2 * Copyright 2000, International Business Machines Corporation and others.
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
10 /* Copyright (C) 1994 Cazamar Systems, Inc. */
12 #include <afsconfig.h>
13 #include <afs/param.h>
25 #include "cm_memmap.h"
28 #define TRACE_BUFFER 1
31 extern void afsi_log(char *pattern, ...);
33 /* This module implements the buffer package used by the local transaction
34 * system (cm). It is initialized by calling cm_Init, which calls buf_Init;
35 * it must be initalized before any of its main routines are called.
37 * Each buffer is hashed into a hash table by file ID and offset, and if its
38 * reference count is zero, it is also in a free list.
40 * There are two locks involved in buffer processing. The global lock
41 * buf_globalLock protects all of the global variables defined in this module,
42 * the reference counts and hash pointers in the actual cm_buf_t structures,
43 * and the LRU queue pointers in the buffer structures.
45 * The mutexes in the buffer structures protect the remaining fields in the
46 * buffers, as well the data itself.
48 * The locking hierarchy here is this:
50 * - resv multiple simul. buffers reservation
51 * - lock buffer I/O flags
52 * - lock buffer's mutex
53 * - lock buf_globalLock
57 /* global debugging log */
58 osi_log_t *buf_logp = NULL;
60 /* Global lock protecting hash tables and free lists */
61 osi_rwlock_t buf_globalLock;
63 /* ptr to head of the free list (most recently used) and the
64 * tail (the guy to remove first). We use osi_Q* functions
65 * to put stuff in buf_freeListp, and maintain the end
69 /* a pointer to a list of all buffers, just so that we can find them
70 * easily for debugging, and for the incr syncer. Locked under
74 /* defaults setup; these variables may be manually assigned into
75 * before calling cm_Init, as a way of changing these defaults.
78 /* callouts for reading and writing data, etc */
79 cm_buf_ops_t *cm_buf_opsp;
82 /* for experimental disk caching support in Win95 client */
83 cm_buf_t *buf_diskFreeListp;
84 cm_buf_t *buf_diskFreeListEndp;
85 cm_buf_t *buf_diskAllp;
86 extern int cm_diskCacheEnabled;
87 #endif /* DISKCACHE95 */
89 /* set this to 1 when we are terminating to prevent access attempts */
90 static int buf_ShutdownFlag = 0;
93 void buf_HoldLockedDbg(cm_buf_t *bp, char *file, long line)
95 void buf_HoldLocked(cm_buf_t *bp)
100 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
101 refCount = InterlockedIncrement(&bp->refCount);
102 #ifdef DEBUG_REFCOUNT
103 osi_Log2(afsd_logp,"buf_HoldLocked bp 0x%p ref %d",bp, refCount);
104 afsi_log("%s:%d buf_HoldLocked bp 0x%p, ref %d", file, line, bp, refCount);
108 /* hold a reference to an already held buffer */
109 #ifdef DEBUG_REFCOUNT
110 void buf_HoldDbg(cm_buf_t *bp, char *file, long line)
112 void buf_Hold(cm_buf_t *bp)
117 lock_ObtainRead(&buf_globalLock);
118 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
119 refCount = InterlockedIncrement(&bp->refCount);
120 #ifdef DEBUG_REFCOUNT
121 osi_Log2(afsd_logp,"buf_Hold bp 0x%p ref %d",bp, refCount);
122 afsi_log("%s:%d buf_Hold bp 0x%p, ref %d", file, line, bp, refCount);
124 lock_ReleaseRead(&buf_globalLock);
127 /* code to drop reference count while holding buf_globalLock */
128 #ifdef DEBUG_REFCOUNT
129 void buf_ReleaseLockedDbg(cm_buf_t *bp, afs_uint32 writeLocked, char *file, long line)
131 void buf_ReleaseLocked(cm_buf_t *bp, afs_uint32 writeLocked)
137 lock_AssertWrite(&buf_globalLock);
139 lock_AssertRead(&buf_globalLock);
141 /* ensure that we're in the LRU queue if our ref count is 0 */
142 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
144 refCount = InterlockedDecrement(&bp->refCount);
145 #ifdef DEBUG_REFCOUNT
146 osi_Log3(afsd_logp,"buf_ReleaseLocked %s bp 0x%p ref %d",writeLocked?"write":"read", bp, refCount);
147 afsi_log("%s:%d buf_ReleaseLocked %s bp 0x%p, ref %d", file, line, writeLocked?"write":"read", bp, refCount);
151 osi_panic("buf refcount 0",__FILE__,__LINE__);;
153 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
157 * If we are read locked there could be a race condition
158 * with buf_Find() so we must obtain a write lock and
159 * double check that the refCount is actually zero
160 * before we remove the buffer from the LRU queue.
163 lock_ConvertRToW(&buf_globalLock);
165 if (bp->refCount == 0 &&
166 !(bp->qFlags & CM_BUF_QINLRU)) {
167 osi_QAddH( (osi_queue_t **) &cm_data.buf_freeListp,
168 (osi_queue_t **) &cm_data.buf_freeListEndp,
170 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
174 lock_ConvertWToR(&buf_globalLock);
178 /* release a buffer. Buffer must be referenced, but unlocked. */
179 #ifdef DEBUG_REFCOUNT
180 void buf_ReleaseDbg(cm_buf_t *bp, char *file, long line)
182 void buf_Release(cm_buf_t *bp)
187 /* ensure that we're in the LRU queue if our ref count is 0 */
188 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
190 refCount = InterlockedDecrement(&bp->refCount);
191 #ifdef DEBUG_REFCOUNT
192 osi_Log2(afsd_logp,"buf_Release bp 0x%p ref %d", bp, refCount);
193 afsi_log("%s:%d buf_ReleaseLocked bp 0x%p, ref %d", file, line, bp, refCount);
197 osi_panic("buf refcount 0",__FILE__,__LINE__);;
199 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
202 lock_ObtainWrite(&buf_globalLock);
203 if (bp->refCount == 0 &&
204 !(bp->qFlags & CM_BUF_QINLRU)) {
205 osi_QAddH( (osi_queue_t **) &cm_data.buf_freeListp,
206 (osi_queue_t **) &cm_data.buf_freeListEndp,
208 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
210 lock_ReleaseWrite(&buf_globalLock);
215 buf_Sync(int quitOnShutdown)
217 cm_buf_t **bpp, *bp, *prevbp;
218 afs_uint32 wasDirty = 0;
221 /* go through all of the dirty buffers */
222 lock_ObtainRead(&buf_globalLock);
223 for (bpp = &cm_data.buf_dirtyListp, prevbp = NULL; bp = *bpp; ) {
224 if (quitOnShutdown && buf_ShutdownFlag)
227 lock_ReleaseRead(&buf_globalLock);
228 /* all dirty buffers are held when they are added to the
229 * dirty list. No need for an additional hold.
231 lock_ObtainMutex(&bp->mx);
233 if (bp->flags & CM_BUF_DIRTY && !(bp->qFlags & CM_BUF_QREDIR)) {
234 /* start cleaning the buffer; don't touch log pages since
235 * the log code counts on knowing exactly who is writing
236 * a log page at any given instant.
238 * only attempt to write the buffer if the volume might
244 volp = cm_GetVolumeByFID(&bp->fid);
245 switch (cm_GetVolumeStatus(volp, bp->fid.volume)) {
249 req.flags |= CM_REQ_NORETRY;
250 buf_CleanAsyncLocked(NULL, bp, &req, 0, &dirty);
256 /* the buffer may or may not have been dirty
257 * and if dirty may or may not have been cleaned
258 * successfully. check the dirty flag again.
260 if (!(bp->flags & CM_BUF_DIRTY)) {
261 /* remove the buffer from the dirty list */
262 lock_ObtainWrite(&buf_globalLock);
263 #ifdef DEBUG_REFCOUNT
264 if (bp->dirtyp == NULL && bp != cm_data.buf_dirtyListEndp) {
265 osi_Log1(afsd_logp,"buf_Sync bp 0x%p list corruption",bp);
266 afsi_log("buf_Sync bp 0x%p list corruption", bp);
271 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINDL);
272 if (cm_data.buf_dirtyListp == NULL)
273 cm_data.buf_dirtyListEndp = NULL;
274 else if (cm_data.buf_dirtyListEndp == bp)
275 cm_data.buf_dirtyListEndp = prevbp;
276 buf_ReleaseLocked(bp, TRUE);
277 lock_ConvertWToR(&buf_globalLock);
279 if (buf_ShutdownFlag) {
282 char volstr[VL_MAXNAMELEN+12]="";
285 volp = cm_GetVolumeByFID(&bp->fid);
288 if (bp->fid.volume == volp->vol[RWVOL].ID)
290 else if (bp->fid.volume == volp->vol[ROVOL].ID)
292 else if (bp->fid.volume == volp->vol[BACKVOL].ID)
296 snprintf(volstr, sizeof(volstr), "%s%s", volp->namep, ext);
298 cellp = cm_FindCellByID(bp->fid.cell, CM_FLAG_NOPROBE);
299 snprintf(volstr, sizeof(volstr), "%u", bp->fid.volume);
302 LogEvent(EVENTLOG_INFORMATION_TYPE, MSG_DIRTY_BUFFER_AT_SHUTDOWN,
303 cellp->name, volstr, bp->fid.vnode, bp->fid.unique,
304 bp->offset.QuadPart+bp->dirty_offset, bp->dirty_length);
307 /* advance the pointer so we don't loop forever */
308 lock_ObtainRead(&buf_globalLock);
312 lock_ReleaseMutex(&bp->mx);
313 } /* for loop over a bunch of buffers */
314 lock_ReleaseRead(&buf_globalLock);
319 /* incremental sync daemon. Writes all dirty buffers every 5000 ms */
320 void buf_IncrSyncer(long parm)
325 while (buf_ShutdownFlag == 0) {
327 i = SleepEx(5000, 1);
334 wasDirty = buf_Sync(1);
335 } /* whole daemon's while loop */
339 buf_ValidateBuffers(void)
341 cm_buf_t * bp, *bpf, *bpa, *bpb;
342 afs_uint64 countb = 0, countf = 0, counta = 0;
344 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
345 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
346 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
347 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
351 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
352 if (bp->magic != CM_BUF_MAGIC) {
353 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
354 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
360 if (countb > cm_data.buf_nbuffers) {
361 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
362 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
367 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
368 if (bp->magic != CM_BUF_MAGIC) {
369 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
370 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
376 if (countf > cm_data.buf_nbuffers) {
377 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
378 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
383 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
384 if (bp->magic != CM_BUF_MAGIC) {
385 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
386 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
392 if (counta > cm_data.buf_nbuffers) {
393 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
394 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
399 if (countb != countf) {
400 afsi_log("cm_ValidateBuffers failure: countb != countf");
401 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
405 if (counta != cm_data.buf_nbuffers) {
406 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
407 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
414 void buf_Shutdown(void)
416 /* disable the buf_IncrSyncer() threads */
417 buf_ShutdownFlag = 1;
419 /* then force all dirty buffers to the file servers */
423 /* initialize the buffer package; called with no locks
424 * held during the initialization phase.
426 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
428 static osi_once_t once;
437 cm_data.buf_nbuffers = nbuffers;
439 /* Have to be able to reserve a whole chunk */
440 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
441 return CM_ERROR_TOOFEWBUFS;
444 /* recall for callouts */
447 if (osi_Once(&once)) {
448 /* initialize global locks */
449 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock", LOCK_HIERARCHY_BUF_GLOBAL);
452 /* remember this for those who want to reset it */
453 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
455 /* lower hash size to a prime number */
456 cm_data.buf_hashSize = osi_PrimeLessThan((afs_uint32)(cm_data.buf_nbuffers/7 + 1));
458 /* create hash table */
459 memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
461 /* another hash table */
462 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
464 /* create buffer headers and put in free list */
465 bp = cm_data.bufHeaderBaseAddress;
466 data = cm_data.bufDataBaseAddress;
467 cm_data.buf_allp = NULL;
469 for (i=0; i<cm_data.buf_nbuffers; i++) {
470 osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress,
471 "invalid cm_buf_t address");
472 osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
473 "invalid cm_buf_t data address");
475 /* allocate and zero some storage */
476 memset(bp, 0, sizeof(cm_buf_t));
477 bp->magic = CM_BUF_MAGIC;
478 /* thread on list of all buffers */
479 bp->allp = cm_data.buf_allp;
480 cm_data.buf_allp = bp;
482 osi_QAddH( (osi_queue_t **) &cm_data.buf_freeListp,
483 (osi_queue_t **) &cm_data.buf_freeListEndp,
485 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
486 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
488 /* grab appropriate number of bytes from aligned zone */
493 data += cm_data.buf_blockSize;
496 /* none reserved at first */
497 cm_data.buf_reservedBufs = 0;
499 /* just for safety's sake */
500 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
502 bp = cm_data.bufHeaderBaseAddress;
503 data = cm_data.bufDataBaseAddress;
505 for (i=0; i<cm_data.buf_nbuffers; i++) {
506 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
509 bp->waitRequests = 0;
510 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
516 buf_ValidateBufQueues();
520 /* init the buffer trace log */
521 buf_logp = osi_LogCreate("buffer", 1000);
522 osi_LogEnable(buf_logp);
527 /* and create the incr-syncer */
528 phandle = thrd_Create(0, 0,
529 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
532 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
533 CloseHandle(phandle);
537 buf_ValidateBufQueues();
542 /* add nbuffers to the buffer pool, if possible.
543 * Called with no locks held.
545 long buf_AddBuffers(afs_uint64 nbuffers)
547 /* The size of a virtual cache cannot be changed after it has
548 * been created. Subsequent calls to MapViewofFile() with
549 * an existing mapping object name would not allow the
550 * object to be resized. Return failure immediately.
552 * A similar problem now occurs with the persistent cache
553 * given that the memory mapped file now contains a complex
556 afsi_log("request to add %d buffers to the existing cache of size %d denied",
557 nbuffers, cm_data.buf_nbuffers);
559 return CM_ERROR_INVAL;
562 /* interface to set the number of buffers to an exact figure.
563 * Called with no locks held.
565 long buf_SetNBuffers(afs_uint64 nbuffers)
568 return CM_ERROR_INVAL;
569 if (nbuffers == cm_data.buf_nbuffers)
571 else if (nbuffers > cm_data.buf_nbuffers)
572 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
574 return CM_ERROR_INVAL;
577 /* wait for reading or writing to clear; called with write-locked
578 * buffer and unlocked scp and returns with locked buffer.
580 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
585 osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
586 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
589 /* if no IO is happening, we're done */
590 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
593 /* otherwise I/O is happening, but some other thread is waiting for
594 * the I/O already. Wait for that guy to figure out what happened,
595 * and then check again.
597 if ( bp->flags & CM_BUF_WAITING ) {
600 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
602 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
603 _InterlockedOr(&bp->flags, CM_BUF_WAITING);
604 bp->waitCount = bp->waitRequests = 1;
606 osi_SleepM((LONG_PTR)bp, &bp->mx);
608 cm_UpdateServerPriority();
610 lock_ObtainMutex(&bp->mx);
611 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
613 if (bp->waitCount == 0) {
614 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
615 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
616 bp->waitRequests = 0;
620 if (scp = cm_FindSCache(&bp->fid))
624 lock_ObtainRead(&scp->rw);
625 if (scp->flags & CM_SCACHEFLAG_WAITING) {
626 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
627 osi_Wakeup((LONG_PTR)&scp->flags);
629 lock_ReleaseRead(&scp->rw);
633 /* if we get here, the IO is done, but we may have to wakeup people waiting for
634 * the I/O to complete. Do so.
636 if (bp->flags & CM_BUF_WAITING) {
637 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
638 osi_Wakeup((LONG_PTR) bp);
640 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
643 cm_ReleaseSCache(scp);
646 /* find a buffer, if any, for a particular file ID and offset. Assumes
647 * that buf_globalLock is write locked when called.
649 cm_buf_t *buf_FindLocked(struct cm_fid *fidp, osi_hyper_t *offsetp)
654 lock_AssertAny(&buf_globalLock);
656 i = BUF_HASH(fidp, offsetp);
657 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
658 if (cm_FidCmp(fidp, &bp->fid) == 0
659 && offsetp->LowPart == bp->offset.LowPart
660 && offsetp->HighPart == bp->offset.HighPart) {
666 /* return whatever we found, if anything */
670 /* find a buffer with offset *offsetp for vnode *scp. Called
671 * with no locks held.
673 cm_buf_t *buf_Find(struct cm_fid *fidp, osi_hyper_t *offsetp)
677 lock_ObtainRead(&buf_globalLock);
678 bp = buf_FindLocked(fidp, offsetp);
679 lock_ReleaseRead(&buf_globalLock);
684 /* find a buffer, if any, for a particular file ID and offset. Assumes
685 * that buf_globalLock is write locked when called. Uses the all buffer
688 cm_buf_t *buf_FindAllLocked(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
693 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
694 if (cm_FidCmp(fidp, &bp->fid) == 0
695 && offsetp->LowPart == bp->offset.LowPart
696 && offsetp->HighPart == bp->offset.HighPart) {
702 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
703 if (cm_FidCmp(fidp, &bp->fid) == 0) {
706 fileOffset = offsetp->QuadPart + cm_data.baseAddress;
707 if (fileOffset == bp->datap) {
714 /* return whatever we found, if anything */
718 /* find a buffer with offset *offsetp for vnode *scp. Called
719 * with no locks held. Use the all buffer list.
721 cm_buf_t *buf_FindAll(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
725 lock_ObtainRead(&buf_globalLock);
726 bp = buf_FindAllLocked(fidp, offsetp, flags);
727 lock_ReleaseRead(&buf_globalLock);
732 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
735 * Makes sure that there's only one person writing this block
736 * at any given time, and also ensures that the log is forced sufficiently far,
737 * if this buffer contains logged data.
739 * Returns non-zero if the buffer was dirty.
741 * 'scp' may or may not be NULL. If it is not NULL, the FID for both cm_scache_t
742 * and cm_buf_t must match.
744 afs_uint32 buf_CleanAsyncLocked(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp,
745 afs_uint32 flags, afs_uint32 *pisdirty)
748 afs_uint32 isdirty = 0;
752 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
753 osi_assertx(scp == NULL || cm_FidCmp(&scp->fid, &bp->fid) == 0, "scp fid != bp fid");
756 * If the matching cm_scache_t was not provided as a parameter
757 * we must either find one or allocate a new one. It is possible
758 * that the cm_scache_t was recycled out of the cache even though
759 * a cm_buf_t with the same FID is in the cache.
762 if ((scp = cm_FindSCache(&bp->fid)) ||
763 (cm_GetSCache(&bp->fid, &scp,
764 bp->userp ? bp->userp : cm_rootUserp,
770 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
772 lock_ReleaseMutex(&bp->mx);
776 * If we didn't find a cm_scache_t object for bp->fid it means
777 * that we no longer have that FID in the cache. It does not
778 * mean that the object does not exist in the cell. That may
779 * in fact be the case but we don't know that until we attempt
780 * a FetchStatus on the FID.
782 osi_Log1(buf_logp, "buf_CleanAsyncLocked unable to start I/O - scp not found buf 0x%p", bp);
783 code = CM_ERROR_NOSUCHFILE;
785 osi_Log2(buf_logp, "buf_CleanAsyncLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
788 LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
789 code = (*cm_buf_opsp->Writep)(scp, &offset,
791 /* we might as well try to write all of the contiguous
792 * dirty buffers in one RPC
798 flags, bp->userp, reqp);
799 osi_Log3(buf_logp, "buf_CleanAsyncLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
801 lock_ObtainMutex(&bp->mx);
802 /* if the Write routine returns No Such File, clear the dirty flag
803 * because we aren't going to be able to write this data to the file
806 if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD || code == CM_ERROR_NOACCESS ||
807 code == CM_ERROR_QUOTA || code == CM_ERROR_SPACE || code == CM_ERROR_TOOBIG ||
808 code == CM_ERROR_READONLY || code == CM_ERROR_NOSUCHPATH){
809 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
810 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
811 bp->dirty_offset = 0;
812 bp->dirty_length = 0;
814 bp->dataVersion = CM_BUF_VERSION_BAD;
820 /* Disk cache support */
821 /* write buffer to disk cache (synchronous for now) */
822 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
823 #endif /* DISKCACHE95 */
825 /* if we get here and retries are not permitted
826 * then we need to exit this loop regardless of
827 * whether or not we were able to clear the dirty bit
829 if (reqp->flags & CM_REQ_NORETRY)
832 /* Ditto if the hardDeadTimeout or idleTimeout was reached */
833 if (code == CM_ERROR_TIMEDOUT || code == CM_ERROR_ALLDOWN ||
834 code == CM_ERROR_ALLBUSY || code == CM_ERROR_ALLOFFLINE ||
835 code == CM_ERROR_CLOCKSKEW) {
841 cm_ReleaseSCache(scp);
843 /* if someone was waiting for the I/O that just completed or failed,
846 if (bp->flags & CM_BUF_WAITING) {
847 /* turn off flags and wakeup users */
848 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
849 osi_Wakeup((LONG_PTR) bp);
858 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
859 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
860 * The buffer must already be clean, and no I/O should be happening to it.
862 void buf_Recycle(cm_buf_t *bp)
867 cm_buf_t *prevBp, *nextBp;
869 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
871 /* if we get here, we know that the buffer still has a 0 ref count,
872 * and that it is clean and has no currently pending I/O. This is
873 * the dude to return.
874 * Remember that as long as the ref count is 0, we know that we won't
875 * have any lock conflicts, so we can grab the buffer lock out of
876 * order in the locking hierarchy.
878 osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
879 bp, bp->offset.HighPart, bp->offset.LowPart);
881 osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
882 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
883 "incorrect cm_buf_t flags");
884 lock_AssertWrite(&buf_globalLock);
886 if (bp->qFlags & CM_BUF_QINHASH) {
887 /* Remove from hash */
889 i = BUF_HASH(&bp->fid, &bp->offset);
890 lbpp = &(cm_data.buf_scacheHashTablepp[i]);
891 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = tbp->hashp) {
896 /* we better find it */
897 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
899 *lbpp = bp->hashp; /* hash out */
902 /* Remove from file hash */
904 i = BUF_FILEHASH(&bp->fid);
905 prevBp = bp->fileHashBackp;
906 bp->fileHashBackp = NULL;
907 nextBp = bp->fileHashp;
908 bp->fileHashp = NULL;
910 prevBp->fileHashp = nextBp;
912 cm_data.buf_fileHashTablepp[i] = nextBp;
914 nextBp->fileHashBackp = prevBp;
916 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINHASH);
919 /* make the fid unrecognizable */
920 memset(&bp->fid, 0, sizeof(cm_fid_t));
922 /* clean up junk flags */
923 _InterlockedAnd(&bp->flags, ~(CM_BUF_EOF | CM_BUF_ERROR));
924 bp->dataVersion = CM_BUF_VERSION_BAD; /* unknown so far */
927 /* recycle a buffer, removing it from the free list, hashing in its new identity
928 * and returning it write-locked so that no one can use it. Called without
929 * any locks held, and can return an error if it loses the race condition and
930 * finds that someone else created the desired buffer.
932 * If success is returned, the buffer is returned write-locked.
934 * May be called with null scp and offsetp, if we're just trying to reclaim some
935 * space from the buffer pool. In that case, the buffer will be returned
936 * without being hashed into the hash table.
938 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
940 cm_buf_t *bp; /* buffer we're dealing with */
941 cm_buf_t *nextBp; /* next buffer in file hash chain */
942 afs_uint32 i; /* temp */
945 buf_ValidateBufQueues();
950 lock_ObtainRead(&scp->bufCreateLock);
951 lock_ObtainWrite(&buf_globalLock);
952 /* check to see if we lost the race */
954 if (bp = buf_FindLocked(&scp->fid, offsetp)) {
955 /* Do not call buf_ReleaseLocked() because we
956 * do not want to allow the buffer to be added
959 afs_int32 refCount = InterlockedDecrement(&bp->refCount);
960 #ifdef DEBUG_REFCOUNT
961 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, refCount);
962 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, refCount);
964 lock_ReleaseWrite(&buf_globalLock);
965 lock_ReleaseRead(&scp->bufCreateLock);
966 return CM_BUF_EXISTS;
970 /* does this fix the problem below? it's a simple solution. */
971 if (!cm_data.buf_freeListEndp)
973 lock_ReleaseWrite(&buf_globalLock);
974 lock_ReleaseRead(&scp->bufCreateLock);
975 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
980 /* for debugging, assert free list isn't empty, although we
981 * really should try waiting for a running tranasction to finish
982 * instead of this; or better, we should have a transaction
983 * throttler prevent us from entering this situation.
985 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
987 /* look at all buffers in free list, some of which may temp.
988 * have high refcounts and which then should be skipped,
989 * starting cleaning I/O for those which are dirty. If we find
990 * a clean buffer, we rehash it, lock it and return it.
992 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
993 /* check to see if it really has zero ref count. This
994 * code can bump refcounts, at least, so it may not be
997 if (bp->refCount > 0)
1000 /* we don't have to lock buffer itself, since the ref
1001 * count is 0 and we know it will stay zero as long as
1002 * we hold the global lock.
1005 /* Don't recycle a buffer held by the redirector. */
1006 if (bp->qFlags & CM_BUF_QREDIR)
1009 /* don't recycle someone in our own chunk */
1010 if (!cm_FidCmp(&bp->fid, &scp->fid)
1011 && (bp->offset.LowPart & (-cm_chunkSize))
1012 == (offsetp->LowPart & (-cm_chunkSize)))
1015 /* if this page is being filled (!) or cleaned, see if
1016 * the I/O has completed. If not, skip it, otherwise
1017 * do the final processing for the I/O.
1019 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
1020 /* probably shouldn't do this much work while
1021 * holding the big lock? Watch for contention
1027 if (bp->flags & CM_BUF_DIRTY) {
1028 /* if the buffer is dirty, start cleaning it and
1029 * move on to the next buffer. We do this with
1030 * just the lock required to minimize contention
1034 lock_ReleaseWrite(&buf_globalLock);
1035 lock_ReleaseRead(&scp->bufCreateLock);
1037 /* grab required lock and clean; this only
1038 * starts the I/O. By the time we're back,
1039 * it'll still be marked dirty, but it will also
1040 * have the WRITING flag set, so we won't get
1043 if (cm_FidCmp(&scp->fid, &bp->fid) == 0)
1044 buf_CleanAsync(scp, bp, reqp, 0, NULL);
1046 buf_CleanAsync(NULL, bp, reqp, 0, NULL);
1048 /* now put it back and go around again */
1053 /* if we get here, we know that the buffer still has a 0
1054 * ref count, and that it is clean and has no currently
1055 * pending I/O. This is the dude to return.
1056 * Remember that as long as the ref count is 0, we know
1057 * that we won't have any lock conflicts, so we can grab
1058 * the buffer lock out of order in the locking hierarchy.
1062 /* now hash in as our new buffer, and give it the
1063 * appropriate label, if requested.
1066 lock_AssertWrite(&buf_globalLock);
1068 _InterlockedOr(&bp->qFlags, CM_BUF_QINHASH);
1073 bp->offset = *offsetp;
1074 i = BUF_HASH(&scp->fid, offsetp);
1075 bp->hashp = cm_data.buf_scacheHashTablepp[i];
1076 cm_data.buf_scacheHashTablepp[i] = bp;
1077 i = BUF_FILEHASH(&scp->fid);
1078 nextBp = cm_data.buf_fileHashTablepp[i];
1079 bp->fileHashp = nextBp;
1080 bp->fileHashBackp = NULL;
1082 nextBp->fileHashBackp = bp;
1083 cm_data.buf_fileHashTablepp[i] = bp;
1086 /* we should remove it from the lru queue. It better still be there,
1087 * since we've held the global (big) lock since we found it there.
1089 osi_assertx(bp->qFlags & CM_BUF_QINLRU,
1090 "buf_GetNewLocked: LRU screwup");
1092 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
1093 (osi_queue_t **) &cm_data.buf_freeListEndp,
1095 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1097 /* prepare to return it. Give it a refcount */
1099 #ifdef DEBUG_REFCOUNT
1100 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, 1);
1101 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, 1);
1103 /* grab the mutex so that people don't use it
1104 * before the caller fills it with data. Again, no one
1105 * should have been able to get to this dude to lock it.
1107 if (!lock_TryMutex(&bp->mx)) {
1108 osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked. refCount %d should be 0",
1110 osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
1113 lock_ReleaseWrite(&buf_globalLock);
1114 lock_ReleaseRead(&scp->bufCreateLock);
1119 buf_ValidateBufQueues();
1120 #endif /* TESTING */
1122 } /* for all buffers in lru queue */
1123 lock_ReleaseWrite(&buf_globalLock);
1124 lock_ReleaseRead(&scp->bufCreateLock);
1125 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List has no buffers with a zero refcount - sleeping 100ms");
1126 Sleep(100); /* give some time for a buffer to be freed */
1127 } /* while loop over everything */
1132 * get a page, returning it held but unlocked. the page may or may not
1133 * contain valid data.
1135 * The scp must be unlocked when passed in unlocked.
1137 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
1141 osi_hyper_t pageOffset;
1142 unsigned long tcount;
1146 cm_diskcache_t *dcp;
1147 #endif /* DISKCACHE95 */
1150 pageOffset.HighPart = offsetp->HighPart;
1151 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
1155 buf_ValidateBufQueues();
1156 #endif /* TESTING */
1158 bp = buf_Find(&scp->fid, &pageOffset);
1160 /* lock it and break out */
1161 lock_ObtainMutex(&bp->mx);
1164 /* touch disk chunk to update LRU info */
1165 diskcache_Touch(bp->dcp);
1166 #endif /* DISKCACHE95 */
1170 /* otherwise, we have to create a page */
1171 code = buf_GetNewLocked(scp, &pageOffset, reqp, &bp);
1174 /* the requested buffer was created */
1179 * the requested buffer existed by the time the
1180 * scp->bufCreateLock and buf_globalLock could be obtained.
1181 * loop again and permit buf_Find() to obtain a reference.
1186 * the requested buffer could not be created.
1187 * return the error to the caller.
1190 buf_ValidateBufQueues();
1191 #endif /* TESTING */
1194 } /* big while loop */
1196 /* if we get here, we have a locked buffer that may have just been
1197 * created, in which case it needs to be filled with data.
1200 /* load the page; freshly created pages should be idle */
1201 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
1204 * start the I/O; may drop lock. as of this writing, the only
1205 * implementation of Readp is cm_BufRead() which simply sets
1206 * tcount to 0 and returns success.
1208 _InterlockedOr(&bp->flags, CM_BUF_READING);
1209 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
1212 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
1213 bp->dcp = dcp; /* pointer to disk cache struct. */
1214 #endif /* DISKCACHE95 */
1217 /* failure or queued */
1218 if (code != ERROR_IO_PENDING) {
1220 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1221 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1222 if (bp->flags & CM_BUF_WAITING) {
1223 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1224 osi_Wakeup((LONG_PTR) bp);
1226 lock_ReleaseMutex(&bp->mx);
1229 buf_ValidateBufQueues();
1230 #endif /* TESTING */
1235 * otherwise, I/O completed instantly and we're done, except
1236 * for padding the xfr out with 0s and checking for EOF
1238 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1239 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1241 _InterlockedOr(&bp->flags, CM_BUF_EOF);
1243 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1244 if (bp->flags & CM_BUF_WAITING) {
1245 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1246 osi_Wakeup((LONG_PTR) bp);
1251 /* wait for reads, either that which we started above, or that someone
1252 * else started. We don't care if we return a buffer being cleaned.
1254 if (bp->flags & CM_BUF_READING)
1255 buf_WaitIO(scp, bp);
1257 /* once it has been read once, we can unlock it and return it, still
1258 * with its refcount held.
1260 lock_ReleaseMutex(&bp->mx);
1263 /* now remove from queue; will be put in at the head (farthest from
1264 * being recycled) when we're done in buf_Release.
1266 lock_ObtainWrite(&buf_globalLock);
1267 if (bp->qFlags & CM_BUF_QINLRU) {
1268 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
1269 (osi_queue_t **) &cm_data.buf_freeListEndp,
1271 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1273 lock_ReleaseWrite(&buf_globalLock);
1275 osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1276 bp, scp, offsetp->HighPart, offsetp->LowPart);
1278 buf_ValidateBufQueues();
1279 #endif /* TESTING */
1283 /* count # of elements in the free list;
1284 * we don't bother doing the proper locking for accessing dataVersion or flags
1285 * since it is a pain, and this is really just an advisory call. If you need
1286 * to do better at some point, rewrite this function.
1288 long buf_CountFreeList(void)
1294 lock_ObtainRead(&buf_globalLock);
1295 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1296 /* if the buffer doesn't have an identity, or if the buffer
1297 * has been invalidate (by having its DV stomped upon), then
1298 * count it as free, since it isn't really being utilized.
1300 if (!(bufp->qFlags & CM_BUF_QINHASH) || bufp->dataVersion == CM_BUF_VERSION_BAD)
1303 lock_ReleaseRead(&buf_globalLock);
1307 /* clean a buffer synchronously */
1308 afs_uint32 buf_CleanAsync(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp, afs_uint32 flags, afs_uint32 *pisdirty)
1311 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1312 osi_assertx(!(flags & CM_BUF_WRITE_SCP_LOCKED), "scp->rw must not be held when calling buf_CleanAsync");
1314 lock_ObtainMutex(&bp->mx);
1315 code = buf_CleanAsyncLocked(scp, bp, reqp, flags, pisdirty);
1316 lock_ReleaseMutex(&bp->mx);
1321 /* wait for a buffer's cleaning to finish */
1322 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp, afs_uint32 locked)
1324 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1327 lock_ObtainMutex(&bp->mx);
1328 if (bp->flags & CM_BUF_WRITING) {
1329 buf_WaitIO(scp, bp);
1332 lock_ReleaseMutex(&bp->mx);
1335 /* set the dirty flag on a buffer, and set associated write-ahead log,
1336 * if there is one. Allow one to be added to a buffer, but not changed.
1338 * The buffer must be locked before calling this routine.
1340 void buf_SetDirty(cm_buf_t *bp, cm_req_t *reqp, afs_uint32 offset, afs_uint32 length, cm_user_t *userp)
1342 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1343 osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
1348 if (bp->flags & CM_BUF_DIRTY) {
1350 osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
1352 if (bp->dirty_offset <= offset) {
1353 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1354 /* dirty_length remains the same */
1356 bp->dirty_length = offset + length - bp->dirty_offset;
1358 } else /* bp->dirty_offset > offset */ {
1359 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1360 bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
1362 bp->dirty_length = length;
1364 bp->dirty_offset = offset;
1367 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1370 _InterlockedOr(&bp->flags, CM_BUF_DIRTY);
1372 /* and turn off EOF flag, since it has associated data now */
1373 _InterlockedAnd(&bp->flags, ~CM_BUF_EOF);
1375 bp->dirty_offset = offset;
1376 bp->dirty_length = length;
1378 /* and add to the dirty list.
1379 * we obtain a hold on the buffer for as long as it remains
1380 * in the list. buffers are only removed from the list by
1381 * the buf_IncrSyncer function regardless of when else the
1382 * dirty flag might be cleared.
1384 * This should never happen but just in case there is a bug
1385 * elsewhere, never add to the dirty list if the buffer is
1388 lock_ObtainWrite(&buf_globalLock);
1389 if (!(bp->qFlags & CM_BUF_QINDL)) {
1391 if (!cm_data.buf_dirtyListp) {
1392 cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
1394 cm_data.buf_dirtyListEndp->dirtyp = bp;
1395 cm_data.buf_dirtyListEndp = bp;
1398 _InterlockedOr(&bp->qFlags, CM_BUF_QINDL);
1400 lock_ReleaseWrite(&buf_globalLock);
1403 /* and record the last writer */
1404 if (bp->userp != userp) {
1407 cm_ReleaseUser(bp->userp);
1412 /* clean all buffers, reset log pointers and invalidate all buffers.
1413 * Called with no locks held, and returns with same.
1415 * This function is guaranteed to clean and remove the log ptr of all the
1416 * buffers that were dirty or had non-zero log ptrs before the call was
1417 * made. That's sufficient to clean up any garbage left around by recovery,
1418 * which is all we're counting on this for; there may be newly created buffers
1419 * added while we're running, but that should be OK.
1421 * In an environment where there are no transactions (artificially imposed, for
1422 * example, when switching the database to raw mode), this function is used to
1423 * make sure that all updates have been written to the disk. In that case, we don't
1424 * really require that we forget the log association between pages and logs, but
1425 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1426 * have to worry about invalidating data in the buffers.
1428 * This function is used at the end of recovery as paranoia to get the recovered
1429 * database out to disk. It removes all references to the recovery log and cleans
1432 long buf_CleanAndReset(void)
1438 lock_ObtainRead(&buf_globalLock);
1439 for(i=0; i<cm_data.buf_hashSize; i++) {
1440 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
1441 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1443 lock_ReleaseRead(&buf_globalLock);
1445 /* now no locks are held; clean buffer and go on */
1447 req.flags |= CM_REQ_NORETRY;
1449 buf_CleanAsync(NULL, bp, &req, 0, NULL);
1450 buf_CleanWait(NULL, bp, FALSE);
1452 /* relock and release buffer */
1453 lock_ObtainRead(&buf_globalLock);
1454 buf_ReleaseLocked(bp, FALSE);
1456 } /* over one bucket */
1457 } /* for loop over all hash buckets */
1460 lock_ReleaseRead(&buf_globalLock);
1463 buf_ValidateBufQueues();
1464 #endif /* TESTING */
1466 /* and we're done */
1470 /* called without global lock being held, reserves buffers for callers
1471 * that need more than one held (not locked) at once.
1473 void buf_ReserveBuffers(afs_uint64 nbuffers)
1475 lock_ObtainWrite(&buf_globalLock);
1477 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1478 cm_data.buf_reserveWaiting = 1;
1479 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1480 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1481 lock_ObtainWrite(&buf_globalLock);
1484 cm_data.buf_reservedBufs += nbuffers;
1488 lock_ReleaseWrite(&buf_globalLock);
1491 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1495 lock_ObtainWrite(&buf_globalLock);
1496 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1500 cm_data.buf_reservedBufs += nbuffers;
1503 lock_ReleaseWrite(&buf_globalLock);
1507 /* called without global lock held, releases reservation held by
1508 * buf_ReserveBuffers.
1510 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1512 lock_ObtainWrite(&buf_globalLock);
1513 cm_data.buf_reservedBufs -= nbuffers;
1514 if (cm_data.buf_reserveWaiting) {
1515 cm_data.buf_reserveWaiting = 0;
1516 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1518 lock_ReleaseWrite(&buf_globalLock);
1521 /* truncate the buffers past sizep, zeroing out the page, if we don't
1522 * end on a page boundary.
1524 * Requires cm_bufCreateLock to be write locked.
1526 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1530 cm_buf_t *nbufp; /* next buffer, if didRelease */
1536 /* assert that cm_bufCreateLock is held in write mode */
1537 lock_AssertWrite(&scp->bufCreateLock);
1539 i = BUF_FILEHASH(&scp->fid);
1541 lock_ObtainRead(&buf_globalLock);
1542 bufp = cm_data.buf_fileHashTablepp[i];
1544 lock_ReleaseRead(&buf_globalLock);
1548 buf_HoldLocked(bufp);
1549 lock_ReleaseRead(&buf_globalLock);
1551 lock_ObtainMutex(&bufp->mx);
1553 bufEnd.HighPart = 0;
1554 bufEnd.LowPart = cm_data.buf_blockSize;
1555 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1557 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1558 LargeIntegerLessThan(*sizep, bufEnd)) {
1559 buf_WaitIO(scp, bufp);
1561 lock_ObtainWrite(&scp->rw);
1563 /* make sure we have a callback (so we have the right value for
1564 * the length), and wait for it to be safe to do a truncate.
1566 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1567 CM_SCACHESYNC_NEEDCALLBACK
1568 | CM_SCACHESYNC_GETSTATUS
1569 | CM_SCACHESYNC_SETSIZE
1570 | CM_SCACHESYNC_BUFLOCKED);
1573 /* if we succeeded in our locking, and this applies to the right
1574 * file, and the truncate request overlaps the buffer either
1575 * totally or partially, then do something.
1577 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1578 && LargeIntegerLessThan(*sizep, bufEnd)) {
1581 /* destroy the buffer, turning off its dirty bit, if
1582 * we're truncating the whole buffer. Otherwise, set
1583 * the dirty bit, and clear out the tail of the buffer
1584 * if we just overlap some.
1586 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1587 /* truncating the entire page */
1588 _InterlockedAnd(&bufp->flags, ~CM_BUF_DIRTY);
1589 bufp->dirty_offset = 0;
1590 bufp->dirty_length = 0;
1591 bufp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1592 bufp->dirtyCounter++;
1595 /* don't set dirty, since dirty implies
1596 * currently up-to-date. Don't need to do this,
1597 * since we'll update the length anyway.
1599 * Zero out remainder of the page, in case we
1600 * seek and write past EOF, and make this data
1603 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1604 osi_assertx(bufferPos != 0, "non-zero bufferPos");
1605 memset(bufp->datap + bufferPos, 0,
1606 cm_data.buf_blockSize - bufferPos);
1610 cm_SyncOpDone( scp, bufp,
1611 CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
1612 | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
1614 lock_ReleaseWrite(&scp->rw);
1615 lock_ReleaseMutex(&bufp->mx);
1618 nbufp = bufp->fileHashp;
1622 /* This forces the loop to end and the error code
1623 * to be returned. */
1631 buf_ValidateBufQueues();
1632 #endif /* TESTING */
1638 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1641 cm_buf_t *bp; /* buffer we're hacking on */
1645 afs_uint32 stable = 0;
1647 i = BUF_FILEHASH(&scp->fid);
1650 lock_ObtainRead(&buf_globalLock);
1651 bp = cm_data.buf_fileHashTablepp[i];
1654 lock_ReleaseRead(&buf_globalLock);
1656 for (; bp; bp = nbp) {
1657 didRelease = 0; /* haven't released this buffer yet */
1659 /* clean buffer synchronously */
1660 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1662 if (code == 0 && !stable && (bp->flags & CM_BUF_DIRTY)) {
1664 * we must stabilize the object to ensure that buffer
1665 * changes cannot occur while the flush is performed.
1666 * However, we do not want to Stabilize if we do not
1667 * need to because Stabilize obtains a callback.
1669 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1670 stable = (code == 0);
1673 if (code == CM_ERROR_BADFD) {
1674 /* if the scp's FID is bad its because we received VNOVNODE
1675 * when attempting to FetchStatus before the write. This
1676 * page therefore contains data that can no longer be stored.
1678 lock_ObtainMutex(&bp->mx);
1679 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
1680 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1681 bp->error = CM_ERROR_BADFD;
1682 bp->dirty_offset = 0;
1683 bp->dirty_length = 0;
1684 bp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1686 lock_ReleaseMutex(&bp->mx);
1687 } else if (!(scp->flags & CM_SCACHEFLAG_RO)) {
1692 lock_ObtainMutex(&bp->mx);
1694 /* start cleaning the buffer, and wait for it to finish */
1695 buf_CleanAsyncLocked(scp, bp, reqp, 0, NULL);
1696 buf_WaitIO(scp, bp);
1698 lock_ReleaseMutex(&bp->mx);
1701 /* actually, we only know that buffer is clean if ref
1702 * count is 1, since we don't have buffer itself locked.
1704 if (!(bp->flags & CM_BUF_DIRTY)) {
1705 lock_ObtainWrite(&buf_globalLock);
1706 if (bp->refCount == 1) { /* bp is held above */
1707 nbp = bp->fileHashp;
1709 buf_HoldLocked(nbp);
1710 buf_ReleaseLocked(bp, TRUE);
1714 lock_ReleaseWrite(&buf_globalLock);
1720 lock_ObtainRead(&buf_globalLock);
1721 nbp = bp->fileHashp;
1723 buf_HoldLocked(nbp);
1724 buf_ReleaseLocked(bp, FALSE);
1725 lock_ReleaseRead(&buf_globalLock);
1727 } /* for loop over a bunch of buffers */
1730 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1733 buf_ValidateBufQueues();
1734 #endif /* TESTING */
1740 /* Must be called with scp->rw held */
1741 long buf_ForceDataVersion(cm_scache_t * scp, afs_uint64 fromVersion, afs_uint64 toVersion)
1747 lock_AssertAny(&scp->rw);
1749 i = BUF_FILEHASH(&scp->fid);
1751 lock_ObtainRead(&buf_globalLock);
1753 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
1754 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1755 if (bp->dataVersion == fromVersion) {
1756 bp->dataVersion = toVersion;
1761 lock_ReleaseRead(&buf_globalLock);
1769 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1773 cm_buf_t *bp; /* buffer we're hacking on */
1774 cm_buf_t *nbp; /* next one */
1777 i = BUF_FILEHASH(&scp->fid);
1779 lock_ObtainRead(&buf_globalLock);
1780 bp = cm_data.buf_fileHashTablepp[i];
1783 lock_ReleaseRead(&buf_globalLock);
1784 for (; bp; bp = nbp) {
1785 /* clean buffer synchronously */
1786 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1787 lock_ObtainMutex(&bp->mx);
1788 if (bp->flags & CM_BUF_DIRTY) {
1789 if (userp && userp != bp->userp) {
1792 cm_ReleaseUser(bp->userp);
1797 case CM_ERROR_NOSUCHFILE:
1798 case CM_ERROR_BADFD:
1799 case CM_ERROR_NOACCESS:
1800 case CM_ERROR_QUOTA:
1801 case CM_ERROR_SPACE:
1802 case CM_ERROR_TOOBIG:
1803 case CM_ERROR_READONLY:
1804 case CM_ERROR_NOSUCHPATH:
1806 * Apply the previous fatal error to this buffer.
1807 * Do not waste the time attempting to store to
1808 * the file server when we know it will fail.
1810 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
1811 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1812 bp->dirty_offset = 0;
1813 bp->dirty_length = 0;
1815 bp->dataVersion = CM_BUF_VERSION_BAD;
1818 case CM_ERROR_TIMEDOUT:
1819 case CM_ERROR_ALLDOWN:
1820 case CM_ERROR_ALLBUSY:
1821 case CM_ERROR_ALLOFFLINE:
1822 case CM_ERROR_CLOCKSKEW:
1823 /* do not mark the buffer in error state but do
1824 * not attempt to complete the rest either.
1828 code = buf_CleanAsyncLocked(scp, bp, reqp, 0, &wasDirty);
1829 if (bp->flags & CM_BUF_ERROR) {
1835 buf_CleanWait(scp, bp, TRUE);
1837 lock_ReleaseMutex(&bp->mx);
1840 lock_ObtainRead(&buf_globalLock);
1841 nbp = bp->fileHashp;
1843 buf_HoldLocked(nbp);
1844 buf_ReleaseLocked(bp, FALSE);
1845 lock_ReleaseRead(&buf_globalLock);
1846 } /* for loop over a bunch of buffers */
1849 buf_ValidateBufQueues();
1850 #endif /* TESTING */
1858 buf_ValidateBufQueues(void)
1860 cm_buf_t * bp, *bpb, *bpf, *bpa;
1861 afs_uint32 countf=0, countb=0, counta=0;
1863 lock_ObtainRead(&buf_globalLock);
1864 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1865 if (bp->magic != CM_BUF_MAGIC)
1866 osi_panic("buf magic error",__FILE__,__LINE__);
1871 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1872 if (bp->magic != CM_BUF_MAGIC)
1873 osi_panic("buf magic error",__FILE__,__LINE__);
1878 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1879 if (bp->magic != CM_BUF_MAGIC)
1880 osi_panic("buf magic error",__FILE__,__LINE__);
1884 lock_ReleaseRead(&buf_globalLock);
1886 if (countb != countf)
1887 osi_panic("buf magic error",__FILE__,__LINE__);
1889 if (counta != cm_data.buf_nbuffers)
1890 osi_panic("buf magic error",__FILE__,__LINE__);
1892 #endif /* TESTING */
1894 /* dump the contents of the buf_scacheHashTablepp. */
1895 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1902 if (cm_data.buf_scacheHashTablepp == NULL)
1906 lock_ObtainRead(&buf_globalLock);
1908 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n",
1909 cookie, cm_data.buf_hashSize);
1910 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1912 for (i = 0; i < cm_data.buf_hashSize; i++)
1914 for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp)
1916 StringCbPrintfA(output, sizeof(output),
1917 "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
1918 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1919 "flags=0x%x, qFlags=0x%x cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1920 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1921 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1922 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1923 bp->cmFlags, bp->error, bp->refCount);
1924 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1928 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
1929 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1931 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
1932 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1933 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1934 StringCbPrintfA(output, sizeof(output),
1935 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1936 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1937 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1938 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1939 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1940 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1941 bp->cmFlags, bp->error, bp->refCount);
1942 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1944 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
1945 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1947 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListp\r\n", cookie);
1948 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1949 for(bp = cm_data.buf_dirtyListp; bp; bp=bp->dirtyp) {
1950 StringCbPrintfA(output, sizeof(output),
1951 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1952 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1953 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1954 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1955 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1956 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1957 bp->cmFlags, bp->error, bp->refCount);
1958 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1960 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListp.\r\n", cookie);
1961 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1964 lock_ReleaseRead(&buf_globalLock);
1968 void buf_ForceTrace(BOOL flush)
1977 len = GetTempPath(sizeof(buf)-10, buf);
1978 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1979 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1980 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1981 if (handle == INVALID_HANDLE_VALUE) {
1982 osi_panic("Cannot create log file", __FILE__, __LINE__);
1984 osi_LogPrint(buf_logp, handle);
1986 FlushFileBuffers(handle);
1987 CloseHandle(handle);
1990 long buf_DirtyBuffersExist(cm_fid_t *fidp)
1993 afs_uint32 bcount = 0;
1996 i = BUF_FILEHASH(fidp);
1998 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
1999 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY))
2006 long buf_CleanDirtyBuffers(cm_scache_t *scp)
2009 afs_uint32 bcount = 0;
2010 cm_fid_t * fidp = &scp->fid;
2012 for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
2013 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
2015 lock_ObtainMutex(&bp->mx);
2016 _InterlockedAnd(&bp->cmFlags, ~CM_BUF_CMSTORING);
2017 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2018 bp->dirty_offset = 0;
2019 bp->dirty_length = 0;
2020 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2021 bp->error = VNOVNODE;
2022 bp->dataVersion = CM_BUF_VERSION_BAD; /* bad */
2024 if (bp->flags & CM_BUF_WAITING) {
2025 osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
2026 osi_Wakeup((long) &bp);
2028 lock_ReleaseMutex(&bp->mx);