1027
1028         osi_Log3(buf_logp, "buf_Get returning bp 0x%x for file 0x%x, offset 0x%x",
1029                 bp, (long) scp, offsetp->LowPart);
1030         return 0;
1031 }
1032
1033 /* count # of elements in the free list;
1034  * we don't bother doing the proper locking for accessing dataVersion or flags
1035  * since it is a pain, and this is really just an advisory call.  If you need
1036  * to do better at some point, rewrite this function.
1037  */
1038 long buf_CountFreeList(void)
1039 {
1040         long count;
1041         cm_buf_t *bufp;
1042
1043         count = 0;
1044         lock_ObtainRead(&buf_globalLock);
1045         for(bufp = buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1046                 /* if the buffer doesn't have an identity, or if the buffer
1047                  * has been invalidate (by having its DV stomped upon), then
1048                  * count it as free, since it isn't really being utilized.
1049                  */
1050                 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion <= 0)
1051                         count++;
1052         }
1053         lock_ReleaseRead(&buf_globalLock);
1054         return count;
1055 }
1056
1057 /* clean a buffer synchronously */
1058 void buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
1059 {
1060         lock_ObtainMutex(&bp->mx);
1061         buf_LockedCleanAsync(bp, reqp);
1062         lock_ReleaseMutex(&bp->mx);
1063 }
1064
1065 /* wait for a buffer's cleaning to finish */
1066 void buf_CleanWait(cm_buf_t *bp)
1067 {
1068         lock_ObtainMutex(&bp->mx);
1069         if (bp->flags & CM_BUF_WRITING) {
1070                 buf_WaitIO(bp);
1071         }
1072         lock_ReleaseMutex(&bp->mx);
1073 }
1074
1075 /* set the dirty flag on a buffer, and set associated write-ahead log,
1076  * if there is one.  Allow one to be added to a buffer, but not changed.
1077  *
1078  * The buffer must be locked before calling this routine.
1079  */
1080 void buf_SetDirty(cm_buf_t *bp)
1081 {
1082         osi_assert(bp->refCount > 0);
1083
1084         osi_Log1(buf_logp, "buf_SetDirty 0x%x", bp);
1085
1086         /* set dirty bit */
1087         bp->flags |= CM_BUF_DIRTY;
1088
1089         /* and turn off EOF flag, since it has associated data now */
1090         bp->flags &= ~CM_BUF_EOF;
1091 }
1092
1093 /* clean all buffers, reset log pointers and invalidate all buffers.
1094  * Called with no locks held, and returns with same.
1095  *
1096  * This function is guaranteed to clean and remove the log ptr of all the
1097  * buffers that were dirty or had non-zero log ptrs before the call was
1098  * made.  That's sufficient to clean up any garbage left around by recovery,
1099  * which is all we're counting on this for; there may be newly created buffers
1100  * added while we're running, but that should be OK.
1101  *
1102  * In an environment where there are no transactions (artificially imposed, for
1103  * example, when switching the database to raw mode), this function is used to
1104  * make sure that all updates have been written to the disk.  In that case, we don't
1105  * really require that we forget the log association between pages and logs, but
1106  * it also doesn't hurt.  Since raw mode I/O goes through this buffer package, we don't
1107  * have to worry about invalidating data in the buffers.
1108  *
1109  * This function is used at the end of recovery as paranoia to get the recovered
1110  * database out to disk.  It removes all references to the recovery log and cleans
1111  * all buffers.
1112  */
1113 long buf_CleanAndReset(void)
1114 {
1115         long i;
1116         cm_buf_t *bp;
1117         cm_req_t req;
1118
1119         lock_ObtainWrite(&buf_globalLock);
1120         for(i=0; i<buf_hashSize; i++) {
1121                 for(bp = buf_hashTablepp[i]; bp; bp = bp->hashp) {
1122                         bp->refCount++;
1123                         lock_ReleaseWrite(&buf_globalLock);
1124
1125                         /* now no locks are held; clean buffer and go on */
1126                         cm_InitReq(&req);
1127                         buf_CleanAsync(bp, &req);
1128                         buf_CleanWait(bp);
1129
1130                         /* relock and release buffer */
1131                         lock_ObtainWrite(&buf_globalLock);
1132                         buf_LockedRelease(bp);
1133                 } /* over one bucket */
1134         }       /* for loop over all hash buckets */
1135
1136         /* release locks */
1137         lock_ReleaseWrite(&buf_globalLock);
1138
1139         /* and we're done */
1140         return 0;
1141 }
1142
1143 /* called without global lock being held, reserves buffers for callers
1144  * that need more than one held (not locked) at once.
1145  */
1146 void buf_ReserveBuffers(long nbuffers)
1147 {
1148         lock_ObtainWrite(&buf_globalLock);
1149         while (1) {
1150                 if (buf_reservedBufs + nbuffers > buf_maxReservedBufs) {
1151                         buf_reserveWaiting = 1;
1152                         osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1153                         osi_SleepW((long) &buf_reservedBufs, &buf_globalLock);
1154                         lock_ObtainWrite(&buf_globalLock);
1155                 }
1156                 else {
1157                         buf_reservedBufs += nbuffers;
1158                         break;
1159                 }
1160         }
1161         lock_ReleaseWrite(&buf_globalLock);
1162 }
1163
1164 int buf_TryReserveBuffers(long nbuffers)
1165 {
1166         int code;
1167
1168         lock_ObtainWrite(&buf_globalLock);
1169         if (buf_reservedBufs + nbuffers > buf_maxReservedBufs) {
1170                 code = 0;
1171         }
1172         else {
1173                 buf_reservedBufs += nbuffers;
1174                 code = 1;
1175         }
1176         lock_ReleaseWrite(&buf_globalLock);
1177         return code;
1178 }
1179
1180 /* called without global lock held, releases reservation held by
1181  * buf_ReserveBuffers.
1182  */
1183 void buf_UnreserveBuffers(long nbuffers)
1184 {
1185         lock_ObtainWrite(&buf_globalLock);
1186         buf_reservedBufs -= nbuffers;
1187         if (buf_reserveWaiting) {
1188                 buf_reserveWaiting = 0;
1189                 osi_Wakeup((long) &buf_reservedBufs);
1190         }
1191         lock_ReleaseWrite(&buf_globalLock);
1192 }
1193
1194 /* truncate the buffers past sizep, zeroing out the page, if we don't
1195  * end on a page boundary.
1196  *
1197  * Requires cm_bufCreateLock to be write locked.
1198  */
1199 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1200         osi_hyper_t *sizep)
1201 {
1202         cm_buf_t *bufp;
1203         cm_buf_t *nbufp;                        /* next buffer, if didRelease */
1204         osi_hyper_t bufEnd;
1205         long code;
1206         long bufferPos;
1207         int didRelease;
1208         long i;
1209
1210         /* assert that cm_bufCreateLock is held in write mode */
1211         lock_AssertWrite(&scp->bufCreateLock);
1212
1213         i = BUF_FILEHASH(&scp->fid);
1214
1215         lock_ObtainWrite(&buf_globalLock);
1216         bufp = buf_fileHashTablepp[i];
1217         if (bufp == NULL) {
1218                 lock_ReleaseWrite(&buf_globalLock);
1219                 return 0;
1220         }
1221
1222         bufp->refCount++;
1223         lock_ReleaseWrite(&buf_globalLock);
1224         for(; bufp; bufp = nbufp) {
1225                 didRelease = 0;
1226                 lock_ObtainMutex(&bufp->mx);
1227
1228                 bufEnd.HighPart = 0;
1229                 bufEnd.LowPart = buf_bufferSize;
1230                 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1231
1232                 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1233                         LargeIntegerLessThan(*sizep, bufEnd)) {
1234                         buf_WaitIO(bufp);
1235                 }
1236                 lock_ObtainMutex(&scp->mx);
1237
1238                 /* make sure we have a callback (so we have the right value for
1239                  * the length), and wait for it to be safe to do a truncate.
1240                  */
1241                 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1242                                  CM_SCACHESYNC_NEEDCALLBACK
1243                                  | CM_SCACHESYNC_GETSTATUS
1244                                  | CM_SCACHESYNC_SETSIZE
1245                                  | CM_SCACHESYNC_BUFLOCKED);
1246                 /* if we succeeded in our locking, and this applies to the right
1247                  * file, and the truncate request overlaps the buffer either
1248                  * totally or partially, then do something.
1249                  */
1250                 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1251                         && LargeIntegerLessThan(*sizep, bufEnd)) {
1252
1253                         lock_ObtainWrite(&buf_globalLock);
1254
1255                         /* destroy the buffer, turning off its dirty bit, if
1256                          * we're truncating the whole buffer.  Otherwise, set
1257                          * the dirty bit, and clear out the tail of the buffer
1258                          * if we just overlap some.
1259                          */
1260                         if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1261                                 /* truncating the entire page */
1262                                 bufp->flags &= ~CM_BUF_DIRTY;
1263                                 bufp->dataVersion = -1; /* known bad */
1264                                 bufp->dirtyCounter++;
1265                         }
1266                         else {
1267                                 /* don't set dirty, since dirty implies
1268                                  * currently up-to-date.  Don't need to do this,
1269                                  * since we'll update the length anyway.
1270                                  *
1271                                  * Zero out remainder of the page, in case we
1272                                  * seek and write past EOF, and make this data
1273                                  * visible again.
1274                                  */
1275                                 bufferPos = sizep->LowPart & (buf_bufferSize - 1);
1276                                 osi_assert(bufferPos != 0);
1277                                 memset(bufp->datap + bufferPos, 0,
1278                                         buf_bufferSize - bufferPos);
1279                         }
1280
1281                         lock_ReleaseWrite(&buf_globalLock);
1282
1283                 }
1284
1285                 lock_ReleaseMutex(&scp->mx);
1286                 lock_ReleaseMutex(&bufp->mx);
1287                 if (!didRelease) {
1288                         lock_ObtainWrite(&buf_globalLock);
1289                         nbufp = bufp->fileHashp;
1290                         if (nbufp) nbufp->refCount++;
1291                         buf_LockedRelease(bufp);
1292                         lock_ReleaseWrite(&buf_globalLock);
1293                 }
1294
1295                 /* bail out early if we fail */
1296                 if (code) {
1297                         /* at this point, nbufp is held; bufp has already been
1298                          * released.
1299                          */
1300                         if (nbufp) buf_Release(nbufp);
1301                         return code;
1302                 }
1303         }
1304
1305         /* success */
1306         return 0;
1307 }
1308
1309 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1310 {
1311         long code;
1312         cm_buf_t *bp;           /* buffer we're hacking on */
1313         cm_buf_t *nbp;
1314         int didRelease;
1315         long i;
1316
1317         i = BUF_FILEHASH(&scp->fid);
1318
1319         code = 0;
1320         lock_ObtainWrite(&buf_globalLock);
1321         bp = buf_fileHashTablepp[i];
1322         if (bp) bp->refCount++;
1323         lock_ReleaseWrite(&buf_globalLock);
1324         for(; bp; bp = nbp) {
1325                 didRelease = 0; /* haven't released this buffer yet */
1326
1327                 /* clean buffer synchronously */
1328                 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1329                         lock_ObtainMutex(&bp->mx);
1330
1331                         /* start cleaning the buffer, and wait for it to finish */
1332                         buf_LockedCleanAsync(bp, reqp);
1333                         buf_WaitIO(bp);
1334                         lock_ReleaseMutex(&bp->mx);
1335
1336                         code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1337                         if (code) goto skip;
1338
1339                         lock_ObtainWrite(&buf_globalLock);
1340                         /* actually, we only know that buffer is clean if ref
1341                          * count is 1, since we don't have buffer itself locked.
1342                          */
1343                         if (!(bp->flags & CM_BUF_DIRTY)) {
1344                                 if (bp->refCount == 1) {        /* bp is held above */
1345                                         buf_LockedRelease(bp);
1346                                         nbp = bp->fileHashp;
1347                                         if (nbp) nbp->refCount++;
1348                                         didRelease = 1;
1349                                         buf_Recycle(bp);
1350                                 }
1351                         }
1352                         lock_ReleaseWrite(&buf_globalLock);
1353
1354                         (*cm_buf_opsp->Unstabilizep)(scp, userp);
1355                 }
1356
1357 skip:
1358                 if (!didRelease) {
1359                         lock_ObtainWrite(&buf_globalLock);
1360                         if (nbp = bp->fileHashp) nbp->refCount++;
1361                         buf_LockedRelease(bp);
1362                         lock_ReleaseWrite(&buf_globalLock);
1363                 }
1364         }       /* for loop over a bunch of buffers */
1365
1366         /* done */
1367         return code;
1368 }
1369
1370 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1371 {
1372         long code;
1373         cm_buf_t *bp;           /* buffer we're hacking on */
1374     cm_buf_t *nbp;              /* next one */
1375         long i;
1376
1377         i = BUF_FILEHASH(&scp->fid);
1378
1379         code = 0;
1380         lock_ObtainWrite(&buf_globalLock);
1381     bp = buf_fileHashTablepp[i];
1382     if (bp) bp->refCount++;
1383     lock_ReleaseWrite(&buf_globalLock);
1384         for(; bp; bp = nbp) {
1385                 /* clean buffer synchronously */
1386                 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1387                         if (userp) {
1388                 cm_HoldUser(userp);
1389                                 lock_ObtainMutex(&bp->mx);
1390                                 if (bp->userp)
1391                     cm_ReleaseUser(bp->userp);
1392                 bp->userp = userp;
1393                                 lock_ReleaseMutex(&bp->mx);
1394             }
1395                         buf_CleanAsync(bp, reqp);
1396             buf_CleanWait(bp);
1397             lock_ObtainMutex(&bp->mx);
1398                         if (bp->flags & CM_BUF_ERROR) {
1399                                 if (code == 0 || code == -1) code = bp->error;
1400                 if (code == 0) code = -1;
1401             }
1402             lock_ReleaseMutex(&bp->mx);
1403                 }
1404
1405                 lock_ObtainWrite(&buf_globalLock);
1406                 buf_LockedRelease(bp);
1407         nbp = bp->fileHashp;
1408         if (nbp) nbp->refCount++;
1409                 lock_ReleaseWrite(&buf_globalLock);
1410         }       /* for loop over a bunch of buffers */
1411
1412     /* done */
1413         return code;
1414 }
1415
1416 /* dump the contents of the buf_hashTablepp. */
1417 int cm_DumpBufHashTable(FILE *outputFile, char *cookie)
1418 {
1419     int zilch;
1420     cm_buf_t *bp;
1421     char output[1024];
1422     int i;
1423
1424         if (buf_hashTablepp == NULL)
1425                 return -1;
1426
1427     lock_ObtainRead(&buf_globalLock);
1428
1429     sprintf(output, "%s - dumping buf_HashTable - buf_hashSize=%d\n", cookie, buf_hashSize);
1430     WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1431
1432     for (i = 0; i < buf_hashSize; i++)
1433     {
1434         for(bp = buf_hashTablepp[i]; bp; bp=bp->hashp)
1435         {
1436             if (bp->refCount)
1437             {
1438                 sprintf(output, "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d,"
1439                         "vnode=%d, unique=%d), size=%d refCount=%d\n",
1440                         cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1441                         bp->fid.vnode, bp->fid.unique, bp->size, bp->refCount);
1442                 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1443             }
1444         }
1445     }
1446
1447     sprintf(output, "%s - Done dumping buf_HashTable.\n", cookie);
1448     WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1449
1450     lock_ReleaseRead(&buf_globalLock);
1451     return 0;
1452 }
1453