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         if ( bp->flags & CM_BUF_WAITING )
 510             osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%x", bp);
 511
 512         bp->flags |= CM_BUF_WAITING;
 513         osi_SleepM((long) bp, &bp->mx);
 514         lock_ObtainMutex(&bp->mx);
 515                 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%x", bp);
 516     }
 517
 518     /* if we get here, the IO is done, but we may have to wakeup people waiting for
 519      * the I/O to complete.  Do so.
 520      */
 521     if (bp->flags & CM_BUF_WAITING) {
 522                 bp->flags &= ~CM_BUF_WAITING;
 523         osi_Wakeup((long) bp);
 524     }
 525     osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%x", (long) bp);
 526 }
 527
 528 /* code to drop reference count while holding buf_globalLock */
 529 void buf_LockedRelease(cm_buf_t *bp)
 530 {
 531         /* ensure that we're in the LRU queue if our ref count is 0 */
 532         osi_assert(bp->refCount > 0);
 533         if (--bp->refCount == 0) {
 534                 if (!(bp->flags & CM_BUF_INLRU)) {
 535                         osi_QAdd((osi_queue_t **) &buf_freeListp, &bp->q);
 536
 537                         /* watch for transition from empty to one element */
 538                         if (!buf_freeListEndp)
 539                                 buf_freeListEndp = buf_freeListp;
 540                         bp->flags |= CM_BUF_INLRU;
 541                 }
 542         }
 543 }
 544
 545 /* find a buffer, if any, for a particular file ID and offset.  Assumes
 546  * that buf_globalLock is write locked when called.
 547  */
 548 cm_buf_t *buf_LockedFind(struct cm_scache *scp, osi_hyper_t *offsetp)
 549 {
 550         long i;
 551         cm_buf_t *bp;
 552
 553         i = BUF_HASH(&scp->fid, offsetp);
 554         for(bp = buf_hashTablepp[i]; bp; bp=bp->hashp) {
 555                 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
 556                         && offsetp->LowPart == bp->offset.LowPart
 557                         && offsetp->HighPart == bp->offset.HighPart) {
 558                         bp->refCount++;
 559                         break;
 560                 }
 561         }
 562
 563         /* return whatever we found, if anything */
 564         return bp;
 565 }
 566
 567 /* find a buffer with offset *offsetp for vnode *scp.  Called
 568  * with no locks held.
 569  */
 570 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
 571 {
 572         cm_buf_t *bp;
 573
 574         lock_ObtainWrite(&buf_globalLock);
 575         bp = buf_LockedFind(scp, offsetp);
 576         lock_ReleaseWrite(&buf_globalLock);
 577
 578         return bp;
 579 }
 580
 581 /* start cleaning I/O on this buffer.  Buffer must be write locked, and is returned
 582  * write-locked.
 583  *
 584  * Makes sure that there's only one person writing this block
 585  * at any given time, and also ensures that the log is forced sufficiently far,
 586  * if this buffer contains logged data.
 587  */
 588 void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
 589 {
 590         long code;
 591
 592         code = 0;
 593         while ((bp->flags & (CM_BUF_WRITING | CM_BUF_DIRTY)) == CM_BUF_DIRTY) {
 594                 lock_ReleaseMutex(&bp->mx);
 595
 596                 code = (*cm_buf_opsp->Writep)(&bp->fid, &bp->offset,
 597                                                 buf_bufferSize, 0, bp->userp,
 598                                                 reqp);
 599
 600                 lock_ObtainMutex(&bp->mx);
 601                 if (code) break;
 602
 603 #ifdef DISKCACHE95
 604                 /* Disk cache support */
 605                 /* write buffer to disk cache (synchronous for now) */
 606                 diskcache_Update(bp->dcp, bp->datap, buf_bufferSize, bp->dataVersion);
 607 #endif /* DISKCACHE95 */
 608         };
 609
 610         /* do logging after call to GetLastError, or else */
 611         osi_Log2(buf_logp, "buf_CleanAsync starts I/O on 0x%x, done=%d", bp, code);
 612
 613         /* if someone was waiting for the I/O that just completed or failed,
 614          * wake them up.
 615          */
 616         if (bp->flags & CM_BUF_WAITING) {
 617                 /* turn off flags and wakeup users */
 618                 bp->flags &= ~CM_BUF_WAITING;
 619                 osi_Wakeup((long) bp);
 620         }
 621 }
 622
 623 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
 624  * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
 625  * The buffer must already be clean, and no I/O should be happening to it.
 626  */
 627 void buf_Recycle(cm_buf_t *bp)
 628 {
 629         int i;
 630         cm_buf_t **lbpp;
 631         cm_buf_t *tbp;
 632         cm_buf_t *prevBp, *nextBp;
 633
 634         /* if we get here, we know that the buffer still has a 0 ref count,
 635          * and that it is clean and has no currently pending I/O.  This is
 636          * the dude to return.
 637          * Remember that as long as the ref count is 0, we know that we won't
 638          * have any lock conflicts, so we can grab the buffer lock out of
 639          * order in the locking hierarchy.
 640          */
 641         osi_Log2(buf_logp,
 642                 "buf_Recycle recycles 0x%x, off 0x%x",
 643                 bp, bp->offset.LowPart);
 644
 645         osi_assert(bp->refCount == 0);
 646         osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)));
 647         lock_AssertWrite(&buf_globalLock);
 648
 649         if (bp->flags & CM_BUF_INHASH) {
 650                 /* Remove from hash */
 651
 652                 i = BUF_HASH(&bp->fid, &bp->offset);
 653                 lbpp = &(buf_hashTablepp[i]);
 654                 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
 655                         if (tbp == bp) break;
 656                 }
 657
 658                 /* we better find it */
 659                 osi_assertx(tbp != NULL, "buf_GetNewLocked: hash table screwup");
 660
 661                 *lbpp = bp->hashp;      /* hash out */
 662
 663                 /* Remove from file hash */
 664
 665                 i = BUF_FILEHASH(&bp->fid);
 666                 prevBp = bp->fileHashBackp;
 667                 nextBp = bp->fileHashp;
 668                 if (prevBp)
 669                         prevBp->fileHashp = nextBp;
 670                 else
 671                         buf_fileHashTablepp[i] = nextBp;
 672                 if (nextBp)
 673                         nextBp->fileHashBackp = prevBp;
 674
 675                 bp->flags &= ~CM_BUF_INHASH;
 676         }
 677
 678         /* bump the soft reference counter now, to invalidate softRefs; no
 679          * wakeup is required since people don't sleep waiting for this
 680          * counter to change.
 681          */
 682         bp->idCounter++;
 683
 684         /* make the fid unrecognizable */
 685         memset(&bp->fid, 0, sizeof(bp->fid));
 686 }
 687
 688 /* recycle a buffer, removing it from the free list, hashing in its new identity
 689  * and returning it write-locked so that no one can use it.  Called without
 690  * any locks held, and can return an error if it loses the race condition and
 691  * finds that someone else created the desired buffer.
 692  *
 693  * If success is returned, the buffer is returned write-locked.
 694  *
 695  * May be called with null scp and offsetp, if we're just trying to reclaim some
 696  * space from the buffer pool.  In that case, the buffer will be returned
 697  * without being hashed into the hash table.
 698  */
 699 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 700 {
 701         cm_buf_t *bp;           /* buffer we're dealing with */
 702         cm_buf_t *nextBp;       /* next buffer in file hash chain */
 703         long i;                 /* temp */
 704         cm_req_t req;
 705
 706         cm_InitReq(&req);       /* just in case */
 707
 708         while(1) {
 709 retry:
 710                 lock_ObtainWrite(&buf_globalLock);
 711                 /* check to see if we lost the race */
 712                 if (scp) {
 713                         if (bp = buf_LockedFind(scp, offsetp)) {
 714                                 bp->refCount--;
 715                                 lock_ReleaseWrite(&buf_globalLock);
 716                                 return CM_BUF_EXISTS;
 717                         }
 718                 }
 719
 720                 /* for debugging, assert free list isn't empty, although we
 721                  * really should try waiting for a running tranasction to finish
 722                  * instead of this; or better, we should have a transaction
 723                  * throttler prevent us from entering this situation.
 724                  */
 725                 osi_assertx(buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
 726
 727                 /* look at all buffers in free list, some of which may temp.
 728                  * have high refcounts and which then should be skipped,
 729                  * starting cleaning I/O for those which are dirty.  If we find
 730                  * a clean buffer, we rehash it, lock it and return it.
 731                  */
 732                 for(bp = buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
 733                         /* check to see if it really has zero ref count.  This
 734                          * code can bump refcounts, at least, so it may not be
 735                          * zero.
 736                          */
 737                         if (bp->refCount > 0) continue;
 738
 739                         /* we don't have to lock buffer itself, since the ref
 740                          * count is 0 and we know it will stay zero as long as
 741                          * we hold the global lock.
 742                          */
 743
 744                         /* don't recycle someone in our own chunk */
 745                         if (!cm_FidCmp(&bp->fid, &scp->fid)
 746                             && (bp->offset.LowPart & (-cm_chunkSize))
 747                                   == (offsetp->LowPart & (-cm_chunkSize)))
 748                                 continue;
 749
 750                         /* if this page is being filled (!) or cleaned, see if
 751                          * the I/O has completed.  If not, skip it, otherwise
 752                          * do the final processing for the I/O.
 753                          */
 754                         if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
 755                                 /* probably shouldn't do this much work while
 756                                  * holding the big lock?  Watch for contention
 757                                  * here.
 758                                  */
 759                                 continue;
 760                         }
 761
 762                         if (bp->flags & CM_BUF_DIRTY) {
 763                                 /* if the buffer is dirty, start cleaning it and
 764                                  * move on to the next buffer.  We do this with
 765                                  * just the lock required to minimize contention
 766                                  * on the big lock.
 767                                  */
 768                                 bp->refCount++;
 769                                 lock_ReleaseWrite(&buf_globalLock);
 770
 771                                 /* grab required lock and clean; this only
 772                                  * starts the I/O.  By the time we're back,
 773                                  * it'll still be marked dirty, but it will also
 774                                  * have the WRITING flag set, so we won't get
 775                                  * back here.
 776                                  */
 777                                 buf_CleanAsync(bp, &req);
 778
 779                                 /* now put it back and go around again */
 780                                 buf_Release(bp);
 781                                 goto retry;
 782                         }
 783
 784                         /* if we get here, we know that the buffer still has a 0
 785                          * ref count, and that it is clean and has no currently
 786                          * pending I/O.  This is the dude to return.
 787                          * Remember that as long as the ref count is 0, we know
 788                          * that we won't have any lock conflicts, so we can grab
 789                          * the buffer lock out of order in the locking hierarchy.
 790                          */
 791                         buf_Recycle(bp);
 792
 793                         /* clean up junk flags */
 794                         bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
 795                         bp->dataVersion = -1;   /* unknown so far */
 796
 797                         /* now hash in as our new buffer, and give it the
 798                          * appropriate label, if requested.
 799                          */
 800                         if (scp) {
 801                                 bp->flags |= CM_BUF_INHASH;
 802                                 bp->fid = scp->fid;
 803                                 bp->offset = *offsetp;
 804                                 i = BUF_HASH(&scp->fid, offsetp);
 805                                 bp->hashp = buf_hashTablepp[i];
 806                                 buf_hashTablepp[i] = bp;
 807                                 i = BUF_FILEHASH(&scp->fid);
 808                                 nextBp = buf_fileHashTablepp[i];
 809                                 bp->fileHashp = nextBp;
 810                                 bp->fileHashBackp = NULL;
 811                                 if (nextBp)
 812                                         nextBp->fileHashBackp = bp;
 813                                 buf_fileHashTablepp[i] = bp;
 814                         }
 815
 816                         /* prepare to return it.  Start by giving it a good
 817                          * refcount */
 818                         bp->refCount = 1;
 819
 820                         /* and since it has a non-zero ref count, we should move
 821                          * it from the lru queue.  It better be still there,
 822                          * since we've held the global (big) lock since we found
 823                          * it there.
 824                          */
 825                         osi_assertx(bp->flags & CM_BUF_INLRU,
 826                                     "buf_GetNewLocked: LRU screwup");
 827                         if (buf_freeListEndp == bp) {
 828                                 /* we're the last guy in this queue, so maintain it */
 829                                 buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
 830                         }
 831                         osi_QRemove((osi_queue_t **) &buf_freeListp, &bp->q);
 832                         bp->flags &= ~CM_BUF_INLRU;
 833
 834                         /* finally, grab the mutex so that people don't use it
 835                          * before the caller fills it with data.  Again, no one
 836                          * should have been able to get to this dude to lock it.
 837                          */
 838                         osi_assertx(lock_TryMutex(&bp->mx),
 839                                     "buf_GetNewLocked: TryMutex failed");
 840
 841                         lock_ReleaseWrite(&buf_globalLock);
 842                         *bufpp = bp;
 843                         return 0;
 844                 } /* for all buffers in lru queue */
 845                 lock_ReleaseWrite(&buf_globalLock);
 846         }       /* while loop over everything */
 847         /* not reached */
 848 } /* the proc */
 849
 850 /* get a page, returning it held but unlocked.  Doesn't fill in the page
 851  * with I/O, since we're going to write the whole thing new.
 852  */
 853 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 854 {
 855         cm_buf_t *bp;
 856         long code;
 857         osi_hyper_t pageOffset;
 858         int created;
 859
 860         created = 0;
 861         pageOffset.HighPart = offsetp->HighPart;
 862         pageOffset.LowPart = offsetp->LowPart & ~(buf_bufferSize-1);
 863         while (1) {
 864                 lock_ObtainWrite(&buf_globalLock);
 865                 bp = buf_LockedFind(scp, &pageOffset);
 866                 lock_ReleaseWrite(&buf_globalLock);
 867                 if (bp) {
 868                         /* lock it and break out */
 869                         lock_ObtainMutex(&bp->mx);
 870                         break;
 871                 }
 872
 873                 /* otherwise, we have to create a page */
 874                 code = buf_GetNewLocked(scp, &pageOffset, &bp);
 875
 876                 /* check if the buffer was created in a race condition branch.
 877                  * If so, go around so we can hold a reference to it.
 878                  */
 879                 if (code == CM_BUF_EXISTS) continue;
 880
 881                 /* something else went wrong */
 882                 if (code != 0) return code;
 883
 884                 /* otherwise, we have a locked buffer that we just created */
 885                 created = 1;
 886                 break;
 887         } /* big while loop */
 888
 889         /* wait for reads */
 890         if (bp->flags & CM_BUF_READING)
 891                 buf_WaitIO(bp);
 892
 893         /* once it has been read once, we can unlock it and return it, still
 894          * with its refcount held.
 895          */
 896         lock_ReleaseMutex(&bp->mx);
 897         *bufpp = bp;
 898         osi_Log3(buf_logp, "buf_GetNew returning bp 0x%x for file 0x%x, offset 0x%x",
 899                 bp, (long) scp, offsetp->LowPart);
 900         return 0;
 901 }
 902
 903 /* get a page, returning it held but unlocked.  Make sure it is complete */
 904 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
 905 {
 906         cm_buf_t *bp;
 907         long code;
 908         osi_hyper_t pageOffset;
 909         unsigned long tcount;
 910         int created;
 911 #ifdef DISKCACHE95
 912         cm_diskcache_t *dcp;
 913 #endif /* DISKCACHE95 */
 914
 915         created = 0;
 916         pageOffset.HighPart = offsetp->HighPart;
 917         pageOffset.LowPart = offsetp->LowPart & ~(buf_bufferSize-1);
 918         while (1) {
 919                 lock_ObtainWrite(&buf_globalLock);
 920                 bp = buf_LockedFind(scp, &pageOffset);
 921                 lock_ReleaseWrite(&buf_globalLock);
 922                 if (bp) {
 923                         /* lock it and break out */
 924                         lock_ObtainMutex(&bp->mx);
 925                         break;
 926
 927 #ifdef DISKCACHE95
 928                         /* touch disk chunk to update LRU info */
 929                         diskcache_Touch(bp->dcp);
 930 #endif /* DISKCACHE95 */
 931                 }
 932
 933                 /* otherwise, we have to create a page */
 934                 code = buf_GetNewLocked(scp, &pageOffset, &bp);
 935
 936                 /* check if the buffer was created in a race condition branch.
 937                  * If so, go around so we can hold a reference to it.
 938                  */
 939                 if (code == CM_BUF_EXISTS) continue;
 940
 941                 /* something else went wrong */
 942                 if (code != 0) return code;
 943
 944                 /* otherwise, we have a locked buffer that we just created */
 945                 created = 1;
 946                 break;
 947         } /* big while loop */
 948
 949         /* if we get here, we have a locked buffer that may have just been
 950          * created, in which case it needs to be filled with data.
 951          */
 952         if (created) {
 953                 /* load the page; freshly created pages should be idle */
 954                 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)));
 955
 956                 /* setup offset, event */
 957 #ifndef DJGPP  /* doesn't seem to be used */
 958                 bp->over.Offset = bp->offset.LowPart;
 959                 bp->over.OffsetHigh = bp->offset.HighPart;
 960 #endif /* !DJGPP */
 961
 962                 /* start the I/O; may drop lock */
 963                 bp->flags |= CM_BUF_READING;
 964                 code = (*cm_buf_opsp->Readp)(bp, buf_bufferSize, &tcount, NULL);
 965
 966 #ifdef DISKCACHE95
 967                 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, buf_bufferSize, &bp->dataVersion, &tcount, &dcp);
 968                 bp->dcp = dcp;    /* pointer to disk cache struct. */
 969 #endif /* DISKCACHE95 */
 970
 971                 if (code != 0) {
 972                         /* failure or queued */
 973 #ifndef DJGPP   /* cm_bufRead always returns 0 */
 974                         if (code != ERROR_IO_PENDING) {
 975 #endif
 976                                 bp->error = code;
 977                                 bp->flags |= CM_BUF_ERROR;
 978                                 bp->flags &= ~CM_BUF_READING;
 979                                 if (bp->flags & CM_BUF_WAITING) {
 980                                         bp->flags &= ~CM_BUF_WAITING;
 981                                         osi_Wakeup((long) bp);
 982                                 }
 983                                 lock_ReleaseMutex(&bp->mx);
 984                                 buf_Release(bp);
 985                                 return code;
 986 #ifndef DJGPP
 987                         }
 988 #endif
 989                 } else {
 990                         /* otherwise, I/O completed instantly and we're done, except
 991                          * for padding the xfr out with 0s and checking for EOF
 992                          */
 993                         if (tcount < (unsigned long) buf_bufferSize) {
 994                                 memset(bp->datap+tcount, 0, buf_bufferSize - tcount);
 995                                 if (tcount == 0)
 996                                         bp->flags |= CM_BUF_EOF;
 997                         }
 998                         bp->flags &= ~CM_BUF_READING;
 999                         if (bp->flags & CM_BUF_WAITING) {
1000                                 bp->flags &= ~CM_BUF_WAITING;
1001                                 osi_Wakeup((long) bp);
1002                         }
1003                 }
1004
1005         } /* if created */
1006
1007         /* wait for reads, either that which we started above, or that someone
1008          * else started.  We don't care if we return a buffer being cleaned.
1009          */
1010         if (bp->flags & CM_BUF_READING)
1011                 buf_WaitIO(bp);
1012
1013         /* once it has been read once, we can unlock it and return it, still
1014          * with its refcount held.
1015          */
1016         lock_ReleaseMutex(&bp->mx);
1017         *bufpp = bp;
1018
1019         /* now remove from queue; will be put in at the head (farthest from
1020          * being recycled) when we're done in buf_Release.
1021          */
1022         lock_ObtainWrite(&buf_globalLock);
1023         if (bp->flags & CM_BUF_INLRU) {
1024                 if (buf_freeListEndp == bp)
1025                         buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1026                 osi_QRemove((osi_queue_t **) &buf_freeListp, &bp->q);
1027                 bp->flags &= ~CM_BUF_INLRU;
1028         }
1029         lock_ReleaseWrite(&buf_globalLock);
1030
1031         osi_Log3(buf_logp, "buf_Get returning bp 0x%x for file 0x%x, offset 0x%x",
1032                 bp, (long) scp, offsetp->LowPart);
1033         return 0;
1034 }
1035
1036 /* count # of elements in the free list;
1037  * we don't bother doing the proper locking for accessing dataVersion or flags
1038  * since it is a pain, and this is really just an advisory call.  If you need
1039  * to do better at some point, rewrite this function.
1040  */
1041 long buf_CountFreeList(void)
1042 {
1043         long count;
1044         cm_buf_t *bufp;
1045
1046         count = 0;
1047         lock_ObtainRead(&buf_globalLock);
1048         for(bufp = buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1049                 /* if the buffer doesn't have an identity, or if the buffer
1050                  * has been invalidate (by having its DV stomped upon), then
1051                  * count it as free, since it isn't really being utilized.
1052                  */
1053                 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion <= 0)
1054                         count++;
1055         }
1056         lock_ReleaseRead(&buf_globalLock);
1057         return count;
1058 }
1059
1060 /* clean a buffer synchronously */
1061 void buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
1062 {
1063         lock_ObtainMutex(&bp->mx);
1064         buf_LockedCleanAsync(bp, reqp);
1065         lock_ReleaseMutex(&bp->mx);
1066 }
1067
1068 /* wait for a buffer's cleaning to finish */
1069 void buf_CleanWait(cm_buf_t *bp)
1070 {
1071         lock_ObtainMutex(&bp->mx);
1072         if (bp->flags & CM_BUF_WRITING) {
1073                 buf_WaitIO(bp);
1074         }
1075         lock_ReleaseMutex(&bp->mx);
1076 }
1077
1078 /* set the dirty flag on a buffer, and set associated write-ahead log,
1079  * if there is one.  Allow one to be added to a buffer, but not changed.
1080  *
1081  * The buffer must be locked before calling this routine.
1082  */
1083 void buf_SetDirty(cm_buf_t *bp)
1084 {
1085         osi_assert(bp->refCount > 0);
1086
1087         osi_Log1(buf_logp, "buf_SetDirty 0x%x", bp);
1088
1089         /* set dirty bit */
1090         bp->flags |= CM_BUF_DIRTY;
1091
1092         /* and turn off EOF flag, since it has associated data now */
1093         bp->flags &= ~CM_BUF_EOF;
1094 }
1095
1096 /* clean all buffers, reset log pointers and invalidate all buffers.
1097  * Called with no locks held, and returns with same.
1098  *
1099  * This function is guaranteed to clean and remove the log ptr of all the
1100  * buffers that were dirty or had non-zero log ptrs before the call was
1101  * made.  That's sufficient to clean up any garbage left around by recovery,
1102  * which is all we're counting on this for; there may be newly created buffers
1103  * added while we're running, but that should be OK.
1104  *
1105  * In an environment where there are no transactions (artificially imposed, for
1106  * example, when switching the database to raw mode), this function is used to
1107  * make sure that all updates have been written to the disk.  In that case, we don't
1108  * really require that we forget the log association between pages and logs, but
1109  * it also doesn't hurt.  Since raw mode I/O goes through this buffer package, we don't
1110  * have to worry about invalidating data in the buffers.
1111  *
1112  * This function is used at the end of recovery as paranoia to get the recovered
1113  * database out to disk.  It removes all references to the recovery log and cleans
1114  * all buffers.
1115  */
1116 long buf_CleanAndReset(void)
1117 {
1118         long i;
1119         cm_buf_t *bp;
1120         cm_req_t req;
1121
1122         lock_ObtainWrite(&buf_globalLock);
1123         for(i=0; i<buf_hashSize; i++) {
1124                 for(bp = buf_hashTablepp[i]; bp; bp = bp->hashp) {
1125                         bp->refCount++;
1126                         lock_ReleaseWrite(&buf_globalLock);
1127
1128                         /* now no locks are held; clean buffer and go on */
1129                         cm_InitReq(&req);
1130                         buf_CleanAsync(bp, &req);
1131                         buf_CleanWait(bp);
1132
1133                         /* relock and release buffer */
1134                         lock_ObtainWrite(&buf_globalLock);
1135                         buf_LockedRelease(bp);
1136                 } /* over one bucket */
1137         }       /* for loop over all hash buckets */
1138
1139         /* release locks */
1140         lock_ReleaseWrite(&buf_globalLock);
1141
1142         /* and we're done */
1143         return 0;
1144 }
1145
1146 /* called without global lock being held, reserves buffers for callers
1147  * that need more than one held (not locked) at once.
1148  */
1149 void buf_ReserveBuffers(long nbuffers)
1150 {
1151         lock_ObtainWrite(&buf_globalLock);
1152         while (1) {
1153                 if (buf_reservedBufs + nbuffers > buf_maxReservedBufs) {
1154                         buf_reserveWaiting = 1;
1155                         osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1156                         osi_SleepW((long) &buf_reservedBufs, &buf_globalLock);
1157                         lock_ObtainWrite(&buf_globalLock);
1158                 }
1159                 else {
1160                         buf_reservedBufs += nbuffers;
1161                         break;
1162                 }
1163         }
1164         lock_ReleaseWrite(&buf_globalLock);
1165 }
1166
1167 int buf_TryReserveBuffers(long nbuffers)
1168 {
1169         int code;
1170
1171         lock_ObtainWrite(&buf_globalLock);
1172         if (buf_reservedBufs + nbuffers > buf_maxReservedBufs) {
1173                 code = 0;
1174         }
1175         else {
1176                 buf_reservedBufs += nbuffers;
1177                 code = 1;
1178         }
1179         lock_ReleaseWrite(&buf_globalLock);
1180         return code;
1181 }
1182
1183 /* called without global lock held, releases reservation held by
1184  * buf_ReserveBuffers.
1185  */
1186 void buf_UnreserveBuffers(long nbuffers)
1187 {
1188         lock_ObtainWrite(&buf_globalLock);
1189         buf_reservedBufs -= nbuffers;
1190         if (buf_reserveWaiting) {
1191                 buf_reserveWaiting = 0;
1192                 osi_Wakeup((long) &buf_reservedBufs);
1193         }
1194         lock_ReleaseWrite(&buf_globalLock);
1195 }
1196
1197 /* truncate the buffers past sizep, zeroing out the page, if we don't
1198  * end on a page boundary.
1199  *
1200  * Requires cm_bufCreateLock to be write locked.
1201  */
1202 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1203         osi_hyper_t *sizep)
1204 {
1205         cm_buf_t *bufp;
1206         cm_buf_t *nbufp;                        /* next buffer, if didRelease */
1207         osi_hyper_t bufEnd;
1208         long code;
1209         long bufferPos;
1210         int didRelease;
1211         long i;
1212
1213         /* assert that cm_bufCreateLock is held in write mode */
1214         lock_AssertWrite(&scp->bufCreateLock);
1215
1216         i = BUF_FILEHASH(&scp->fid);
1217
1218         lock_ObtainWrite(&buf_globalLock);
1219         bufp = buf_fileHashTablepp[i];
1220         if (bufp == NULL) {
1221                 lock_ReleaseWrite(&buf_globalLock);
1222                 return 0;
1223         }
1224
1225         bufp->refCount++;
1226         lock_ReleaseWrite(&buf_globalLock);
1227         for(; bufp; bufp = nbufp) {
1228                 didRelease = 0;
1229                 lock_ObtainMutex(&bufp->mx);
1230
1231                 bufEnd.HighPart = 0;
1232                 bufEnd.LowPart = buf_bufferSize;
1233                 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1234
1235                 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1236                         LargeIntegerLessThan(*sizep, bufEnd)) {
1237                         buf_WaitIO(bufp);
1238                 }
1239                 lock_ObtainMutex(&scp->mx);
1240
1241                 /* make sure we have a callback (so we have the right value for
1242                  * the length), and wait for it to be safe to do a truncate.
1243                  */
1244                 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1245                                  CM_SCACHESYNC_NEEDCALLBACK
1246                                  | CM_SCACHESYNC_GETSTATUS
1247                                  | CM_SCACHESYNC_SETSIZE
1248                                  | CM_SCACHESYNC_BUFLOCKED);
1249                 /* if we succeeded in our locking, and this applies to the right
1250                  * file, and the truncate request overlaps the buffer either
1251                  * totally or partially, then do something.
1252                  */
1253                 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1254                         && LargeIntegerLessThan(*sizep, bufEnd)) {
1255
1256                         lock_ObtainWrite(&buf_globalLock);
1257
1258                         /* destroy the buffer, turning off its dirty bit, if
1259                          * we're truncating the whole buffer.  Otherwise, set
1260                          * the dirty bit, and clear out the tail of the buffer
1261                          * if we just overlap some.
1262                          */
1263                         if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1264                                 /* truncating the entire page */
1265                                 bufp->flags &= ~CM_BUF_DIRTY;
1266                                 bufp->dataVersion = -1; /* known bad */
1267                                 bufp->dirtyCounter++;
1268                         }
1269                         else {
1270                                 /* don't set dirty, since dirty implies
1271                                  * currently up-to-date.  Don't need to do this,
1272                                  * since we'll update the length anyway.
1273                                  *
1274                                  * Zero out remainder of the page, in case we
1275                                  * seek and write past EOF, and make this data
1276                                  * visible again.
1277                                  */
1278                                 bufferPos = sizep->LowPart & (buf_bufferSize - 1);
1279                                 osi_assert(bufferPos != 0);
1280                                 memset(bufp->datap + bufferPos, 0,
1281                                         buf_bufferSize - bufferPos);
1282                         }
1283
1284                         lock_ReleaseWrite(&buf_globalLock);
1285
1286                 }
1287
1288                 lock_ReleaseMutex(&scp->mx);
1289                 lock_ReleaseMutex(&bufp->mx);
1290                 if (!didRelease) {
1291                         lock_ObtainWrite(&buf_globalLock);
1292                         nbufp = bufp->fileHashp;
1293                         if (nbufp) nbufp->refCount++;
1294                         buf_LockedRelease(bufp);
1295                         lock_ReleaseWrite(&buf_globalLock);
1296                 }
1297
1298                 /* bail out early if we fail */
1299                 if (code) {
1300                         /* at this point, nbufp is held; bufp has already been
1301                          * released.
1302                          */
1303                         if (nbufp) buf_Release(nbufp);
1304                         return code;
1305                 }
1306         }
1307
1308         /* success */
1309         return 0;
1310 }
1311
1312 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1313 {
1314         long code;
1315         cm_buf_t *bp;           /* buffer we're hacking on */
1316         cm_buf_t *nbp;
1317         int didRelease;
1318         long i;
1319
1320         i = BUF_FILEHASH(&scp->fid);
1321
1322         code = 0;
1323         lock_ObtainWrite(&buf_globalLock);
1324         bp = buf_fileHashTablepp[i];
1325         if (bp) bp->refCount++;
1326         lock_ReleaseWrite(&buf_globalLock);
1327         for(; bp; bp = nbp) {
1328                 didRelease = 0; /* haven't released this buffer yet */
1329
1330                 /* clean buffer synchronously */
1331                 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1332                         lock_ObtainMutex(&bp->mx);
1333
1334                         /* start cleaning the buffer, and wait for it to finish */
1335                         buf_LockedCleanAsync(bp, reqp);
1336                         buf_WaitIO(bp);
1337                         lock_ReleaseMutex(&bp->mx);
1338
1339                         code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1340                         if (code) goto skip;
1341
1342                         lock_ObtainWrite(&buf_globalLock);
1343                         /* actually, we only know that buffer is clean if ref
1344                          * count is 1, since we don't have buffer itself locked.
1345                          */
1346                         if (!(bp->flags & CM_BUF_DIRTY)) {
1347                                 if (bp->refCount == 1) {        /* bp is held above */
1348                                         buf_LockedRelease(bp);
1349                                         nbp = bp->fileHashp;
1350                                         if (nbp) nbp->refCount++;
1351                                         didRelease = 1;
1352                                         buf_Recycle(bp);
1353                                 }
1354                         }
1355                         lock_ReleaseWrite(&buf_globalLock);
1356
1357                         (*cm_buf_opsp->Unstabilizep)(scp, userp);
1358                 }
1359
1360 skip:
1361                 if (!didRelease) {
1362                         lock_ObtainWrite(&buf_globalLock);
1363                         if (nbp = bp->fileHashp) nbp->refCount++;
1364                         buf_LockedRelease(bp);
1365                         lock_ReleaseWrite(&buf_globalLock);
1366                 }
1367         }       /* for loop over a bunch of buffers */
1368
1369         /* done */
1370         return code;
1371 }
1372
1373 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1374 {
1375         long code;
1376         cm_buf_t *bp;           /* buffer we're hacking on */
1377     cm_buf_t *nbp;              /* next one */
1378         long i;
1379
1380         i = BUF_FILEHASH(&scp->fid);
1381
1382         code = 0;
1383         lock_ObtainWrite(&buf_globalLock);
1384     bp = buf_fileHashTablepp[i];
1385     if (bp) bp->refCount++;
1386     lock_ReleaseWrite(&buf_globalLock);
1387         for(; bp; bp = nbp) {
1388                 /* clean buffer synchronously */
1389                 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1390                         if (userp) {
1391                 cm_HoldUser(userp);
1392                                 lock_ObtainMutex(&bp->mx);
1393                                 if (bp->userp)
1394                     cm_ReleaseUser(bp->userp);
1395                 bp->userp = userp;
1396                                 lock_ReleaseMutex(&bp->mx);
1397             }
1398                         buf_CleanAsync(bp, reqp);
1399             buf_CleanWait(bp);
1400             lock_ObtainMutex(&bp->mx);
1401                         if (bp->flags & CM_BUF_ERROR) {
1402                                 if (code == 0 || code == -1) code = bp->error;
1403                 if (code == 0) code = -1;
1404             }
1405             lock_ReleaseMutex(&bp->mx);
1406                 }
1407
1408                 lock_ObtainWrite(&buf_globalLock);
1409                 buf_LockedRelease(bp);
1410         nbp = bp->fileHashp;
1411         if (nbp) nbp->refCount++;
1412                 lock_ReleaseWrite(&buf_globalLock);
1413         }       /* for loop over a bunch of buffers */
1414
1415     /* done */
1416         return code;
1417 }
1418
1419 /* dump the contents of the buf_hashTablepp. */
1420 int cm_DumpBufHashTable(FILE *outputFile, char *cookie)
1421 {
1422     int zilch;
1423     cm_buf_t *bp;
1424     char output[1024];
1425     int i;
1426
1427         if (buf_hashTablepp == NULL)
1428                 return -1;
1429
1430     lock_ObtainRead(&buf_globalLock);
1431
1432     sprintf(output, "%s - dumping buf_HashTable - buf_hashSize=%d\n", cookie, buf_hashSize);
1433     WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1434
1435     for (i = 0; i < buf_hashSize; i++)
1436     {
1437         for(bp = buf_hashTablepp[i]; bp; bp=bp->hashp)
1438         {
1439             if (bp->refCount)
1440             {
1441                 sprintf(output, "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d,"
1442                         "vnode=%d, unique=%d), size=%d refCount=%d\n",
1443                         cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1444                         bp->fid.vnode, bp->fid.unique, bp->size, bp->refCount);
1445                 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1446             }
1447         }
1448     }
1449
1450     sprintf(output, "%s - Done dumping buf_HashTable.\n", cookie);
1451     WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1452
1453     lock_ReleaseRead(&buf_globalLock);
1454     return 0;
1455 }
1456