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>
23 #include <hcrypto\md5.h>
26 #include "cm_memmap.h"
29 #define TRACE_BUFFER 1
32 extern void afsi_log(char *pattern, ...);
34 /* This module implements the buffer package used by the local transaction
35 * system (cm). It is initialized by calling cm_Init, which calls buf_Init;
36 * it must be initalized before any of its main routines are called.
38 * Each buffer is hashed into a hash table by file ID and offset, and if its
39 * reference count is zero, it is also in a free list.
41 * There are two locks involved in buffer processing. The global lock
42 * buf_globalLock protects all of the global variables defined in this module,
43 * the reference counts and hash pointers in the actual cm_buf_t structures,
44 * and the LRU queue pointers in the buffer structures.
46 * The mutexes in the buffer structures protect the remaining fields in the
47 * buffers, as well the data itself.
49 * The locking hierarchy here is this:
51 * - resv multiple simul. buffers reservation
52 * - lock buffer I/O flags
53 * - lock buffer's mutex
54 * - lock buf_globalLock
58 /* global debugging log */
59 osi_log_t *buf_logp = NULL;
61 /* Global lock protecting hash tables and free lists */
62 osi_rwlock_t buf_globalLock;
64 /* Global lock used to limit the number of RDR Release
65 * Extents requests to one. */
66 osi_mutex_t buf_rdrReleaseExtentsLock;
68 /* ptr to head of the free list (most recently used) and the
69 * tail (the guy to remove first). We use osi_Q* functions
70 * to put stuff in buf_freeListp, and maintain the end
74 /* a pointer to a list of all buffers, just so that we can find them
75 * easily for debugging, and for the incr syncer. Locked under
79 /* defaults setup; these variables may be manually assigned into
80 * before calling cm_Init, as a way of changing these defaults.
83 /* callouts for reading and writing data, etc */
84 cm_buf_ops_t *cm_buf_opsp;
87 /* for experimental disk caching support in Win95 client */
88 cm_buf_t *buf_diskFreeListp;
89 cm_buf_t *buf_diskFreeListEndp;
90 cm_buf_t *buf_diskAllp;
91 extern int cm_diskCacheEnabled;
92 #endif /* DISKCACHE95 */
94 /* set this to 1 when we are terminating to prevent access attempts */
95 static int buf_ShutdownFlag = 0;
98 void buf_HoldLockedDbg(cm_buf_t *bp, char *file, long line)
100 void buf_HoldLocked(cm_buf_t *bp)
105 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
106 refCount = InterlockedIncrement(&bp->refCount);
107 #ifdef DEBUG_REFCOUNT
108 osi_Log2(afsd_logp,"buf_HoldLocked bp 0x%p ref %d",bp, refCount);
109 afsi_log("%s:%d buf_HoldLocked bp 0x%p, ref %d", file, line, bp, refCount);
113 /* hold a reference to an already held buffer */
114 #ifdef DEBUG_REFCOUNT
115 void buf_HoldDbg(cm_buf_t *bp, char *file, long line)
117 void buf_Hold(cm_buf_t *bp)
122 lock_ObtainRead(&buf_globalLock);
123 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
124 refCount = InterlockedIncrement(&bp->refCount);
125 #ifdef DEBUG_REFCOUNT
126 osi_Log2(afsd_logp,"buf_Hold bp 0x%p ref %d",bp, refCount);
127 afsi_log("%s:%d buf_Hold bp 0x%p, ref %d", file, line, bp, refCount);
129 lock_ReleaseRead(&buf_globalLock);
132 /* code to drop reference count while holding buf_globalLock */
133 #ifdef DEBUG_REFCOUNT
134 void buf_ReleaseLockedDbg(cm_buf_t *bp, afs_uint32 writeLocked, char *file, long line)
136 void buf_ReleaseLocked(cm_buf_t *bp, afs_uint32 writeLocked)
142 lock_AssertWrite(&buf_globalLock);
144 lock_AssertRead(&buf_globalLock);
146 /* ensure that we're in the LRU queue if our ref count is 0 */
147 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
149 refCount = InterlockedDecrement(&bp->refCount);
150 #ifdef DEBUG_REFCOUNT
151 osi_Log3(afsd_logp,"buf_ReleaseLocked %s bp 0x%p ref %d",writeLocked?"write":"read", bp, refCount);
152 afsi_log("%s:%d buf_ReleaseLocked %s bp 0x%p, ref %d", file, line, writeLocked?"write":"read", bp, refCount);
156 osi_panic("buf refcount 0",__FILE__,__LINE__);;
158 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
162 * If we are read locked there could be a race condition
163 * with buf_Find() so we must obtain a write lock and
164 * double check that the refCount is actually zero
165 * before we remove the buffer from the LRU queue.
168 lock_ConvertRToW(&buf_globalLock);
170 if (bp->refCount == 0 &&
171 !(bp->qFlags & (CM_BUF_QINLRU|CM_BUF_QREDIR))) {
172 osi_QAddH( (osi_queue_t **) &cm_data.buf_freeListp,
173 (osi_queue_t **) &cm_data.buf_freeListEndp,
175 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
176 buf_IncrementFreeCount();
180 lock_ConvertWToR(&buf_globalLock);
184 /* release a buffer. Buffer must be referenced, but unlocked. */
185 #ifdef DEBUG_REFCOUNT
186 void buf_ReleaseDbg(cm_buf_t *bp, char *file, long line)
188 void buf_Release(cm_buf_t *bp)
193 /* ensure that we're in the LRU queue if our ref count is 0 */
194 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
196 refCount = InterlockedDecrement(&bp->refCount);
197 #ifdef DEBUG_REFCOUNT
198 osi_Log2(afsd_logp,"buf_Release bp 0x%p ref %d", bp, refCount);
199 afsi_log("%s:%d buf_ReleaseLocked bp 0x%p, ref %d", file, line, bp, refCount);
203 osi_panic("buf refcount 0",__FILE__,__LINE__);;
205 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
208 lock_ObtainWrite(&buf_globalLock);
209 if (bp->refCount == 0 &&
210 !(bp->qFlags & (CM_BUF_QINLRU|CM_BUF_QREDIR))) {
211 osi_QAddH( (osi_queue_t **) &cm_data.buf_freeListp,
212 (osi_queue_t **) &cm_data.buf_freeListEndp,
214 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
215 buf_IncrementFreeCount();
217 lock_ReleaseWrite(&buf_globalLock);
222 buf_Sync(int quitOnShutdown)
224 cm_buf_t **bpp, *bp, *prevbp;
225 afs_uint32 wasDirty = 0;
228 /* go through all of the dirty buffers */
229 lock_ObtainRead(&buf_globalLock);
230 for (bpp = &cm_data.buf_dirtyListp, prevbp = NULL; bp = *bpp; ) {
231 if (quitOnShutdown && buf_ShutdownFlag)
235 * If the buffer is held be the redirector we must fetch
236 * it back in order to determine whether or not it is in
239 if (bp->qFlags & CM_BUF_QREDIR) {
240 osi_Log1(buf_logp,"buf_Sync buffer held by redirector bp 0x%p", bp);
242 /* Request single buffer from the redirector */
243 buf_RDRShakeAnExtentFree(bp, &req);
246 lock_ReleaseRead(&buf_globalLock);
248 * all dirty buffers are held when they are added to the
249 * dirty list. No need for an additional hold.
251 lock_ObtainMutex(&bp->mx);
253 if ((bp->flags & CM_BUF_DIRTY)) {
254 /* start cleaning the buffer; don't touch log pages since
255 * the log code counts on knowing exactly who is writing
256 * a log page at any given instant.
258 * only attempt to write the buffer if the volume might
264 volp = cm_GetVolumeByFID(&bp->fid);
265 switch (cm_GetVolumeStatus(volp, bp->fid.volume)) {
269 req.flags |= CM_REQ_NORETRY;
270 buf_CleanAsyncLocked(NULL, bp, &req, 0, &dirty);
276 /* the buffer may or may not have been dirty
277 * and if dirty may or may not have been cleaned
278 * successfully. check the dirty flag again.
280 if (!(bp->flags & CM_BUF_DIRTY)) {
281 /* remove the buffer from the dirty list */
282 lock_ObtainWrite(&buf_globalLock);
283 #ifdef DEBUG_REFCOUNT
284 if (bp->dirtyp == NULL && bp != cm_data.buf_dirtyListEndp) {
285 osi_Log1(afsd_logp,"buf_Sync bp 0x%p list corruption",bp);
286 afsi_log("buf_Sync bp 0x%p list corruption", bp);
291 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINDL);
292 if (cm_data.buf_dirtyListp == NULL)
293 cm_data.buf_dirtyListEndp = NULL;
294 else if (cm_data.buf_dirtyListEndp == bp)
295 cm_data.buf_dirtyListEndp = prevbp;
296 buf_ReleaseLocked(bp, TRUE);
297 lock_ConvertWToR(&buf_globalLock);
299 if (buf_ShutdownFlag) {
302 char volstr[VL_MAXNAMELEN+12]="";
305 volp = cm_GetVolumeByFID(&bp->fid);
308 if (bp->fid.volume == volp->vol[RWVOL].ID)
310 else if (bp->fid.volume == volp->vol[ROVOL].ID)
312 else if (bp->fid.volume == volp->vol[BACKVOL].ID)
316 snprintf(volstr, sizeof(volstr), "%s%s", volp->namep, ext);
318 cellp = cm_FindCellByID(bp->fid.cell, CM_FLAG_NOPROBE);
319 snprintf(volstr, sizeof(volstr), "%u", bp->fid.volume);
322 LogEvent(EVENTLOG_INFORMATION_TYPE, MSG_DIRTY_BUFFER_AT_SHUTDOWN,
323 cellp->name, volstr, bp->fid.vnode, bp->fid.unique,
324 bp->offset.QuadPart+bp->dirty_offset, bp->dirty_length);
327 /* advance the pointer so we don't loop forever */
328 lock_ObtainRead(&buf_globalLock);
332 lock_ReleaseMutex(&bp->mx);
333 } /* for loop over a bunch of buffers */
334 lock_ReleaseRead(&buf_globalLock);
339 /* incremental sync daemon. Writes all dirty buffers every 5000 ms */
340 void buf_IncrSyncer(long parm)
345 while (buf_ShutdownFlag == 0) {
347 i = SleepEx(5000, 1);
354 wasDirty = buf_Sync(1);
355 } /* whole daemon's while loop */
359 buf_ValidateBuffers(void)
361 cm_buf_t * bp, *bpf, *bpa, *bpb;
362 afs_uint64 countb = 0, countf = 0, counta = 0, countr = 0;
364 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
365 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
366 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
367 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
371 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
372 if (bp->magic != CM_BUF_MAGIC) {
373 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
374 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
380 if (countb > cm_data.buf_nbuffers) {
381 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
382 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
387 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
388 if (bp->magic != CM_BUF_MAGIC) {
389 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
390 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
396 if (countf > cm_data.buf_nbuffers) {
397 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
398 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
403 for ( bp = cm_data.buf_redirListp; bp; bp = (cm_buf_t *) osi_QNext(&bp->q)) {
404 if (!(bp->qFlags & CM_BUF_QREDIR)) {
405 afsi_log("CM_BUF_QREDIR not set on cm_buf_t in buf_redirListp");
406 fprintf(stderr, "CM_BUF_QREDIR not set on cm_buf_t in buf_redirListp");
410 if (countr > cm_data.buf_nbuffers) {
411 afsi_log("cm_ValidateBuffers failure: countr > cm_data.buf_nbuffers");
412 fprintf(stderr, "cm_ValidateBuffers failure: countr > cm_data.buf_nbuffers\n");
417 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
418 if (bp->magic != CM_BUF_MAGIC) {
419 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
420 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
426 if (counta > cm_data.buf_nbuffers) {
427 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
428 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
433 if (countb != countf) {
434 afsi_log("cm_ValidateBuffers failure: countb != countf");
435 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
439 if (counta != cm_data.buf_nbuffers) {
440 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
441 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
448 void buf_Shutdown(void)
450 /* disable the buf_IncrSyncer() threads */
451 buf_ShutdownFlag = 1;
453 /* then force all dirty buffers to the file servers */
457 /* initialize the buffer package; called with no locks
458 * held during the initialization phase.
460 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
462 static osi_once_t once;
471 cm_data.buf_nbuffers = nbuffers;
473 /* Have to be able to reserve a whole chunk */
474 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
475 return CM_ERROR_TOOFEWBUFS;
478 /* recall for callouts */
481 if (osi_Once(&once)) {
482 /* initialize global locks */
483 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock", LOCK_HIERARCHY_BUF_GLOBAL);
484 lock_InitializeMutex(&buf_rdrReleaseExtentsLock, "RDR Release Extents lock", LOCK_HIERARCHY_RDR_EXTENTS);
487 /* remember this for those who want to reset it */
488 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
490 /* lower hash size to a prime number */
491 cm_data.buf_hashSize = osi_PrimeLessThan((afs_uint32)(cm_data.buf_nbuffers/7 + 1));
493 /* create hash table */
494 memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
496 /* another hash table */
497 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
499 /* create buffer headers and put in free list */
500 bp = cm_data.bufHeaderBaseAddress;
501 data = cm_data.bufDataBaseAddress;
502 cm_data.buf_allp = NULL;
504 for (i=0; i<cm_data.buf_nbuffers; i++) {
505 osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress,
506 "invalid cm_buf_t address");
507 osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
508 "invalid cm_buf_t data address");
510 /* allocate and zero some storage */
511 memset(bp, 0, sizeof(cm_buf_t));
512 bp->magic = CM_BUF_MAGIC;
513 /* thread on list of all buffers */
514 bp->allp = cm_data.buf_allp;
515 cm_data.buf_allp = bp;
517 osi_QAddH( (osi_queue_t **) &cm_data.buf_freeListp,
518 (osi_queue_t **) &cm_data.buf_freeListEndp,
520 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
521 buf_IncrementFreeCount();
522 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
524 /* grab appropriate number of bytes from aligned zone */
529 data += cm_data.buf_blockSize;
532 /* none reserved at first */
533 cm_data.buf_reservedBufs = 0;
535 /* just for safety's sake */
536 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
538 bp = cm_data.bufHeaderBaseAddress;
539 data = cm_data.bufDataBaseAddress;
541 for (i=0; i<cm_data.buf_nbuffers; i++) {
542 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
545 bp->waitRequests = 0;
546 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
548 if (bp->qFlags & CM_BUF_QREDIR) {
550 * extent was not returned by the file system driver.
553 bp->dataVersion = CM_BUF_VERSION_BAD;
554 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QREDIR);
555 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_redirListp,
556 (osi_queue_t **) &cm_data.buf_redirListEndp,
558 buf_DecrementRedirCount();
559 bp->redirq.nextp = bp->redirq.prevp = NULL;
560 bp->redirLastAccess = 0;
561 bp->redirReleaseRequested = 0;
568 * There should be nothing left in cm_data.buf_redirListp
569 * but double check just to be sure.
571 for ( bp = cm_data.buf_redirListp;
573 bp = cm_data.buf_redirListp)
576 * extent was not returned by the file system driver.
579 bp->dataVersion = CM_BUF_VERSION_BAD;
580 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QREDIR);
581 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_redirListp,
582 (osi_queue_t **) &cm_data.buf_redirListEndp,
584 buf_DecrementRedirCount();
585 bp->redirq.nextp = bp->redirq.prevp = NULL;
586 bp->redirLastAccess = 0;
587 bp->redirReleaseRequested = 0;
593 buf_ValidateBufQueues();
597 /* init the buffer trace log */
598 buf_logp = osi_LogCreate("buffer", 1000);
599 osi_LogEnable(buf_logp);
604 /* and create the incr-syncer */
605 phandle = thrd_Create(0, 0,
606 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
609 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
610 CloseHandle(phandle);
614 buf_ValidateBufQueues();
619 /* add nbuffers to the buffer pool, if possible.
620 * Called with no locks held.
622 long buf_AddBuffers(afs_uint64 nbuffers)
624 /* The size of a virtual cache cannot be changed after it has
625 * been created. Subsequent calls to MapViewofFile() with
626 * an existing mapping object name would not allow the
627 * object to be resized. Return failure immediately.
629 * A similar problem now occurs with the persistent cache
630 * given that the memory mapped file now contains a complex
633 afsi_log("request to add %d buffers to the existing cache of size %d denied",
634 nbuffers, cm_data.buf_nbuffers);
636 return CM_ERROR_INVAL;
639 /* interface to set the number of buffers to an exact figure.
640 * Called with no locks held.
642 long buf_SetNBuffers(afs_uint64 nbuffers)
645 return CM_ERROR_INVAL;
646 if (nbuffers == cm_data.buf_nbuffers)
648 else if (nbuffers > cm_data.buf_nbuffers)
649 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
651 return CM_ERROR_INVAL;
654 /* wait for reading or writing to clear; called with write-locked
655 * buffer and unlocked scp and returns with locked buffer.
657 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
662 osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
663 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
666 /* if no IO is happening, we're done */
667 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
670 /* otherwise I/O is happening, but some other thread is waiting for
671 * the I/O already. Wait for that guy to figure out what happened,
672 * and then check again.
674 if ( bp->flags & CM_BUF_WAITING ) {
677 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
679 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
680 _InterlockedOr(&bp->flags, CM_BUF_WAITING);
681 bp->waitCount = bp->waitRequests = 1;
683 osi_SleepM((LONG_PTR)bp, &bp->mx);
685 cm_UpdateServerPriority();
687 lock_ObtainMutex(&bp->mx);
688 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
690 if (bp->waitCount == 0) {
691 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
692 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
693 bp->waitRequests = 0;
697 if (scp = cm_FindSCache(&bp->fid))
701 lock_ObtainRead(&scp->rw);
702 if (scp->flags & CM_SCACHEFLAG_WAITING) {
703 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
704 osi_Wakeup((LONG_PTR)&scp->flags);
706 lock_ReleaseRead(&scp->rw);
710 /* if we get here, the IO is done, but we may have to wakeup people waiting for
711 * the I/O to complete. Do so.
713 if (bp->flags & CM_BUF_WAITING) {
714 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
715 osi_Wakeup((LONG_PTR) bp);
717 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
720 cm_ReleaseSCache(scp);
723 /* find a buffer, if any, for a particular file ID and offset. Assumes
724 * that buf_globalLock is write locked when called.
726 cm_buf_t *buf_FindLocked(struct cm_fid *fidp, osi_hyper_t *offsetp)
731 lock_AssertAny(&buf_globalLock);
733 i = BUF_HASH(fidp, offsetp);
734 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
735 if (cm_FidCmp(fidp, &bp->fid) == 0
736 && offsetp->LowPart == bp->offset.LowPart
737 && offsetp->HighPart == bp->offset.HighPart) {
743 /* return whatever we found, if anything */
747 /* find a buffer with offset *offsetp for vnode *scp. Called
748 * with no locks held.
750 cm_buf_t *buf_Find(struct cm_fid *fidp, osi_hyper_t *offsetp)
754 lock_ObtainRead(&buf_globalLock);
755 bp = buf_FindLocked(fidp, offsetp);
756 lock_ReleaseRead(&buf_globalLock);
761 /* find a buffer, if any, for a particular file ID and offset. Assumes
762 * that buf_globalLock is write locked when called. Uses the all buffer
765 cm_buf_t *buf_FindAllLocked(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
770 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
771 if (cm_FidCmp(fidp, &bp->fid) == 0
772 && offsetp->LowPart == bp->offset.LowPart
773 && offsetp->HighPart == bp->offset.HighPart) {
779 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
780 if (cm_FidCmp(fidp, &bp->fid) == 0) {
783 fileOffset = offsetp->QuadPart + cm_data.baseAddress;
784 if (fileOffset == bp->datap) {
791 /* return whatever we found, if anything */
795 /* find a buffer with offset *offsetp for vnode *scp. Called
796 * with no locks held. Use the all buffer list.
798 cm_buf_t *buf_FindAll(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
802 lock_ObtainRead(&buf_globalLock);
803 bp = buf_FindAllLocked(fidp, offsetp, flags);
804 lock_ReleaseRead(&buf_globalLock);
809 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
812 * Makes sure that there's only one person writing this block
813 * at any given time, and also ensures that the log is forced sufficiently far,
814 * if this buffer contains logged data.
816 * Returns non-zero if the buffer was dirty.
818 * 'scp' may or may not be NULL. If it is not NULL, the FID for both cm_scache_t
819 * and cm_buf_t must match.
821 afs_uint32 buf_CleanAsyncLocked(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp,
822 afs_uint32 flags, afs_uint32 *pisdirty)
825 afs_uint32 isdirty = 0;
829 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
830 osi_assertx(scp == NULL || cm_FidCmp(&scp->fid, &bp->fid) == 0, "scp fid != bp fid");
833 * If the matching cm_scache_t was not provided as a parameter
834 * we must either find one or allocate a new one. It is possible
835 * that the cm_scache_t was recycled out of the cache even though
836 * a cm_buf_t with the same FID is in the cache.
839 if ((scp = cm_FindSCache(&bp->fid)) ||
840 (cm_GetSCache(&bp->fid, &scp,
841 bp->userp ? bp->userp : cm_rootUserp,
847 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
849 lock_ReleaseMutex(&bp->mx);
853 * If we didn't find a cm_scache_t object for bp->fid it means
854 * that we no longer have that FID in the cache. It does not
855 * mean that the object does not exist in the cell. That may
856 * in fact be the case but we don't know that until we attempt
857 * a FetchStatus on the FID.
859 osi_Log1(buf_logp, "buf_CleanAsyncLocked unable to start I/O - scp not found buf 0x%p", bp);
860 code = CM_ERROR_NOSUCHFILE;
862 osi_Log2(buf_logp, "buf_CleanAsyncLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
865 LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
866 code = (*cm_buf_opsp->Writep)(scp, &offset,
868 /* we might as well try to write all of the contiguous
869 * dirty buffers in one RPC
875 flags, bp->userp, reqp);
876 osi_Log3(buf_logp, "buf_CleanAsyncLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
878 lock_ObtainMutex(&bp->mx);
879 /* if the Write routine returns No Such File, clear the dirty flag
880 * because we aren't going to be able to write this data to the file
883 if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD || code == CM_ERROR_NOACCESS ||
884 code == CM_ERROR_QUOTA || code == CM_ERROR_SPACE || code == CM_ERROR_TOOBIG ||
885 code == CM_ERROR_READONLY || code == CM_ERROR_NOSUCHPATH){
886 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
887 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
888 bp->dirty_offset = 0;
889 bp->dirty_length = 0;
891 bp->dataVersion = CM_BUF_VERSION_BAD;
897 /* Disk cache support */
898 /* write buffer to disk cache (synchronous for now) */
899 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
900 #endif /* DISKCACHE95 */
902 /* if we get here and retries are not permitted
903 * then we need to exit this loop regardless of
904 * whether or not we were able to clear the dirty bit
906 if (reqp->flags & CM_REQ_NORETRY)
909 /* Ditto if the hardDeadTimeout or idleTimeout was reached */
910 if (code == CM_ERROR_TIMEDOUT || code == CM_ERROR_ALLDOWN ||
911 code == CM_ERROR_ALLBUSY || code == CM_ERROR_ALLOFFLINE ||
912 code == CM_ERROR_CLOCKSKEW) {
918 cm_ReleaseSCache(scp);
920 /* if someone was waiting for the I/O that just completed or failed,
923 if (bp->flags & CM_BUF_WAITING) {
924 /* turn off flags and wakeup users */
925 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
926 osi_Wakeup((LONG_PTR) bp);
935 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
936 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
937 * The buffer must already be clean, and no I/O should be happening to it.
939 void buf_Recycle(cm_buf_t *bp)
944 cm_buf_t *prevBp, *nextBp;
946 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
948 osi_assertx(!(bp->qFlags & CM_BUF_QREDIR), "can't recycle redir held buffers");
950 /* if we get here, we know that the buffer still has a 0 ref count,
951 * and that it is clean and has no currently pending I/O. This is
952 * the dude to return.
953 * Remember that as long as the ref count is 0, we know that we won't
954 * have any lock conflicts, so we can grab the buffer lock out of
955 * order in the locking hierarchy.
957 osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
958 bp, bp->offset.HighPart, bp->offset.LowPart);
960 osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
961 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
962 "incorrect cm_buf_t flags");
963 lock_AssertWrite(&buf_globalLock);
965 if (bp->qFlags & CM_BUF_QINHASH) {
966 /* Remove from hash */
968 i = BUF_HASH(&bp->fid, &bp->offset);
969 lbpp = &(cm_data.buf_scacheHashTablepp[i]);
970 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = tbp->hashp) {
975 /* we better find it */
976 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
978 *lbpp = bp->hashp; /* hash out */
981 /* Remove from file hash */
983 i = BUF_FILEHASH(&bp->fid);
984 prevBp = bp->fileHashBackp;
985 bp->fileHashBackp = NULL;
986 nextBp = bp->fileHashp;
987 bp->fileHashp = NULL;
989 prevBp->fileHashp = nextBp;
991 cm_data.buf_fileHashTablepp[i] = nextBp;
993 nextBp->fileHashBackp = prevBp;
995 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINHASH);
998 /* make the fid unrecognizable */
999 memset(&bp->fid, 0, sizeof(cm_fid_t));
1001 /* clean up junk flags */
1002 _InterlockedAnd(&bp->flags, ~(CM_BUF_EOF | CM_BUF_ERROR));
1003 bp->dataVersion = CM_BUF_VERSION_BAD; /* unknown so far */
1008 * buf_RDRShakeAnExtentFree
1009 * called with buf_globalLock read locked
1012 buf_RDRShakeAnExtentFree(cm_buf_t *rbp, cm_req_t *reqp)
1014 afs_uint32 code = 0;
1015 LARGE_INTEGER heldExtents = {0,0};
1016 AFSFileExtentCB extentList[1];
1017 DWORD extentCount = 0;
1018 BOOL locked = FALSE;
1020 if (!(rbp->qFlags & CM_BUF_QREDIR))
1023 lock_ReleaseRead(&buf_globalLock);
1025 if (!lock_TryMutex(&buf_rdrReleaseExtentsLock)) {
1026 osi_Log0(afsd_logp, "Waiting for prior RDR_RequestExtentRelease request to complete");
1027 if (reqp->flags & CM_REQ_NORETRY) {
1028 code = CM_ERROR_WOULDBLOCK;
1032 lock_ObtainMutex(&buf_rdrReleaseExtentsLock);
1035 extentList[0].Flags = 0;
1036 extentList[0].Length = cm_data.blockSize;
1037 extentList[0].FileOffset.QuadPart = rbp->offset.QuadPart;
1038 extentList[0].CacheOffset.QuadPart = rbp->datap - cm_data.baseAddress;
1041 code = RDR_RequestExtentRelease(&rbp->fid, heldExtents, extentCount, extentList);
1043 lock_ReleaseMutex(&buf_rdrReleaseExtentsLock);
1046 lock_ObtainRead(&buf_globalLock);
1051 * buf_RDRShakeFileExtentsFree
1052 * requests all extents held by the redirector to be returned for
1053 * the specified cm_scache_t. This function is called with no
1057 buf_RDRShakeFileExtentsFree(cm_scache_t *rscp, cm_req_t *reqp)
1059 afs_uint32 code = 0;
1060 afs_uint64 n_redir = 0;
1062 if (!lock_TryMutex(&buf_rdrReleaseExtentsLock)) {
1063 osi_Log0(afsd_logp, "Waiting for prior RDR_RequestExtentRelease request to complete");
1064 if (reqp->flags & CM_REQ_NORETRY)
1065 return CM_ERROR_WOULDBLOCK;
1067 lock_ObtainMutex(&buf_rdrReleaseExtentsLock);
1070 for ( code = CM_ERROR_RETRY; code == CM_ERROR_RETRY; ) {
1071 LARGE_INTEGER heldExtents = {0,0};
1072 AFSFileExtentCB extentList[1024];
1073 DWORD extentCount = 0;
1077 /* only retry if a call to RDR_RequestExtentRelease says to */
1079 lock_ObtainWrite(&buf_globalLock);
1081 if (rscp->redirBufCount == 0)
1083 lock_ReleaseWrite(&buf_globalLock);
1088 for ( srbp = redirq_to_cm_buf_t(rscp->redirQueueT);
1090 srbp = ((code == 0 && extentCount == 0) ? redirq_to_cm_buf_t(rscp->redirQueueT) :
1091 redirq_to_cm_buf_t(osi_QPrev(&srbp->redirq))))
1093 extentList[extentCount].Flags = 0;
1094 extentList[extentCount].Length = cm_data.blockSize;
1095 extentList[extentCount].FileOffset.QuadPart = srbp->offset.QuadPart;
1096 extentList[extentCount].CacheOffset.QuadPart = srbp->datap - cm_data.baseAddress;
1097 srbp->redirReleaseRequested = now;
1100 if (extentCount == 1024) {
1101 lock_ReleaseWrite(&buf_globalLock);
1102 heldExtents.QuadPart = cm_data.buf_redirCount;
1103 code = RDR_RequestExtentRelease(&rscp->fid, heldExtents, extentCount, extentList);
1105 if (code == CM_ERROR_RETRY) {
1107 * The redirector either is not holding the extents or cannot let them
1108 * go because they are otherwise in use. At the moment, do nothing.
1114 lock_ObtainWrite(&buf_globalLock);
1117 lock_ReleaseWrite(&buf_globalLock);
1119 if (code == 0 && extentCount > 0) {
1120 heldExtents.QuadPart = cm_data.buf_redirCount;
1121 code = RDR_RequestExtentRelease(&rscp->fid, heldExtents, extentCount, extentList);
1124 if ((code == CM_ERROR_RETRY) && (reqp->flags & CM_REQ_NORETRY)) {
1125 code = CM_ERROR_WOULDBLOCK;
1129 lock_ReleaseMutex(&buf_rdrReleaseExtentsLock);
1134 buf_RDRShakeSomeExtentsFree(cm_req_t *reqp, afs_uint32 oneFid, afs_uint32 minage)
1136 afs_uint32 code = 0;
1138 if (!lock_TryMutex(&buf_rdrReleaseExtentsLock)) {
1139 if (reqp->flags & CM_REQ_NORETRY)
1140 return CM_ERROR_WOULDBLOCK;
1142 osi_Log0(afsd_logp, "Waiting for prior RDR_RequestExtentRelease request to complete");
1143 lock_ObtainMutex(&buf_rdrReleaseExtentsLock);
1146 for ( code = CM_ERROR_RETRY; code == CM_ERROR_RETRY; ) {
1147 LARGE_INTEGER heldExtents;
1148 AFSFileExtentCB extentList[1024];
1149 DWORD extentCount = 0;
1150 cm_buf_t *rbp, *srbp;
1153 BOOL locked = FALSE;
1155 /* only retry if a call to RDR_RequestExtentRelease says to */
1157 lock_ObtainWrite(&buf_globalLock);
1160 for ( rbp = cm_data.buf_redirListEndp;
1161 code == 0 && rbp && (!oneFid || extentCount == 0);
1162 rbp = (cm_buf_t *) osi_QPrev(&rbp->q))
1167 if (rbp->redirLastAccess >= rbp->redirReleaseRequested) {
1168 rscp = cm_FindSCache(&rbp->fid);
1173 for ( srbp = redirq_to_cm_buf_t(rscp->redirQueueT);
1174 srbp && extentCount < 1024;
1175 srbp = redirq_to_cm_buf_t(osi_QPrev(&srbp->redirq)))
1178 * Do not request a release if we have already done so
1179 * or if the extent was delivered to windows less than
1180 * 'minage' seconds ago.
1182 if (srbp->redirLastAccess >= srbp->redirReleaseRequested &&
1183 srbp->redirLastAccess < now - minage) {
1184 extentList[extentCount].Flags = 0;
1185 extentList[extentCount].Length = cm_data.blockSize;
1186 extentList[extentCount].FileOffset.QuadPart = srbp->offset.QuadPart;
1187 extentList[extentCount].CacheOffset.QuadPart = srbp->datap - cm_data.baseAddress;
1188 srbp->redirReleaseRequested = now;
1192 cm_ReleaseSCache(rscp);
1195 if ( !oneFid && extentCount > 0) {
1197 lock_ReleaseWrite(&buf_globalLock);
1200 heldExtents.QuadPart = cm_data.buf_redirCount;
1201 code = RDR_RequestExtentRelease(&rbp->fid, heldExtents, extentCount, extentList);
1204 lock_ObtainWrite(&buf_globalLock);
1209 lock_ReleaseWrite(&buf_globalLock);
1212 heldExtents.QuadPart = cm_data.buf_redirCount;
1213 if (rbp && extentCount)
1214 code = RDR_RequestExtentRelease(&rbp->fid, heldExtents, extentCount, extentList);
1216 code = RDR_RequestExtentRelease(NULL, heldExtents, 1024, NULL);
1222 if ((code == CM_ERROR_RETRY) && (reqp->flags & CM_REQ_NORETRY)) {
1223 code = CM_ERROR_WOULDBLOCK;
1227 lock_ReleaseMutex(&buf_rdrReleaseExtentsLock);
1231 /* returns 0 if the buffer does not exist, and non-0 if it does */
1233 buf_ExistsLocked(struct cm_scache *scp, osi_hyper_t *offsetp)
1237 if (bp = buf_FindLocked(&scp->fid, offsetp)) {
1238 /* Do not call buf_ReleaseLocked() because we
1239 * do not want to allow the buffer to be added
1242 afs_int32 refCount = InterlockedDecrement(&bp->refCount);
1243 #ifdef DEBUG_REFCOUNT
1244 osi_Log2(afsd_logp,"buf_ExistsLocked bp 0x%p ref %d", bp, refCount);
1245 afsi_log("%s:%d buf_ExistsLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, refCount);
1247 return CM_BUF_EXISTS;
1253 /* recycle a buffer, removing it from the free list, hashing in its new identity
1254 * and returning it write-locked so that no one can use it. Called without
1255 * any locks held, and can return an error if it loses the race condition and
1256 * finds that someone else created the desired buffer.
1258 * If success is returned, the buffer is returned write-locked.
1260 * May be called with null scp and offsetp, if we're just trying to reclaim some
1261 * space from the buffer pool. In that case, the buffer will be returned
1262 * without being hashed into the hash table.
1264 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
1266 cm_buf_t *bp; /* buffer we're dealing with */
1267 cm_buf_t *nextBp; /* next buffer in file hash chain */
1268 afs_uint32 i; /* temp */
1269 afs_uint64 n_bufs, n_nonzero, n_busy, n_dirty, n_own;
1272 buf_ValidateBufQueues();
1273 #endif /* TESTING */
1283 lock_ObtainRead(&scp->bufCreateLock);
1284 lock_ObtainWrite(&buf_globalLock);
1285 /* check to see if we lost the race */
1286 if (buf_ExistsLocked(scp, offsetp)) {
1287 lock_ReleaseWrite(&buf_globalLock);
1288 lock_ReleaseRead(&scp->bufCreateLock);
1289 return CM_BUF_EXISTS;
1292 /* does this fix the problem below? it's a simple solution. */
1293 if (!cm_data.buf_freeListEndp)
1295 lock_ReleaseWrite(&buf_globalLock);
1296 lock_ReleaseRead(&scp->bufCreateLock);
1298 if ( RDR_Initialized )
1301 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
1306 /* for debugging, assert free list isn't empty, although we
1307 * really should try waiting for a running tranasction to finish
1308 * instead of this; or better, we should have a transaction
1309 * throttler prevent us from entering this situation.
1311 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
1313 /* look at all buffers in free list, some of which may temp.
1314 * have high refcounts and which then should be skipped,
1315 * starting cleaning I/O for those which are dirty. If we find
1316 * a clean buffer, we rehash it, lock it and return it.
1318 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1321 /* check to see if it really has zero ref count. This
1322 * code can bump refcounts, at least, so it may not be
1325 if (bp->refCount > 0) {
1330 /* we don't have to lock buffer itself, since the ref
1331 * count is 0 and we know it will stay zero as long as
1332 * we hold the global lock.
1335 /* don't recycle someone in our own chunk */
1336 if (!cm_FidCmp(&bp->fid, &scp->fid)
1337 && (bp->offset.LowPart & (-cm_chunkSize))
1338 == (offsetp->LowPart & (-cm_chunkSize))) {
1343 /* if this page is being filled (!) or cleaned, see if
1344 * the I/O has completed. If not, skip it, otherwise
1345 * do the final processing for the I/O.
1347 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
1348 /* probably shouldn't do this much work while
1349 * holding the big lock? Watch for contention
1356 if (bp->flags & CM_BUF_DIRTY) {
1359 /* leave the buffer alone if held by the redirector */
1360 if (bp->qFlags & CM_BUF_QREDIR)
1363 /* if the buffer is dirty, start cleaning it and
1364 * move on to the next buffer. We do this with
1365 * just the lock required to minimize contention
1369 lock_ReleaseWrite(&buf_globalLock);
1370 lock_ReleaseRead(&scp->bufCreateLock);
1372 /* grab required lock and clean; this only
1373 * starts the I/O. By the time we're back,
1374 * it'll still be marked dirty, but it will also
1375 * have the WRITING flag set, so we won't get
1378 if (cm_FidCmp(&scp->fid, &bp->fid) == 0)
1379 buf_CleanAsync(scp, bp, reqp, 0, NULL);
1381 buf_CleanAsync(NULL, bp, reqp, 0, NULL);
1383 /* now put it back and go around again */
1386 /* but first obtain the locks we gave up
1387 * before the buf_CleanAsync() call */
1388 lock_ObtainRead(&scp->bufCreateLock);
1389 lock_ObtainWrite(&buf_globalLock);
1392 * Since we dropped the locks we need to verify that
1393 * another thread has not allocated the buffer for us.
1395 if (buf_ExistsLocked(scp, offsetp)) {
1396 lock_ReleaseWrite(&buf_globalLock);
1397 lock_ReleaseRead(&scp->bufCreateLock);
1398 return CM_BUF_EXISTS;
1403 osi_Log3(afsd_logp, "buf_GetNewLocked: scp 0x%p examined %u buffers before recycling bufp 0x%p",
1405 osi_Log4(afsd_logp, "... nonzero %u; own %u; busy %u; dirty %u", n_nonzero, n_own, n_busy, n_dirty);
1407 /* if we get here, we know that the buffer still has a 0
1408 * ref count, and that it is clean and has no currently
1409 * pending I/O. This is the dude to return.
1410 * Remember that as long as the ref count is 0, we know
1411 * that we won't have any lock conflicts, so we can grab
1412 * the buffer lock out of order in the locking hierarchy.
1416 /* now hash in as our new buffer, and give it the
1417 * appropriate label, if requested.
1420 lock_AssertWrite(&buf_globalLock);
1422 _InterlockedOr(&bp->qFlags, CM_BUF_QINHASH);
1427 bp->offset = *offsetp;
1428 i = BUF_HASH(&scp->fid, offsetp);
1429 bp->hashp = cm_data.buf_scacheHashTablepp[i];
1430 cm_data.buf_scacheHashTablepp[i] = bp;
1431 i = BUF_FILEHASH(&scp->fid);
1432 nextBp = cm_data.buf_fileHashTablepp[i];
1433 bp->fileHashp = nextBp;
1434 bp->fileHashBackp = NULL;
1436 nextBp->fileHashBackp = bp;
1437 cm_data.buf_fileHashTablepp[i] = bp;
1440 /* we should remove it from the lru queue. It better still be there,
1441 * since we've held the global (big) lock since we found it there.
1443 osi_assertx(bp->qFlags & CM_BUF_QINLRU,
1444 "buf_GetNewLocked: LRU screwup");
1446 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
1447 (osi_queue_t **) &cm_data.buf_freeListEndp,
1449 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1450 buf_DecrementFreeCount();
1452 /* prepare to return it. Give it a refcount */
1454 #ifdef DEBUG_REFCOUNT
1455 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, 1);
1456 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, 1);
1458 /* grab the mutex so that people don't use it
1459 * before the caller fills it with data. Again, no one
1460 * should have been able to get to this dude to lock it.
1462 if (!lock_TryMutex(&bp->mx)) {
1463 osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked. refCount %d should be 0",
1465 osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
1468 lock_ReleaseWrite(&buf_globalLock);
1469 lock_ReleaseRead(&scp->bufCreateLock);
1474 buf_ValidateBufQueues();
1475 #endif /* TESTING */
1477 } /* for all buffers in lru queue */
1478 lock_ReleaseWrite(&buf_globalLock);
1479 lock_ReleaseRead(&scp->bufCreateLock);
1481 osi_Log1(afsd_logp, "buf_GetNewLocked: Free Buffer List has %u buffers none free", n_bufs);
1482 osi_Log4(afsd_logp, "... nonzero %u; own %u; busy %u; dirty %u", n_nonzero, n_own, n_busy, n_dirty);
1484 if (RDR_Initialized) {
1487 code = buf_RDRShakeSomeExtentsFree(reqp, TRUE, 2 /* seconds */);
1489 case CM_ERROR_RETRY:
1492 case CM_ERROR_WOULDBLOCK:
1493 return CM_ERROR_WOULDBLOCK;
1497 Sleep(100); /* give some time for a buffer to be freed */
1498 } /* while loop over everything */
1503 * get a page, returning it held but unlocked. the page may or may not
1504 * contain valid data.
1506 * The scp must be unlocked when passed in unlocked.
1508 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
1512 osi_hyper_t pageOffset;
1513 unsigned long tcount;
1517 cm_diskcache_t *dcp;
1518 #endif /* DISKCACHE95 */
1521 pageOffset.HighPart = offsetp->HighPart;
1522 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
1526 buf_ValidateBufQueues();
1527 #endif /* TESTING */
1529 bp = buf_Find(&scp->fid, &pageOffset);
1531 /* lock it and break out */
1532 lock_ObtainMutex(&bp->mx);
1535 /* touch disk chunk to update LRU info */
1536 diskcache_Touch(bp->dcp);
1537 #endif /* DISKCACHE95 */
1541 /* otherwise, we have to create a page */
1542 code = buf_GetNewLocked(scp, &pageOffset, reqp, &bp);
1545 /* the requested buffer was created */
1550 * the requested buffer existed by the time the
1551 * scp->bufCreateLock and buf_globalLock could be obtained.
1552 * loop again and permit buf_Find() to obtain a reference.
1557 * the requested buffer could not be created.
1558 * return the error to the caller.
1561 buf_ValidateBufQueues();
1562 #endif /* TESTING */
1565 } /* big while loop */
1567 /* if we get here, we have a locked buffer that may have just been
1568 * created, in which case it needs to be filled with data.
1571 /* load the page; freshly created pages should be idle */
1572 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
1575 * start the I/O; may drop lock. as of this writing, the only
1576 * implementation of Readp is cm_BufRead() which simply sets
1577 * tcount to 0 and returns success.
1579 _InterlockedOr(&bp->flags, CM_BUF_READING);
1580 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
1583 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
1584 bp->dcp = dcp; /* pointer to disk cache struct. */
1585 #endif /* DISKCACHE95 */
1588 /* failure or queued */
1590 /* unless cm_BufRead() is altered, this path cannot be hit */
1591 if (code != ERROR_IO_PENDING) {
1593 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1594 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1595 if (bp->flags & CM_BUF_WAITING) {
1596 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1597 osi_Wakeup((LONG_PTR) bp);
1599 lock_ReleaseMutex(&bp->mx);
1602 buf_ValidateBufQueues();
1603 #endif /* TESTING */
1608 * otherwise, I/O completed instantly and we're done, except
1609 * for padding the xfr out with 0s and checking for EOF
1611 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1612 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1614 _InterlockedOr(&bp->flags, CM_BUF_EOF);
1616 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1617 if (bp->flags & CM_BUF_WAITING) {
1618 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1619 osi_Wakeup((LONG_PTR) bp);
1624 /* wait for reads, either that which we started above, or that someone
1625 * else started. We don't care if we return a buffer being cleaned.
1627 if (bp->flags & CM_BUF_READING)
1628 buf_WaitIO(scp, bp);
1630 /* once it has been read once, we can unlock it and return it, still
1631 * with its refcount held.
1633 lock_ReleaseMutex(&bp->mx);
1636 /* now remove from queue; will be put in at the head (farthest from
1637 * being recycled) when we're done in buf_Release.
1639 lock_ObtainWrite(&buf_globalLock);
1640 if (bp->qFlags & CM_BUF_QINLRU) {
1641 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
1642 (osi_queue_t **) &cm_data.buf_freeListEndp,
1644 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1645 buf_DecrementFreeCount();
1647 lock_ReleaseWrite(&buf_globalLock);
1649 osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1650 bp, scp, offsetp->HighPart, offsetp->LowPart);
1652 buf_ValidateBufQueues();
1653 #endif /* TESTING */
1657 /* clean a buffer synchronously */
1658 afs_uint32 buf_CleanAsync(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp, afs_uint32 flags, afs_uint32 *pisdirty)
1661 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1662 osi_assertx(!(flags & CM_BUF_WRITE_SCP_LOCKED), "scp->rw must not be held when calling buf_CleanAsync");
1664 lock_ObtainMutex(&bp->mx);
1665 code = buf_CleanAsyncLocked(scp, bp, reqp, flags, pisdirty);
1666 lock_ReleaseMutex(&bp->mx);
1671 /* wait for a buffer's cleaning to finish */
1672 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp, afs_uint32 locked)
1674 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1677 lock_ObtainMutex(&bp->mx);
1678 if (bp->flags & CM_BUF_WRITING) {
1679 buf_WaitIO(scp, bp);
1682 lock_ReleaseMutex(&bp->mx);
1685 /* set the dirty flag on a buffer, and set associated write-ahead log,
1686 * if there is one. Allow one to be added to a buffer, but not changed.
1688 * The buffer must be locked before calling this routine.
1690 void buf_SetDirty(cm_buf_t *bp, cm_req_t *reqp, afs_uint32 offset, afs_uint32 length, cm_user_t *userp)
1692 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1693 osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
1698 if (bp->flags & CM_BUF_DIRTY) {
1700 osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
1702 if (bp->dirty_offset <= offset) {
1703 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1704 /* dirty_length remains the same */
1706 bp->dirty_length = offset + length - bp->dirty_offset;
1708 } else /* bp->dirty_offset > offset */ {
1709 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1710 bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
1712 bp->dirty_length = length;
1714 bp->dirty_offset = offset;
1717 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1720 _InterlockedOr(&bp->flags, CM_BUF_DIRTY);
1722 /* and turn off EOF flag, since it has associated data now */
1723 _InterlockedAnd(&bp->flags, ~CM_BUF_EOF);
1725 bp->dirty_offset = offset;
1726 bp->dirty_length = length;
1729 * if the request is not from the afs redirector,
1730 * add to the dirty list. The redirector interface ensures
1731 * that a background store operation is queued for each and
1732 * every dirty extent that is released. Therefore, the
1733 * buf_IncrSyncer thread is not required to ensure that
1734 * dirty buffers are written to the file server.
1736 * we obtain a hold on the buffer for as long as it remains
1737 * in the list. buffers are only removed from the list by
1738 * the buf_IncrSyncer function regardless of when else the
1739 * dirty flag might be cleared.
1741 * This should never happen but just in case there is a bug
1742 * elsewhere, never add to the dirty list if the buffer is
1745 if (!(reqp->flags & CM_REQ_SOURCE_REDIR)) {
1746 lock_ObtainWrite(&buf_globalLock);
1747 if (!(bp->qFlags & CM_BUF_QINDL)) {
1749 if (!cm_data.buf_dirtyListp) {
1750 cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
1752 cm_data.buf_dirtyListEndp->dirtyp = bp;
1753 cm_data.buf_dirtyListEndp = bp;
1756 _InterlockedOr(&bp->qFlags, CM_BUF_QINDL);
1758 lock_ReleaseWrite(&buf_globalLock);
1762 /* and record the last writer */
1763 if (bp->userp != userp) {
1766 cm_ReleaseUser(bp->userp);
1771 /* clean all buffers, reset log pointers and invalidate all buffers.
1772 * Called with no locks held, and returns with same.
1774 * This function is guaranteed to clean and remove the log ptr of all the
1775 * buffers that were dirty or had non-zero log ptrs before the call was
1776 * made. That's sufficient to clean up any garbage left around by recovery,
1777 * which is all we're counting on this for; there may be newly created buffers
1778 * added while we're running, but that should be OK.
1780 * In an environment where there are no transactions (artificially imposed, for
1781 * example, when switching the database to raw mode), this function is used to
1782 * make sure that all updates have been written to the disk. In that case, we don't
1783 * really require that we forget the log association between pages and logs, but
1784 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1785 * have to worry about invalidating data in the buffers.
1787 * This function is used at the end of recovery as paranoia to get the recovered
1788 * database out to disk. It removes all references to the recovery log and cleans
1791 long buf_CleanAndReset(void)
1797 lock_ObtainRead(&buf_globalLock);
1798 for(i=0; i<cm_data.buf_hashSize; i++) {
1799 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
1800 if (bp->qFlags & CM_BUF_QREDIR) {
1801 osi_Log1(buf_logp,"buf_CleanAndReset buffer held by redirector bp 0x%p", bp);
1803 /* Request single extent from the redirector */
1804 buf_RDRShakeAnExtentFree(bp, &req);
1807 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1809 lock_ReleaseRead(&buf_globalLock);
1811 /* now no locks are held; clean buffer and go on */
1813 req.flags |= CM_REQ_NORETRY;
1815 buf_CleanAsync(NULL, bp, &req, 0, NULL);
1816 buf_CleanWait(NULL, bp, FALSE);
1818 /* relock and release buffer */
1819 lock_ObtainRead(&buf_globalLock);
1820 buf_ReleaseLocked(bp, FALSE);
1822 } /* over one bucket */
1823 } /* for loop over all hash buckets */
1826 lock_ReleaseRead(&buf_globalLock);
1829 buf_ValidateBufQueues();
1830 #endif /* TESTING */
1832 /* and we're done */
1836 /* called without global lock being held, reserves buffers for callers
1837 * that need more than one held (not locked) at once.
1839 void buf_ReserveBuffers(afs_uint64 nbuffers)
1841 lock_ObtainWrite(&buf_globalLock);
1843 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1844 cm_data.buf_reserveWaiting = 1;
1845 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1846 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1847 lock_ObtainWrite(&buf_globalLock);
1850 cm_data.buf_reservedBufs += nbuffers;
1854 lock_ReleaseWrite(&buf_globalLock);
1857 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1861 lock_ObtainWrite(&buf_globalLock);
1862 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1866 cm_data.buf_reservedBufs += nbuffers;
1869 lock_ReleaseWrite(&buf_globalLock);
1873 /* called without global lock held, releases reservation held by
1874 * buf_ReserveBuffers.
1876 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1878 lock_ObtainWrite(&buf_globalLock);
1879 cm_data.buf_reservedBufs -= nbuffers;
1880 if (cm_data.buf_reserveWaiting) {
1881 cm_data.buf_reserveWaiting = 0;
1882 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1884 lock_ReleaseWrite(&buf_globalLock);
1887 /* truncate the buffers past sizep, zeroing out the page, if we don't
1888 * end on a page boundary.
1890 * Requires cm_bufCreateLock to be write locked.
1892 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1896 cm_buf_t *nbufp; /* next buffer, if didRelease */
1902 /* assert that cm_bufCreateLock is held in write mode */
1903 lock_AssertWrite(&scp->bufCreateLock);
1905 i = BUF_FILEHASH(&scp->fid);
1907 lock_ObtainRead(&buf_globalLock);
1908 bufp = cm_data.buf_fileHashTablepp[i];
1910 lock_ReleaseRead(&buf_globalLock);
1914 buf_HoldLocked(bufp);
1915 lock_ReleaseRead(&buf_globalLock);
1918 lock_ObtainMutex(&bufp->mx);
1920 bufEnd.HighPart = 0;
1921 bufEnd.LowPart = cm_data.buf_blockSize;
1922 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1924 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1925 LargeIntegerLessThan(*sizep, bufEnd)) {
1926 buf_WaitIO(scp, bufp);
1928 lock_ObtainWrite(&scp->rw);
1930 /* make sure we have a callback (so we have the right value for
1931 * the length), and wait for it to be safe to do a truncate.
1933 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1934 CM_SCACHESYNC_NEEDCALLBACK
1935 | CM_SCACHESYNC_GETSTATUS
1936 | CM_SCACHESYNC_SETSIZE
1937 | CM_SCACHESYNC_BUFLOCKED);
1940 /* if we succeeded in our locking, and this applies to the right
1941 * file, and the truncate request overlaps the buffer either
1942 * totally or partially, then do something.
1944 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1945 && LargeIntegerLessThan(*sizep, bufEnd)) {
1948 /* destroy the buffer, turning off its dirty bit, if
1949 * we're truncating the whole buffer. Otherwise, set
1950 * the dirty bit, and clear out the tail of the buffer
1951 * if we just overlap some.
1953 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1954 /* truncating the entire page */
1955 if (reqp->flags & CM_REQ_SOURCE_REDIR) {
1957 * Implicitly clear the redirector flag
1958 * and release the matching hold.
1960 if (bufp->qFlags & CM_BUF_QREDIR) {
1961 osi_Log4(buf_logp,"buf_Truncate taking from file system bufp 0x%p vno 0x%x foffset 0x%x:%x",
1962 bufp, bufp->fid.vnode, bufp->offset.HighPart, bufp->offset.LowPart);
1963 lock_ObtainWrite(&buf_globalLock);
1964 if (bufp->qFlags & CM_BUF_QREDIR) {
1965 buf_RemoveFromRedirQueue(scp, bufp);
1966 buf_ReleaseLocked(bufp, TRUE);
1968 lock_ReleaseWrite(&buf_globalLock);
1971 if (RDR_Initialized)
1972 RDR_InvalidateObject(scp->fid.cell, scp->fid.volume, scp->fid.vnode,
1973 scp->fid.unique, scp->fid.hash,
1974 scp->fileType, AFS_INVALIDATE_SMB);
1976 _InterlockedAnd(&bufp->flags, ~CM_BUF_DIRTY);
1978 bufp->dirty_offset = 0;
1979 bufp->dirty_length = 0;
1980 bufp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1981 bufp->dirtyCounter++;
1984 /* don't set dirty, since dirty implies
1985 * currently up-to-date. Don't need to do this,
1986 * since we'll update the length anyway.
1988 * Zero out remainder of the page, in case we
1989 * seek and write past EOF, and make this data
1992 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1993 osi_assertx(bufferPos != 0, "non-zero bufferPos");
1994 memset(bufp->datap + bufferPos, 0,
1995 cm_data.buf_blockSize - bufferPos);
1999 cm_SyncOpDone( scp, bufp,
2000 CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
2001 | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
2003 lock_ReleaseWrite(&scp->rw);
2004 lock_ReleaseMutex(&bufp->mx);
2007 nbufp = bufp->fileHashp;
2011 /* This forces the loop to end and the error code
2012 * to be returned. */
2020 buf_ValidateBufQueues();
2021 #endif /* TESTING */
2027 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
2030 cm_buf_t *bp; /* buffer we're hacking on */
2034 afs_uint32 stable = 0;
2036 i = BUF_FILEHASH(&scp->fid);
2039 lock_ObtainRead(&buf_globalLock);
2040 bp = cm_data.buf_fileHashTablepp[i];
2043 lock_ReleaseRead(&buf_globalLock);
2045 for (; bp; bp = nbp) {
2046 didRelease = 0; /* haven't released this buffer yet */
2048 /* clean buffer synchronously */
2049 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
2051 if (code == 0 && !stable && (bp->flags & CM_BUF_DIRTY)) {
2053 * we must stabilize the object to ensure that buffer
2054 * changes cannot occur while the flush is performed.
2055 * However, we do not want to Stabilize if we do not
2056 * need to because Stabilize obtains a callback.
2058 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
2059 stable = (code == 0);
2062 if (code == CM_ERROR_BADFD) {
2063 /* if the scp's FID is bad its because we received VNOVNODE
2064 * when attempting to FetchStatus before the write. This
2065 * page therefore contains data that can no longer be stored.
2067 lock_ObtainMutex(&bp->mx);
2068 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2069 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2070 bp->error = CM_ERROR_BADFD;
2071 bp->dirty_offset = 0;
2072 bp->dirty_length = 0;
2073 bp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
2075 lock_ReleaseMutex(&bp->mx);
2076 } else if (!(scp->flags & CM_SCACHEFLAG_RO)) {
2081 lock_ObtainMutex(&bp->mx);
2083 /* start cleaning the buffer, and wait for it to finish */
2084 buf_CleanAsyncLocked(scp, bp, reqp, 0, NULL);
2085 buf_WaitIO(scp, bp);
2087 lock_ReleaseMutex(&bp->mx);
2090 /* actually, we only know that buffer is clean if ref
2091 * count is 1, since we don't have buffer itself locked.
2093 if (!(bp->flags & CM_BUF_DIRTY) && !(bp->qFlags & CM_BUF_QREDIR)) {
2094 lock_ObtainWrite(&buf_globalLock);
2095 if (!(bp->flags & CM_BUF_DIRTY) && !(bp->qFlags & CM_BUF_QREDIR)) {
2096 if (bp->refCount == 1) { /* bp is held above */
2097 nbp = bp->fileHashp;
2099 buf_HoldLocked(nbp);
2100 buf_ReleaseLocked(bp, TRUE);
2105 lock_ReleaseWrite(&buf_globalLock);
2111 lock_ObtainRead(&buf_globalLock);
2112 nbp = bp->fileHashp;
2114 buf_HoldLocked(nbp);
2115 buf_ReleaseLocked(bp, FALSE);
2116 lock_ReleaseRead(&buf_globalLock);
2118 } /* for loop over a bunch of buffers */
2121 (*cm_buf_opsp->Unstabilizep)(scp, userp);
2124 buf_ValidateBufQueues();
2125 #endif /* TESTING */
2131 /* Must be called with scp->rw held */
2132 long buf_ForceDataVersion(cm_scache_t * scp, afs_uint64 fromVersion, afs_uint64 toVersion)
2138 lock_AssertAny(&scp->rw);
2140 i = BUF_FILEHASH(&scp->fid);
2142 lock_ObtainRead(&buf_globalLock);
2144 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
2145 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
2146 if (bp->dataVersion == fromVersion) {
2147 bp->dataVersion = toVersion;
2152 lock_ReleaseRead(&buf_globalLock);
2160 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
2164 cm_buf_t *bp; /* buffer we're hacking on */
2165 cm_buf_t *nbp; /* next one */
2168 if (RDR_Initialized && scp->redirBufCount > 0) {
2169 /* Retrieve all extents for this file from the redirector */
2170 buf_RDRShakeFileExtentsFree(scp, reqp);
2173 i = BUF_FILEHASH(&scp->fid);
2175 lock_ObtainRead(&buf_globalLock);
2176 bp = cm_data.buf_fileHashTablepp[i];
2179 lock_ReleaseRead(&buf_globalLock);
2180 for (; bp; bp = nbp) {
2181 /* clean buffer synchronously */
2182 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
2184 * If the buffer is held by the redirector we must fetch
2185 * it back in order to determine whether or not it is in
2188 lock_ObtainRead(&buf_globalLock);
2189 if (bp->qFlags & CM_BUF_QREDIR) {
2190 osi_Log1(buf_logp,"buf_CleanVnode buffer held by redirector bp 0x%p", bp);
2192 /* Retrieve single extent from the redirector */
2193 buf_RDRShakeAnExtentFree(bp, reqp);
2195 lock_ReleaseRead(&buf_globalLock);
2197 lock_ObtainMutex(&bp->mx);
2198 if ((bp->flags & CM_BUF_DIRTY)) {
2199 if (userp && userp != bp->userp) {
2202 cm_ReleaseUser(bp->userp);
2207 case CM_ERROR_NOSUCHFILE:
2208 case CM_ERROR_BADFD:
2209 case CM_ERROR_NOACCESS:
2210 case CM_ERROR_QUOTA:
2211 case CM_ERROR_SPACE:
2212 case CM_ERROR_TOOBIG:
2213 case CM_ERROR_READONLY:
2214 case CM_ERROR_NOSUCHPATH:
2216 * Apply the previous fatal error to this buffer.
2217 * Do not waste the time attempting to store to
2218 * the file server when we know it will fail.
2220 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2221 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2222 bp->dirty_offset = 0;
2223 bp->dirty_length = 0;
2225 bp->dataVersion = CM_BUF_VERSION_BAD;
2228 case CM_ERROR_TIMEDOUT:
2229 case CM_ERROR_ALLDOWN:
2230 case CM_ERROR_ALLBUSY:
2231 case CM_ERROR_ALLOFFLINE:
2232 case CM_ERROR_CLOCKSKEW:
2233 /* do not mark the buffer in error state but do
2234 * not attempt to complete the rest either.
2238 code = buf_CleanAsyncLocked(scp, bp, reqp, 0, &wasDirty);
2239 if (bp->flags & CM_BUF_ERROR) {
2245 buf_CleanWait(scp, bp, TRUE);
2247 lock_ReleaseMutex(&bp->mx);
2250 lock_ObtainRead(&buf_globalLock);
2251 nbp = bp->fileHashp;
2253 buf_HoldLocked(nbp);
2254 buf_ReleaseLocked(bp, FALSE);
2255 lock_ReleaseRead(&buf_globalLock);
2256 } /* for loop over a bunch of buffers */
2259 buf_ValidateBufQueues();
2260 #endif /* TESTING */
2268 buf_ValidateBufQueues(void)
2270 cm_buf_t * bp, *bpb, *bpf, *bpa;
2271 afs_uint32 countf=0, countb=0, counta=0;
2273 lock_ObtainRead(&buf_globalLock);
2274 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
2275 if (bp->magic != CM_BUF_MAGIC)
2276 osi_panic("buf magic error",__FILE__,__LINE__);
2281 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
2282 if (bp->magic != CM_BUF_MAGIC)
2283 osi_panic("buf magic error",__FILE__,__LINE__);
2288 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
2289 if (bp->magic != CM_BUF_MAGIC)
2290 osi_panic("buf magic error",__FILE__,__LINE__);
2294 lock_ReleaseRead(&buf_globalLock);
2296 if (countb != countf)
2297 osi_panic("buf magic error",__FILE__,__LINE__);
2299 if (counta != cm_data.buf_nbuffers)
2300 osi_panic("buf magic error",__FILE__,__LINE__);
2302 #endif /* TESTING */
2304 /* dump the contents of the buf_scacheHashTablepp. */
2305 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
2312 if (cm_data.buf_scacheHashTablepp == NULL)
2316 lock_ObtainRead(&buf_globalLock);
2318 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n",
2319 cookie, cm_data.buf_hashSize);
2320 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2322 for (i = 0; i < cm_data.buf_hashSize; i++)
2324 for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp)
2326 StringCbPrintfA(output, sizeof(output),
2327 "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
2328 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
2329 "flags=0x%x, qFlags=0x%x cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
2330 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
2331 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
2332 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
2333 bp->cmFlags, bp->error, bp->refCount);
2334 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2338 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
2339 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2341 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
2342 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2343 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
2344 StringCbPrintfA(output, sizeof(output),
2345 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
2346 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
2347 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
2348 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
2349 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
2350 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
2351 bp->cmFlags, bp->error, bp->refCount);
2352 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2354 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
2355 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2357 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListp\r\n", cookie);
2358 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2359 for(bp = cm_data.buf_dirtyListp; bp; bp=bp->dirtyp) {
2360 StringCbPrintfA(output, sizeof(output),
2361 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
2362 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
2363 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
2364 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
2365 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
2366 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
2367 bp->cmFlags, bp->error, bp->refCount);
2368 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2370 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListp.\r\n", cookie);
2371 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2374 lock_ReleaseRead(&buf_globalLock);
2378 void buf_ForceTrace(BOOL flush)
2387 len = GetTempPath(sizeof(buf)-10, buf);
2388 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
2389 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
2390 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
2391 if (handle == INVALID_HANDLE_VALUE) {
2392 osi_panic("Cannot create log file", __FILE__, __LINE__);
2394 osi_LogPrint(buf_logp, handle);
2396 FlushFileBuffers(handle);
2397 CloseHandle(handle);
2400 long buf_DirtyBuffersExist(cm_fid_t *fidp)
2403 afs_uint32 bcount = 0;
2407 i = BUF_FILEHASH(fidp);
2409 lock_ObtainRead(&buf_globalLock);
2410 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
2411 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
2416 lock_ReleaseRead(&buf_globalLock);
2420 long buf_RDRBuffersExist(cm_fid_t *fidp)
2423 afs_uint32 bcount = 0;
2427 if (!RDR_Initialized)
2430 i = BUF_FILEHASH(fidp);
2432 lock_ObtainRead(&buf_globalLock);
2433 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
2434 if (!cm_FidCmp(fidp, &bp->fid) && (bp->qFlags & CM_BUF_QREDIR)) {
2439 lock_ReleaseRead(&buf_globalLock);
2443 long buf_ClearRDRFlag(cm_scache_t *scp, char *reason)
2445 cm_fid_t *fidp = &scp->fid;
2447 afs_uint32 bcount = 0;
2450 i = BUF_FILEHASH(fidp);
2452 lock_ObtainWrite(&scp->rw);
2453 lock_ObtainRead(&buf_globalLock);
2454 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
2455 if (!cm_FidCmp(fidp, &bp->fid) && (bp->qFlags & CM_BUF_QREDIR)) {
2456 lock_ConvertRToW(&buf_globalLock);
2457 if (bp->qFlags & CM_BUF_QREDIR) {
2458 osi_Log4(buf_logp,"buf_ClearRDRFlag taking from file system bp 0x%p vno 0x%x foffset 0x%x:%x",
2459 bp, bp->fid.vnode, bp->offset.HighPart, bp->offset.LowPart);
2460 buf_RemoveFromRedirQueue(scp, bp);
2461 buf_ReleaseLocked(bp, TRUE);
2463 lock_ConvertWToR(&buf_globalLock);
2467 /* Confirm that there are none left */
2468 lock_ConvertRToW(&buf_globalLock);
2469 for ( bp = redirq_to_cm_buf_t(scp->redirQueueT);
2471 bp = redirq_to_cm_buf_t(scp->redirQueueT))
2473 if (bp->qFlags & CM_BUF_QREDIR) {
2474 osi_Log4(buf_logp,"buf_ClearRDRFlag taking from file system bufp 0x%p vno 0x%x foffset 0x%x:%x",
2475 bp, bp->fid.vnode, bp->offset.HighPart, bp->offset.LowPart);
2476 buf_RemoveFromRedirQueue(scp, bp);
2477 buf_ReleaseLocked(bp, TRUE);
2481 lock_ReleaseWrite(&buf_globalLock);
2482 lock_ReleaseWrite(&scp->rw);
2487 long buf_CleanDirtyBuffers(cm_scache_t *scp)
2490 afs_uint32 bcount = 0;
2491 cm_fid_t * fidp = &scp->fid;
2493 for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
2494 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
2496 lock_ObtainMutex(&bp->mx);
2497 _InterlockedAnd(&bp->cmFlags, ~CM_BUF_CMSTORING);
2498 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2499 bp->dirty_offset = 0;
2500 bp->dirty_length = 0;
2501 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2502 bp->error = VNOVNODE;
2503 bp->dataVersion = CM_BUF_VERSION_BAD; /* bad */
2505 if (bp->flags & CM_BUF_WAITING) {
2506 osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
2507 osi_Wakeup((long) &bp);
2509 lock_ReleaseMutex(&bp->mx);
2518 * The following routines will not be used on a
2519 * regular basis but are very useful in a variety
2520 * of scenarios when debugging data corruption.
2523 buf_HexCheckSum(cm_buf_t * bp)
2526 static char buf[33];
2527 static char tr[16] = {'0','1','2','3','4','5','6','7','8','9','A','B','C','D','E','F'};
2529 for (i=0;i<16;i++) {
2530 k = bp->md5cksum[i];
2532 buf[i*2] = tr[k / 16];
2533 buf[i*2+1] = tr[k % 16];
2541 buf_ComputeCheckSum(cm_buf_t * bp)
2546 MD5_Update(&md5, bp->datap, cm_data.blockSize);
2547 MD5_Final(bp->md5cksum, &md5);
2549 osi_Log4(buf_logp, "CheckSum bp 0x%p md5 %s, dirty: offset %u length %u",
2550 bp, osi_LogSaveString(buf_logp, buf_HexCheckSum(bp)),
2551 bp->dirty_offset, bp->dirty_length);
2555 buf_ValidateCheckSum(cm_buf_t * bp)
2558 unsigned char tmp[16];
2561 MD5_Update(&md5, bp->datap, cm_data.blockSize);
2562 MD5_Final(tmp, &md5);
2564 if (memcmp(tmp, bp->md5cksum, 16) == 0)
2570 buf_InsertToRedirQueue(cm_scache_t *scp, cm_buf_t *bufp)
2572 lock_AssertWrite(&buf_globalLock);
2574 lock_AssertWrite(&scp->rw);
2576 if (bufp->qFlags & CM_BUF_QINLRU) {
2577 _InterlockedAnd(&bufp->qFlags, ~CM_BUF_QINLRU);
2578 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
2579 (osi_queue_t **) &cm_data.buf_freeListEndp,
2581 buf_DecrementFreeCount();
2583 _InterlockedOr(&bufp->qFlags, CM_BUF_QREDIR);
2584 osi_QAddH( (osi_queue_t **) &cm_data.buf_redirListp,
2585 (osi_queue_t **) &cm_data.buf_redirListEndp,
2587 buf_IncrementRedirCount();
2588 bufp->redirLastAccess = time(NULL);
2590 osi_QAddH( (osi_queue_t **) &scp->redirQueueH,
2591 (osi_queue_t **) &scp->redirQueueT,
2593 scp->redirLastAccess = bufp->redirLastAccess;
2594 InterlockedIncrement(&scp->redirBufCount);
2599 buf_RemoveFromRedirQueue(cm_scache_t *scp, cm_buf_t *bufp)
2601 lock_AssertWrite(&buf_globalLock);
2603 lock_AssertWrite(&scp->rw);
2605 _InterlockedAnd(&bufp->qFlags, ~CM_BUF_QREDIR);
2606 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_redirListp,
2607 (osi_queue_t **) &cm_data.buf_redirListEndp,
2609 buf_DecrementRedirCount();
2611 osi_QRemoveHT( (osi_queue_t **) &scp->redirQueueH,
2612 (osi_queue_t **) &scp->redirQueueT,
2614 InterlockedDecrement(&scp->redirBufCount);
2619 buf_MoveToHeadOfRedirQueue(cm_scache_t *scp, cm_buf_t *bufp)
2621 lock_AssertWrite(&buf_globalLock);
2622 osi_assertx(bufp->qFlags & CM_BUF_QREDIR,
2623 "buf_MoveToHeadOfRedirQueue buffer not held by redirector");
2626 lock_AssertWrite(&scp->rw);
2628 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_redirListp,
2629 (osi_queue_t **) &cm_data.buf_redirListEndp,
2631 osi_QAddH( (osi_queue_t **) &cm_data.buf_redirListp,
2632 (osi_queue_t **) &cm_data.buf_redirListEndp,
2634 bufp->redirLastAccess = time(NULL);
2636 osi_QRemoveHT( (osi_queue_t **) &scp->redirQueueH,
2637 (osi_queue_t **) &scp->redirQueueT,
2639 osi_QAddH( (osi_queue_t **) &scp->redirQueueH,
2640 (osi_queue_t **) &scp->redirQueueT,
2642 scp->redirLastAccess = bufp->redirLastAccess;