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_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
169 /* watch for transition from empty to one element */
170 if (!cm_data.buf_freeListEndp)
171 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
172 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
176 lock_ConvertWToR(&buf_globalLock);
180 /* release a buffer. Buffer must be referenced, but unlocked. */
181 #ifdef DEBUG_REFCOUNT
182 void buf_ReleaseDbg(cm_buf_t *bp, char *file, long line)
184 void buf_Release(cm_buf_t *bp)
189 /* ensure that we're in the LRU queue if our ref count is 0 */
190 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
192 refCount = InterlockedDecrement(&bp->refCount);
193 #ifdef DEBUG_REFCOUNT
194 osi_Log2(afsd_logp,"buf_Release bp 0x%p ref %d", bp, refCount);
195 afsi_log("%s:%d buf_ReleaseLocked bp 0x%p, ref %d", file, line, bp, refCount);
199 osi_panic("buf refcount 0",__FILE__,__LINE__);;
201 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
204 lock_ObtainWrite(&buf_globalLock);
205 if (bp->refCount == 0 &&
206 !(bp->qFlags & CM_BUF_QINLRU)) {
207 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
209 /* watch for transition from empty to one element */
210 if (!cm_data.buf_freeListEndp)
211 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
212 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
214 lock_ReleaseWrite(&buf_globalLock);
219 buf_Sync(int quitOnShutdown)
221 cm_buf_t **bpp, *bp, *prevbp;
222 afs_uint32 wasDirty = 0;
225 /* go through all of the dirty buffers */
226 lock_ObtainRead(&buf_globalLock);
227 for (bpp = &cm_data.buf_dirtyListp, prevbp = NULL; bp = *bpp; ) {
228 if (quitOnShutdown && buf_ShutdownFlag)
231 lock_ReleaseRead(&buf_globalLock);
232 /* all dirty buffers are held when they are added to the
233 * dirty list. No need for an additional hold.
235 lock_ObtainMutex(&bp->mx);
237 if (bp->flags & CM_BUF_DIRTY && !(bp->qFlags & CM_BUF_QREDIR)) {
238 /* start cleaning the buffer; don't touch log pages since
239 * the log code counts on knowing exactly who is writing
240 * a log page at any given instant.
242 * only attempt to write the buffer if the volume might
248 volp = cm_GetVolumeByFID(&bp->fid);
249 switch (cm_GetVolumeStatus(volp, bp->fid.volume)) {
253 req.flags |= CM_REQ_NORETRY;
254 buf_CleanAsyncLocked(NULL, bp, &req, 0, &dirty);
260 /* the buffer may or may not have been dirty
261 * and if dirty may or may not have been cleaned
262 * successfully. check the dirty flag again.
264 if (!(bp->flags & CM_BUF_DIRTY)) {
265 /* remove the buffer from the dirty list */
266 lock_ObtainWrite(&buf_globalLock);
267 #ifdef DEBUG_REFCOUNT
268 if (bp->dirtyp == NULL && bp != cm_data.buf_dirtyListEndp) {
269 osi_Log1(afsd_logp,"buf_IncrSyncer bp 0x%p list corruption",bp);
270 afsi_log("buf_IncrSyncer bp 0x%p list corruption", bp);
275 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINDL);
276 if (cm_data.buf_dirtyListp == NULL)
277 cm_data.buf_dirtyListEndp = NULL;
278 else if (cm_data.buf_dirtyListEndp == bp)
279 cm_data.buf_dirtyListEndp = prevbp;
280 buf_ReleaseLocked(bp, TRUE);
281 lock_ConvertWToR(&buf_globalLock);
283 if (buf_ShutdownFlag) {
286 char volstr[VL_MAXNAMELEN+12]="";
289 volp = cm_GetVolumeByFID(&bp->fid);
292 if (bp->fid.volume == volp->vol[RWVOL].ID)
294 else if (bp->fid.volume == volp->vol[ROVOL].ID)
296 else if (bp->fid.volume == volp->vol[BACKVOL].ID)
300 snprintf(volstr, sizeof(volstr), "%s%s", volp->namep, ext);
302 cellp = cm_FindCellByID(bp->fid.cell, CM_FLAG_NOPROBE);
303 snprintf(volstr, sizeof(volstr), "%u", bp->fid.volume);
306 LogEvent(EVENTLOG_INFORMATION_TYPE, MSG_DIRTY_BUFFER_AT_SHUTDOWN,
307 cellp->name, volstr, bp->fid.vnode, bp->fid.unique,
308 bp->offset.QuadPart+bp->dirty_offset, bp->dirty_length);
311 /* advance the pointer so we don't loop forever */
312 lock_ObtainRead(&buf_globalLock);
316 lock_ReleaseMutex(&bp->mx);
317 } /* for loop over a bunch of buffers */
318 lock_ReleaseRead(&buf_globalLock);
323 /* incremental sync daemon. Writes all dirty buffers every 5000 ms */
324 void buf_IncrSyncer(long parm)
329 while (buf_ShutdownFlag == 0) {
332 i = SleepEx(5000, 1);
339 wasDirty = buf_Sync(1);
340 } /* whole daemon's while loop */
344 buf_ValidateBuffers(void)
346 cm_buf_t * bp, *bpf, *bpa, *bpb;
347 afs_uint64 countb = 0, countf = 0, counta = 0;
349 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
350 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
351 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
352 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
356 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
357 if (bp->magic != CM_BUF_MAGIC) {
358 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
359 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
365 if (countb > cm_data.buf_nbuffers) {
366 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
367 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
372 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
373 if (bp->magic != CM_BUF_MAGIC) {
374 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
375 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
381 if (countf > cm_data.buf_nbuffers) {
382 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
383 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
388 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
389 if (bp->magic != CM_BUF_MAGIC) {
390 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
391 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
397 if (counta > cm_data.buf_nbuffers) {
398 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
399 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
404 if (countb != countf) {
405 afsi_log("cm_ValidateBuffers failure: countb != countf");
406 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
410 if (counta != cm_data.buf_nbuffers) {
411 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
412 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
419 void buf_Shutdown(void)
421 /* disable the buf_IncrSyncer() threads */
422 buf_ShutdownFlag = 1;
424 /* then force all dirty buffers to the file servers */
428 /* initialize the buffer package; called with no locks
429 * held during the initialization phase.
431 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
433 static osi_once_t once;
442 cm_data.buf_nbuffers = nbuffers;
444 /* Have to be able to reserve a whole chunk */
445 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
446 return CM_ERROR_TOOFEWBUFS;
449 /* recall for callouts */
452 if (osi_Once(&once)) {
453 /* initialize global locks */
454 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock", LOCK_HIERARCHY_BUF_GLOBAL);
457 /* remember this for those who want to reset it */
458 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
460 /* lower hash size to a prime number */
461 cm_data.buf_hashSize = osi_PrimeLessThan((afs_uint32)(cm_data.buf_nbuffers/7 + 1));
463 /* create hash table */
464 memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
466 /* another hash table */
467 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
469 /* create buffer headers and put in free list */
470 bp = cm_data.bufHeaderBaseAddress;
471 data = cm_data.bufDataBaseAddress;
472 cm_data.buf_allp = NULL;
474 for (i=0; i<cm_data.buf_nbuffers; i++) {
475 osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress,
476 "invalid cm_buf_t address");
477 osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
478 "invalid cm_buf_t data address");
480 /* allocate and zero some storage */
481 memset(bp, 0, sizeof(cm_buf_t));
482 bp->magic = CM_BUF_MAGIC;
483 /* thread on list of all buffers */
484 bp->allp = cm_data.buf_allp;
485 cm_data.buf_allp = bp;
487 osi_QAdd((osi_queue_t **)&cm_data.buf_freeListp, &bp->q);
488 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
489 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
491 /* grab appropriate number of bytes from aligned zone */
494 /* setup last buffer pointer */
496 cm_data.buf_freeListEndp = bp;
500 data += cm_data.buf_blockSize;
503 /* none reserved at first */
504 cm_data.buf_reservedBufs = 0;
506 /* just for safety's sake */
507 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
509 bp = cm_data.bufHeaderBaseAddress;
510 data = cm_data.bufDataBaseAddress;
512 for (i=0; i<cm_data.buf_nbuffers; i++) {
513 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
516 bp->waitRequests = 0;
517 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
523 buf_ValidateBufQueues();
527 /* init the buffer trace log */
528 buf_logp = osi_LogCreate("buffer", 1000);
529 osi_LogEnable(buf_logp);
534 /* and create the incr-syncer */
535 phandle = thrd_Create(0, 0,
536 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
539 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
540 CloseHandle(phandle);
544 buf_ValidateBufQueues();
549 /* add nbuffers to the buffer pool, if possible.
550 * Called with no locks held.
552 long buf_AddBuffers(afs_uint64 nbuffers)
554 /* The size of a virtual cache cannot be changed after it has
555 * been created. Subsequent calls to MapViewofFile() with
556 * an existing mapping object name would not allow the
557 * object to be resized. Return failure immediately.
559 * A similar problem now occurs with the persistent cache
560 * given that the memory mapped file now contains a complex
563 afsi_log("request to add %d buffers to the existing cache of size %d denied",
564 nbuffers, cm_data.buf_nbuffers);
566 return CM_ERROR_INVAL;
569 /* interface to set the number of buffers to an exact figure.
570 * Called with no locks held.
572 long buf_SetNBuffers(afs_uint64 nbuffers)
575 return CM_ERROR_INVAL;
576 if (nbuffers == cm_data.buf_nbuffers)
578 else if (nbuffers > cm_data.buf_nbuffers)
579 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
581 return CM_ERROR_INVAL;
584 /* wait for reading or writing to clear; called with write-locked
585 * buffer and unlocked scp and returns with locked buffer.
587 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
592 osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
593 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
596 /* if no IO is happening, we're done */
597 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
600 /* otherwise I/O is happening, but some other thread is waiting for
601 * the I/O already. Wait for that guy to figure out what happened,
602 * and then check again.
604 if ( bp->flags & CM_BUF_WAITING ) {
607 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
609 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
610 _InterlockedOr(&bp->flags, CM_BUF_WAITING);
611 bp->waitCount = bp->waitRequests = 1;
613 osi_SleepM((LONG_PTR)bp, &bp->mx);
615 cm_UpdateServerPriority();
617 lock_ObtainMutex(&bp->mx);
618 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
620 if (bp->waitCount == 0) {
621 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
622 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
623 bp->waitRequests = 0;
627 if (scp = cm_FindSCache(&bp->fid))
631 lock_ObtainRead(&scp->rw);
632 if (scp->flags & CM_SCACHEFLAG_WAITING) {
633 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
634 osi_Wakeup((LONG_PTR)&scp->flags);
636 lock_ReleaseRead(&scp->rw);
640 /* if we get here, the IO is done, but we may have to wakeup people waiting for
641 * the I/O to complete. Do so.
643 if (bp->flags & CM_BUF_WAITING) {
644 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
645 osi_Wakeup((LONG_PTR) bp);
647 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
650 cm_ReleaseSCache(scp);
653 /* find a buffer, if any, for a particular file ID and offset. Assumes
654 * that buf_globalLock is write locked when called.
656 cm_buf_t *buf_FindLocked(struct cm_fid *fidp, osi_hyper_t *offsetp)
661 lock_AssertAny(&buf_globalLock);
663 i = BUF_HASH(fidp, offsetp);
664 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
665 if (cm_FidCmp(fidp, &bp->fid) == 0
666 && offsetp->LowPart == bp->offset.LowPart
667 && offsetp->HighPart == bp->offset.HighPart) {
673 /* return whatever we found, if anything */
677 /* find a buffer with offset *offsetp for vnode *scp. Called
678 * with no locks held.
680 cm_buf_t *buf_Find(struct cm_fid *fidp, osi_hyper_t *offsetp)
684 lock_ObtainRead(&buf_globalLock);
685 bp = buf_FindLocked(fidp, offsetp);
686 lock_ReleaseRead(&buf_globalLock);
691 /* find a buffer, if any, for a particular file ID and offset. Assumes
692 * that buf_globalLock is write locked when called. Uses the all buffer
695 cm_buf_t *buf_FindAllLocked(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
700 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
701 if (cm_FidCmp(fidp, &bp->fid) == 0
702 && offsetp->LowPart == bp->offset.LowPart
703 && offsetp->HighPart == bp->offset.HighPart) {
709 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
710 if (cm_FidCmp(fidp, &bp->fid) == 0) {
713 fileOffset = offsetp->QuadPart + cm_data.baseAddress;
714 if (fileOffset == bp->datap) {
721 /* return whatever we found, if anything */
725 /* find a buffer with offset *offsetp for vnode *scp. Called
726 * with no locks held. Use the all buffer list.
728 cm_buf_t *buf_FindAll(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
732 lock_ObtainRead(&buf_globalLock);
733 bp = buf_FindAllLocked(fidp, offsetp, flags);
734 lock_ReleaseRead(&buf_globalLock);
739 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
742 * Makes sure that there's only one person writing this block
743 * at any given time, and also ensures that the log is forced sufficiently far,
744 * if this buffer contains logged data.
746 * Returns non-zero if the buffer was dirty.
748 * 'scp' may or may not be NULL. If it is not NULL, the FID for both cm_scache_t
749 * and cm_buf_t must match.
751 afs_uint32 buf_CleanAsyncLocked(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp,
752 afs_uint32 flags, afs_uint32 *pisdirty)
755 afs_uint32 isdirty = 0;
759 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
760 osi_assertx(scp == NULL || cm_FidCmp(&scp->fid, &bp->fid) == 0, "scp fid != bp fid");
763 * If the matching cm_scache_t was not provided as a parameter
764 * we must either find one or allocate a new one. It is possible
765 * that the cm_scache_t was recycled out of the cache even though
766 * a cm_buf_t with the same FID is in the cache.
769 if ((scp = cm_FindSCache(&bp->fid)) ||
770 (cm_GetSCache(&bp->fid, &scp,
771 bp->userp ? bp->userp : cm_rootUserp,
777 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
779 lock_ReleaseMutex(&bp->mx);
783 * If we didn't find a cm_scache_t object for bp->fid it means
784 * that we no longer have that FID in the cache. It does not
785 * mean that the object does not exist in the cell. That may
786 * in fact be the case but we don't know that until we attempt
787 * a FetchStatus on the FID.
789 osi_Log1(buf_logp, "buf_CleanAsyncLocked unable to start I/O - scp not found buf 0x%p", bp);
790 code = CM_ERROR_NOSUCHFILE;
792 osi_Log2(buf_logp, "buf_CleanAsyncLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
795 LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
796 code = (*cm_buf_opsp->Writep)(scp, &offset,
798 /* we might as well try to write all of the contiguous
799 * dirty buffers in one RPC
805 flags, bp->userp, reqp);
806 osi_Log3(buf_logp, "buf_CleanAsyncLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
808 lock_ObtainMutex(&bp->mx);
809 /* if the Write routine returns No Such File, clear the dirty flag
810 * because we aren't going to be able to write this data to the file
813 if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD || code == CM_ERROR_NOACCESS ||
814 code == CM_ERROR_QUOTA || code == CM_ERROR_SPACE || code == CM_ERROR_TOOBIG ||
815 code == CM_ERROR_READONLY || code == CM_ERROR_NOSUCHPATH){
816 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
817 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
818 bp->dirty_offset = 0;
819 bp->dirty_length = 0;
821 bp->dataVersion = CM_BUF_VERSION_BAD;
827 /* Disk cache support */
828 /* write buffer to disk cache (synchronous for now) */
829 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
830 #endif /* DISKCACHE95 */
832 /* if we get here and retries are not permitted
833 * then we need to exit this loop regardless of
834 * whether or not we were able to clear the dirty bit
836 if (reqp->flags & CM_REQ_NORETRY)
839 /* Ditto if the hardDeadTimeout or idleTimeout was reached */
840 if (code == CM_ERROR_TIMEDOUT || code == CM_ERROR_ALLDOWN ||
841 code == CM_ERROR_ALLBUSY || code == CM_ERROR_ALLOFFLINE ||
842 code == CM_ERROR_CLOCKSKEW) {
848 cm_ReleaseSCache(scp);
850 /* if someone was waiting for the I/O that just completed or failed,
853 if (bp->flags & CM_BUF_WAITING) {
854 /* turn off flags and wakeup users */
855 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
856 osi_Wakeup((LONG_PTR) bp);
865 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
866 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
867 * The buffer must already be clean, and no I/O should be happening to it.
869 void buf_Recycle(cm_buf_t *bp)
874 cm_buf_t *prevBp, *nextBp;
876 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
878 /* if we get here, we know that the buffer still has a 0 ref count,
879 * and that it is clean and has no currently pending I/O. This is
880 * the dude to return.
881 * Remember that as long as the ref count is 0, we know that we won't
882 * have any lock conflicts, so we can grab the buffer lock out of
883 * order in the locking hierarchy.
885 osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
886 bp, bp->offset.HighPart, bp->offset.LowPart);
888 osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
889 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
890 "incorrect cm_buf_t flags");
891 lock_AssertWrite(&buf_globalLock);
893 if (bp->qFlags & CM_BUF_QINHASH) {
894 /* Remove from hash */
896 i = BUF_HASH(&bp->fid, &bp->offset);
897 lbpp = &(cm_data.buf_scacheHashTablepp[i]);
898 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
903 /* we better find it */
904 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
906 *lbpp = bp->hashp; /* hash out */
909 /* Remove from file hash */
911 i = BUF_FILEHASH(&bp->fid);
912 prevBp = bp->fileHashBackp;
913 bp->fileHashBackp = NULL;
914 nextBp = bp->fileHashp;
915 bp->fileHashp = NULL;
917 prevBp->fileHashp = nextBp;
919 cm_data.buf_fileHashTablepp[i] = nextBp;
921 nextBp->fileHashBackp = prevBp;
923 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINHASH);
926 /* make the fid unrecognizable */
927 memset(&bp->fid, 0, sizeof(cm_fid_t));
929 /* clean up junk flags */
930 _InterlockedAnd(&bp->flags, ~(CM_BUF_EOF | CM_BUF_ERROR));
931 bp->dataVersion = CM_BUF_VERSION_BAD; /* unknown so far */
934 /* recycle a buffer, removing it from the free list, hashing in its new identity
935 * and returning it write-locked so that no one can use it. Called without
936 * any locks held, and can return an error if it loses the race condition and
937 * finds that someone else created the desired buffer.
939 * If success is returned, the buffer is returned write-locked.
941 * May be called with null scp and offsetp, if we're just trying to reclaim some
942 * space from the buffer pool. In that case, the buffer will be returned
943 * without being hashed into the hash table.
945 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
947 cm_buf_t *bp; /* buffer we're dealing with */
948 cm_buf_t *nextBp; /* next buffer in file hash chain */
949 afs_uint32 i; /* temp */
952 buf_ValidateBufQueues();
957 lock_ObtainRead(&scp->bufCreateLock);
958 lock_ObtainWrite(&buf_globalLock);
959 /* check to see if we lost the race */
961 if (bp = buf_FindLocked(&scp->fid, offsetp)) {
962 /* Do not call buf_ReleaseLocked() because we
963 * do not want to allow the buffer to be added
966 afs_int32 refCount = InterlockedDecrement(&bp->refCount);
967 #ifdef DEBUG_REFCOUNT
968 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, refCount);
969 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, refCount);
971 lock_ReleaseWrite(&buf_globalLock);
972 lock_ReleaseRead(&scp->bufCreateLock);
973 return CM_BUF_EXISTS;
977 /* does this fix the problem below? it's a simple solution. */
978 if (!cm_data.buf_freeListEndp)
980 lock_ReleaseWrite(&buf_globalLock);
981 lock_ReleaseRead(&scp->bufCreateLock);
982 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
987 /* for debugging, assert free list isn't empty, although we
988 * really should try waiting for a running tranasction to finish
989 * instead of this; or better, we should have a transaction
990 * throttler prevent us from entering this situation.
992 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
994 /* look at all buffers in free list, some of which may temp.
995 * have high refcounts and which then should be skipped,
996 * starting cleaning I/O for those which are dirty. If we find
997 * a clean buffer, we rehash it, lock it and return it.
999 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1000 /* check to see if it really has zero ref count. This
1001 * code can bump refcounts, at least, so it may not be
1004 if (bp->refCount > 0)
1007 /* we don't have to lock buffer itself, since the ref
1008 * count is 0 and we know it will stay zero as long as
1009 * we hold the global lock.
1012 /* Don't recycle a buffer held by the redirector. */
1013 if (bp->qFlags & CM_BUF_QREDIR)
1016 /* don't recycle someone in our own chunk */
1017 if (!cm_FidCmp(&bp->fid, &scp->fid)
1018 && (bp->offset.LowPart & (-cm_chunkSize))
1019 == (offsetp->LowPart & (-cm_chunkSize)))
1022 /* if this page is being filled (!) or cleaned, see if
1023 * the I/O has completed. If not, skip it, otherwise
1024 * do the final processing for the I/O.
1026 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
1027 /* probably shouldn't do this much work while
1028 * holding the big lock? Watch for contention
1034 if (bp->flags & CM_BUF_DIRTY) {
1035 /* if the buffer is dirty, start cleaning it and
1036 * move on to the next buffer. We do this with
1037 * just the lock required to minimize contention
1041 lock_ReleaseWrite(&buf_globalLock);
1042 lock_ReleaseRead(&scp->bufCreateLock);
1044 /* grab required lock and clean; this only
1045 * starts the I/O. By the time we're back,
1046 * it'll still be marked dirty, but it will also
1047 * have the WRITING flag set, so we won't get
1050 if (cm_FidCmp(&scp->fid, &bp->fid) == 0)
1051 buf_CleanAsync(scp, bp, reqp, 0, NULL);
1053 buf_CleanAsync(NULL, bp, reqp, 0, NULL);
1055 /* now put it back and go around again */
1060 /* if we get here, we know that the buffer still has a 0
1061 * ref count, and that it is clean and has no currently
1062 * pending I/O. This is the dude to return.
1063 * Remember that as long as the ref count is 0, we know
1064 * that we won't have any lock conflicts, so we can grab
1065 * the buffer lock out of order in the locking hierarchy.
1069 /* now hash in as our new buffer, and give it the
1070 * appropriate label, if requested.
1073 lock_AssertWrite(&buf_globalLock);
1075 _InterlockedOr(&bp->qFlags, CM_BUF_QINHASH);
1080 bp->offset = *offsetp;
1081 i = BUF_HASH(&scp->fid, offsetp);
1082 bp->hashp = cm_data.buf_scacheHashTablepp[i];
1083 cm_data.buf_scacheHashTablepp[i] = bp;
1084 i = BUF_FILEHASH(&scp->fid);
1085 nextBp = cm_data.buf_fileHashTablepp[i];
1086 bp->fileHashp = nextBp;
1087 bp->fileHashBackp = NULL;
1089 nextBp->fileHashBackp = bp;
1090 cm_data.buf_fileHashTablepp[i] = bp;
1093 /* we should move it from the lru queue. It better still be there,
1094 * since we've held the global (big) lock since we found it there.
1096 osi_assertx(bp->qFlags & CM_BUF_QINLRU,
1097 "buf_GetNewLocked: LRU screwup");
1099 if (cm_data.buf_freeListEndp == bp) {
1100 /* we're the last guy in this queue, so maintain it */
1101 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1103 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1104 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1106 /* prepare to return it. Give it a refcount */
1108 #ifdef DEBUG_REFCOUNT
1109 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, 1);
1110 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, 1);
1112 /* grab the mutex so that people don't use it
1113 * before the caller fills it with data. Again, no one
1114 * should have been able to get to this dude to lock it.
1116 if (!lock_TryMutex(&bp->mx)) {
1117 osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked. refCount %d should be 0",
1119 osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
1122 lock_ReleaseWrite(&buf_globalLock);
1123 lock_ReleaseRead(&scp->bufCreateLock);
1128 buf_ValidateBufQueues();
1129 #endif /* TESTING */
1131 } /* for all buffers in lru queue */
1132 lock_ReleaseWrite(&buf_globalLock);
1133 lock_ReleaseRead(&scp->bufCreateLock);
1134 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List has no buffers with a zero refcount - sleeping 100ms");
1135 Sleep(100); /* give some time for a buffer to be freed */
1136 } /* while loop over everything */
1141 * get a page, returning it held but unlocked. the page may or may not
1142 * contain valid data.
1144 * The scp must be unlocked when passed in unlocked.
1146 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
1150 osi_hyper_t pageOffset;
1151 unsigned long tcount;
1155 cm_diskcache_t *dcp;
1156 #endif /* DISKCACHE95 */
1159 pageOffset.HighPart = offsetp->HighPart;
1160 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
1164 buf_ValidateBufQueues();
1165 #endif /* TESTING */
1167 bp = buf_Find(&scp->fid, &pageOffset);
1169 /* lock it and break out */
1170 lock_ObtainMutex(&bp->mx);
1173 /* touch disk chunk to update LRU info */
1174 diskcache_Touch(bp->dcp);
1175 #endif /* DISKCACHE95 */
1179 /* otherwise, we have to create a page */
1180 code = buf_GetNewLocked(scp, &pageOffset, reqp, &bp);
1183 /* the requested buffer was created */
1188 * the requested buffer existed by the time the
1189 * scp->bufCreateLock and buf_globalLock could be obtained.
1190 * loop again and permit buf_Find() to obtain a reference.
1195 * the requested buffer could not be created.
1196 * return the error to the caller.
1199 buf_ValidateBufQueues();
1200 #endif /* TESTING */
1203 } /* big while loop */
1205 /* if we get here, we have a locked buffer that may have just been
1206 * created, in which case it needs to be filled with data.
1209 /* load the page; freshly created pages should be idle */
1210 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
1213 * start the I/O; may drop lock. as of this writing, the only
1214 * implementation of Readp is cm_BufRead() which simply sets
1215 * tcount to 0 and returns success.
1217 _InterlockedOr(&bp->flags, CM_BUF_READING);
1218 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
1221 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
1222 bp->dcp = dcp; /* pointer to disk cache struct. */
1223 #endif /* DISKCACHE95 */
1226 /* failure or queued */
1227 if (code != ERROR_IO_PENDING) {
1229 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1230 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1231 if (bp->flags & CM_BUF_WAITING) {
1232 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1233 osi_Wakeup((LONG_PTR) bp);
1235 lock_ReleaseMutex(&bp->mx);
1238 buf_ValidateBufQueues();
1239 #endif /* TESTING */
1244 * otherwise, I/O completed instantly and we're done, except
1245 * for padding the xfr out with 0s and checking for EOF
1247 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1248 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1250 _InterlockedOr(&bp->flags, CM_BUF_EOF);
1252 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1253 if (bp->flags & CM_BUF_WAITING) {
1254 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1255 osi_Wakeup((LONG_PTR) bp);
1260 /* wait for reads, either that which we started above, or that someone
1261 * else started. We don't care if we return a buffer being cleaned.
1263 if (bp->flags & CM_BUF_READING)
1264 buf_WaitIO(scp, bp);
1266 /* once it has been read once, we can unlock it and return it, still
1267 * with its refcount held.
1269 lock_ReleaseMutex(&bp->mx);
1272 /* now remove from queue; will be put in at the head (farthest from
1273 * being recycled) when we're done in buf_Release.
1275 lock_ObtainWrite(&buf_globalLock);
1276 if (bp->qFlags & CM_BUF_QINLRU) {
1277 if (cm_data.buf_freeListEndp == bp)
1278 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1279 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1280 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1282 lock_ReleaseWrite(&buf_globalLock);
1284 osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1285 bp, scp, offsetp->HighPart, offsetp->LowPart);
1287 buf_ValidateBufQueues();
1288 #endif /* TESTING */
1292 /* count # of elements in the free list;
1293 * we don't bother doing the proper locking for accessing dataVersion or flags
1294 * since it is a pain, and this is really just an advisory call. If you need
1295 * to do better at some point, rewrite this function.
1297 long buf_CountFreeList(void)
1303 lock_ObtainRead(&buf_globalLock);
1304 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1305 /* if the buffer doesn't have an identity, or if the buffer
1306 * has been invalidate (by having its DV stomped upon), then
1307 * count it as free, since it isn't really being utilized.
1309 if (!(bufp->qFlags & CM_BUF_QINHASH) || bufp->dataVersion == CM_BUF_VERSION_BAD)
1312 lock_ReleaseRead(&buf_globalLock);
1316 /* clean a buffer synchronously */
1317 afs_uint32 buf_CleanAsync(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp, afs_uint32 flags, afs_uint32 *pisdirty)
1320 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1321 osi_assertx(!(flags & CM_BUF_WRITE_SCP_LOCKED), "scp->rw must not be held when calling buf_CleanAsync");
1323 lock_ObtainMutex(&bp->mx);
1324 code = buf_CleanAsyncLocked(scp, bp, reqp, flags, pisdirty);
1325 lock_ReleaseMutex(&bp->mx);
1330 /* wait for a buffer's cleaning to finish */
1331 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp, afs_uint32 locked)
1333 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1336 lock_ObtainMutex(&bp->mx);
1337 if (bp->flags & CM_BUF_WRITING) {
1338 buf_WaitIO(scp, bp);
1341 lock_ReleaseMutex(&bp->mx);
1344 /* set the dirty flag on a buffer, and set associated write-ahead log,
1345 * if there is one. Allow one to be added to a buffer, but not changed.
1347 * The buffer must be locked before calling this routine.
1349 void buf_SetDirty(cm_buf_t *bp, afs_uint32 offset, afs_uint32 length, cm_user_t *userp)
1351 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1352 osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
1354 if (bp->flags & CM_BUF_DIRTY) {
1356 osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
1358 if (bp->dirty_offset <= offset) {
1359 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1360 /* dirty_length remains the same */
1362 bp->dirty_length = offset + length - bp->dirty_offset;
1364 } else /* bp->dirty_offset > offset */ {
1365 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1366 bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
1368 bp->dirty_length = length;
1370 bp->dirty_offset = offset;
1373 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1376 _InterlockedOr(&bp->flags, CM_BUF_DIRTY);
1378 /* and turn off EOF flag, since it has associated data now */
1379 _InterlockedAnd(&bp->flags, ~CM_BUF_EOF);
1381 bp->dirty_offset = offset;
1382 bp->dirty_length = length;
1384 /* and add to the dirty list.
1385 * we obtain a hold on the buffer for as long as it remains
1386 * in the list. buffers are only removed from the list by
1387 * the buf_IncrSyncer function regardless of when else the
1388 * dirty flag might be cleared.
1390 * This should never happen but just in case there is a bug
1391 * elsewhere, never add to the dirty list if the buffer is
1394 lock_ObtainWrite(&buf_globalLock);
1395 if (!(bp->qFlags & CM_BUF_QINDL)) {
1397 if (!cm_data.buf_dirtyListp) {
1398 cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
1400 cm_data.buf_dirtyListEndp->dirtyp = bp;
1401 cm_data.buf_dirtyListEndp = bp;
1404 _InterlockedOr(&bp->qFlags, CM_BUF_QINDL);
1406 lock_ReleaseWrite(&buf_globalLock);
1409 /* and record the last writer */
1410 if (bp->userp != userp) {
1413 cm_ReleaseUser(bp->userp);
1418 /* clean all buffers, reset log pointers and invalidate all buffers.
1419 * Called with no locks held, and returns with same.
1421 * This function is guaranteed to clean and remove the log ptr of all the
1422 * buffers that were dirty or had non-zero log ptrs before the call was
1423 * made. That's sufficient to clean up any garbage left around by recovery,
1424 * which is all we're counting on this for; there may be newly created buffers
1425 * added while we're running, but that should be OK.
1427 * In an environment where there are no transactions (artificially imposed, for
1428 * example, when switching the database to raw mode), this function is used to
1429 * make sure that all updates have been written to the disk. In that case, we don't
1430 * really require that we forget the log association between pages and logs, but
1431 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1432 * have to worry about invalidating data in the buffers.
1434 * This function is used at the end of recovery as paranoia to get the recovered
1435 * database out to disk. It removes all references to the recovery log and cleans
1438 long buf_CleanAndReset(void)
1444 lock_ObtainRead(&buf_globalLock);
1445 for(i=0; i<cm_data.buf_hashSize; i++) {
1446 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
1447 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1449 lock_ReleaseRead(&buf_globalLock);
1451 /* now no locks are held; clean buffer and go on */
1453 req.flags |= CM_REQ_NORETRY;
1455 buf_CleanAsync(NULL, bp, &req, 0, NULL);
1456 buf_CleanWait(NULL, bp, FALSE);
1458 /* relock and release buffer */
1459 lock_ObtainRead(&buf_globalLock);
1460 buf_ReleaseLocked(bp, FALSE);
1462 } /* over one bucket */
1463 } /* for loop over all hash buckets */
1466 lock_ReleaseRead(&buf_globalLock);
1469 buf_ValidateBufQueues();
1470 #endif /* TESTING */
1472 /* and we're done */
1476 /* called without global lock being held, reserves buffers for callers
1477 * that need more than one held (not locked) at once.
1479 void buf_ReserveBuffers(afs_uint64 nbuffers)
1481 lock_ObtainWrite(&buf_globalLock);
1483 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1484 cm_data.buf_reserveWaiting = 1;
1485 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1486 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1487 lock_ObtainWrite(&buf_globalLock);
1490 cm_data.buf_reservedBufs += nbuffers;
1494 lock_ReleaseWrite(&buf_globalLock);
1497 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1501 lock_ObtainWrite(&buf_globalLock);
1502 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1506 cm_data.buf_reservedBufs += nbuffers;
1509 lock_ReleaseWrite(&buf_globalLock);
1513 /* called without global lock held, releases reservation held by
1514 * buf_ReserveBuffers.
1516 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1518 lock_ObtainWrite(&buf_globalLock);
1519 cm_data.buf_reservedBufs -= nbuffers;
1520 if (cm_data.buf_reserveWaiting) {
1521 cm_data.buf_reserveWaiting = 0;
1522 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1524 lock_ReleaseWrite(&buf_globalLock);
1527 /* truncate the buffers past sizep, zeroing out the page, if we don't
1528 * end on a page boundary.
1530 * Requires cm_bufCreateLock to be write locked.
1532 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1536 cm_buf_t *nbufp; /* next buffer, if didRelease */
1542 /* assert that cm_bufCreateLock is held in write mode */
1543 lock_AssertWrite(&scp->bufCreateLock);
1545 i = BUF_FILEHASH(&scp->fid);
1547 lock_ObtainRead(&buf_globalLock);
1548 bufp = cm_data.buf_fileHashTablepp[i];
1550 lock_ReleaseRead(&buf_globalLock);
1554 buf_HoldLocked(bufp);
1555 lock_ReleaseRead(&buf_globalLock);
1557 lock_ObtainMutex(&bufp->mx);
1559 bufEnd.HighPart = 0;
1560 bufEnd.LowPart = cm_data.buf_blockSize;
1561 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1563 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1564 LargeIntegerLessThan(*sizep, bufEnd)) {
1565 buf_WaitIO(scp, bufp);
1567 lock_ObtainWrite(&scp->rw);
1569 /* make sure we have a callback (so we have the right value for
1570 * the length), and wait for it to be safe to do a truncate.
1572 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1573 CM_SCACHESYNC_NEEDCALLBACK
1574 | CM_SCACHESYNC_GETSTATUS
1575 | CM_SCACHESYNC_SETSIZE
1576 | CM_SCACHESYNC_BUFLOCKED);
1579 /* if we succeeded in our locking, and this applies to the right
1580 * file, and the truncate request overlaps the buffer either
1581 * totally or partially, then do something.
1583 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1584 && LargeIntegerLessThan(*sizep, bufEnd)) {
1587 /* destroy the buffer, turning off its dirty bit, if
1588 * we're truncating the whole buffer. Otherwise, set
1589 * the dirty bit, and clear out the tail of the buffer
1590 * if we just overlap some.
1592 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1593 /* truncating the entire page */
1594 _InterlockedAnd(&bufp->flags, ~CM_BUF_DIRTY);
1595 bufp->dirty_offset = 0;
1596 bufp->dirty_length = 0;
1597 bufp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1598 bufp->dirtyCounter++;
1601 /* don't set dirty, since dirty implies
1602 * currently up-to-date. Don't need to do this,
1603 * since we'll update the length anyway.
1605 * Zero out remainder of the page, in case we
1606 * seek and write past EOF, and make this data
1609 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1610 osi_assertx(bufferPos != 0, "non-zero bufferPos");
1611 memset(bufp->datap + bufferPos, 0,
1612 cm_data.buf_blockSize - bufferPos);
1616 cm_SyncOpDone( scp, bufp,
1617 CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
1618 | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
1620 lock_ReleaseWrite(&scp->rw);
1621 lock_ReleaseMutex(&bufp->mx);
1624 nbufp = bufp->fileHashp;
1628 /* This forces the loop to end and the error code
1629 * to be returned. */
1637 buf_ValidateBufQueues();
1638 #endif /* TESTING */
1644 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1647 cm_buf_t *bp; /* buffer we're hacking on */
1651 afs_uint32 stable = 0;
1653 i = BUF_FILEHASH(&scp->fid);
1656 lock_ObtainRead(&buf_globalLock);
1657 bp = cm_data.buf_fileHashTablepp[i];
1660 lock_ReleaseRead(&buf_globalLock);
1662 for (; bp; bp = nbp) {
1663 didRelease = 0; /* haven't released this buffer yet */
1665 /* clean buffer synchronously */
1666 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1668 if (code == 0 && !stable && (bp->flags & CM_BUF_DIRTY)) {
1670 * we must stabilize the object to ensure that buffer
1671 * changes cannot occur while the flush is performed.
1672 * However, we do not want to Stabilize if we do not
1673 * need to because Stabilize obtains a callback.
1675 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1676 stable = (code == 0);
1679 if (code == CM_ERROR_BADFD) {
1680 /* if the scp's FID is bad its because we received VNOVNODE
1681 * when attempting to FetchStatus before the write. This
1682 * page therefore contains data that can no longer be stored.
1684 lock_ObtainMutex(&bp->mx);
1685 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
1686 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1687 bp->error = CM_ERROR_BADFD;
1688 bp->dirty_offset = 0;
1689 bp->dirty_length = 0;
1690 bp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1692 lock_ReleaseMutex(&bp->mx);
1693 } else if (!(scp->flags & CM_SCACHEFLAG_RO)) {
1698 lock_ObtainMutex(&bp->mx);
1700 /* start cleaning the buffer, and wait for it to finish */
1701 buf_CleanAsyncLocked(scp, bp, reqp, 0, NULL);
1702 buf_WaitIO(scp, bp);
1704 lock_ReleaseMutex(&bp->mx);
1707 /* actually, we only know that buffer is clean if ref
1708 * count is 1, since we don't have buffer itself locked.
1710 if (!(bp->flags & CM_BUF_DIRTY)) {
1711 lock_ObtainWrite(&buf_globalLock);
1712 if (bp->refCount == 1) { /* bp is held above */
1713 nbp = bp->fileHashp;
1715 buf_HoldLocked(nbp);
1716 buf_ReleaseLocked(bp, TRUE);
1720 lock_ReleaseWrite(&buf_globalLock);
1726 lock_ObtainRead(&buf_globalLock);
1727 nbp = bp->fileHashp;
1729 buf_HoldLocked(nbp);
1730 buf_ReleaseLocked(bp, FALSE);
1731 lock_ReleaseRead(&buf_globalLock);
1733 } /* for loop over a bunch of buffers */
1736 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1739 buf_ValidateBufQueues();
1740 #endif /* TESTING */
1746 /* Must be called with scp->rw held */
1747 long buf_ForceDataVersion(cm_scache_t * scp, afs_uint64 fromVersion, afs_uint64 toVersion)
1753 lock_AssertAny(&scp->rw);
1755 i = BUF_FILEHASH(&scp->fid);
1757 lock_ObtainRead(&buf_globalLock);
1759 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
1760 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1761 if (bp->dataVersion == fromVersion) {
1762 bp->dataVersion = toVersion;
1767 lock_ReleaseRead(&buf_globalLock);
1775 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1779 cm_buf_t *bp; /* buffer we're hacking on */
1780 cm_buf_t *nbp; /* next one */
1783 i = BUF_FILEHASH(&scp->fid);
1785 lock_ObtainRead(&buf_globalLock);
1786 bp = cm_data.buf_fileHashTablepp[i];
1789 lock_ReleaseRead(&buf_globalLock);
1790 for (; bp; bp = nbp) {
1791 /* clean buffer synchronously */
1792 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1793 lock_ObtainMutex(&bp->mx);
1794 if (bp->flags & CM_BUF_DIRTY) {
1795 if (userp && userp != bp->userp) {
1798 cm_ReleaseUser(bp->userp);
1803 case CM_ERROR_NOSUCHFILE:
1804 case CM_ERROR_BADFD:
1805 case CM_ERROR_NOACCESS:
1806 case CM_ERROR_QUOTA:
1807 case CM_ERROR_SPACE:
1808 case CM_ERROR_TOOBIG:
1809 case CM_ERROR_READONLY:
1810 case CM_ERROR_NOSUCHPATH:
1812 * Apply the previous fatal error to this buffer.
1813 * Do not waste the time attempting to store to
1814 * the file server when we know it will fail.
1816 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
1817 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1818 bp->dirty_offset = 0;
1819 bp->dirty_length = 0;
1821 bp->dataVersion = CM_BUF_VERSION_BAD;
1824 case CM_ERROR_TIMEDOUT:
1825 case CM_ERROR_ALLDOWN:
1826 case CM_ERROR_ALLBUSY:
1827 case CM_ERROR_ALLOFFLINE:
1828 case CM_ERROR_CLOCKSKEW:
1829 /* do not mark the buffer in error state but do
1830 * not attempt to complete the rest either.
1834 code = buf_CleanAsyncLocked(scp, bp, reqp, 0, &wasDirty);
1835 if (bp->flags & CM_BUF_ERROR) {
1841 buf_CleanWait(scp, bp, TRUE);
1843 lock_ReleaseMutex(&bp->mx);
1846 lock_ObtainRead(&buf_globalLock);
1847 nbp = bp->fileHashp;
1849 buf_HoldLocked(nbp);
1850 buf_ReleaseLocked(bp, FALSE);
1851 lock_ReleaseRead(&buf_globalLock);
1852 } /* for loop over a bunch of buffers */
1855 buf_ValidateBufQueues();
1856 #endif /* TESTING */
1864 buf_ValidateBufQueues(void)
1866 cm_buf_t * bp, *bpb, *bpf, *bpa;
1867 afs_uint32 countf=0, countb=0, counta=0;
1869 lock_ObtainRead(&buf_globalLock);
1870 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1871 if (bp->magic != CM_BUF_MAGIC)
1872 osi_panic("buf magic error",__FILE__,__LINE__);
1877 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1878 if (bp->magic != CM_BUF_MAGIC)
1879 osi_panic("buf magic error",__FILE__,__LINE__);
1884 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1885 if (bp->magic != CM_BUF_MAGIC)
1886 osi_panic("buf magic error",__FILE__,__LINE__);
1890 lock_ReleaseRead(&buf_globalLock);
1892 if (countb != countf)
1893 osi_panic("buf magic error",__FILE__,__LINE__);
1895 if (counta != cm_data.buf_nbuffers)
1896 osi_panic("buf magic error",__FILE__,__LINE__);
1898 #endif /* TESTING */
1900 /* dump the contents of the buf_scacheHashTablepp. */
1901 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1908 if (cm_data.buf_scacheHashTablepp == NULL)
1912 lock_ObtainRead(&buf_globalLock);
1914 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n",
1915 cookie, cm_data.buf_hashSize);
1916 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1918 for (i = 0; i < cm_data.buf_hashSize; i++)
1920 for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp)
1922 StringCbPrintfA(output, sizeof(output),
1923 "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
1924 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1925 "flags=0x%x, qFlags=0x%x cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1926 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1927 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1928 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1929 bp->cmFlags, bp->error, bp->refCount);
1930 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1934 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
1935 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1937 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
1938 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1939 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1940 StringCbPrintfA(output, sizeof(output),
1941 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1942 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1943 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1944 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1945 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1946 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1947 bp->cmFlags, bp->error, bp->refCount);
1948 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1950 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
1951 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1953 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListp\r\n", cookie);
1954 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1955 for(bp = cm_data.buf_dirtyListp; bp; bp=bp->dirtyp) {
1956 StringCbPrintfA(output, sizeof(output),
1957 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1958 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1959 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1960 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1961 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1962 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1963 bp->cmFlags, bp->error, bp->refCount);
1964 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1966 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListp.\r\n", cookie);
1967 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1970 lock_ReleaseRead(&buf_globalLock);
1974 void buf_ForceTrace(BOOL flush)
1983 len = GetTempPath(sizeof(buf)-10, buf);
1984 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1985 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1986 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1987 if (handle == INVALID_HANDLE_VALUE) {
1988 osi_panic("Cannot create log file", __FILE__, __LINE__);
1990 osi_LogPrint(buf_logp, handle);
1992 FlushFileBuffers(handle);
1993 CloseHandle(handle);
1996 long buf_DirtyBuffersExist(cm_fid_t *fidp)
1999 afs_uint32 bcount = 0;
2002 i = BUF_FILEHASH(fidp);
2004 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
2005 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY))
2012 long buf_CleanDirtyBuffers(cm_scache_t *scp)
2015 afs_uint32 bcount = 0;
2016 cm_fid_t * fidp = &scp->fid;
2018 for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
2019 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
2021 lock_ObtainMutex(&bp->mx);
2022 _InterlockedAnd(&bp->cmFlags, ~CM_BUF_CMSTORING);
2023 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2024 bp->dirty_offset = 0;
2025 bp->dirty_length = 0;
2026 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2027 bp->error = VNOVNODE;
2028 bp->dataVersion = CM_BUF_VERSION_BAD; /* bad */
2030 if (bp->flags & CM_BUF_WAITING) {
2031 osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
2032 osi_Wakeup((long) &bp);
2034 lock_ReleaseMutex(&bp->mx);