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)
191 lock_ObtainRead(&buf_globalLock);
192 buf_ReleaseLocked(bp, FALSE);
193 lock_ReleaseRead(&buf_globalLock);
197 buf_Sync(int quitOnShutdown)
199 cm_buf_t **bpp, *bp, *prevbp;
200 afs_uint32 wasDirty = 0;
203 /* go through all of the dirty buffers */
204 lock_ObtainRead(&buf_globalLock);
205 for (bpp = &cm_data.buf_dirtyListp, prevbp = NULL; bp = *bpp; ) {
206 if (quitOnShutdown && buf_ShutdownFlag)
210 * If the buffer is held be the redirector we must fetch
211 * it back in order to determine whether or not it is in
214 if (bp->qFlags & CM_BUF_QREDIR) {
215 osi_Log1(buf_logp,"buf_Sync buffer held by redirector bp 0x%p", bp);
217 /* Request single buffer from the redirector */
218 buf_RDRShakeAnExtentFree(bp, &req);
221 lock_ReleaseRead(&buf_globalLock);
223 * all dirty buffers are held when they are added to the
224 * dirty list. No need for an additional hold.
226 lock_ObtainMutex(&bp->mx);
228 if ((bp->flags & CM_BUF_DIRTY)) {
229 /* start cleaning the buffer; don't touch log pages since
230 * the log code counts on knowing exactly who is writing
231 * a log page at any given instant.
233 * only attempt to write the buffer if the volume might
239 volp = cm_GetVolumeByFID(&bp->fid);
240 switch (cm_GetVolumeStatus(volp, bp->fid.volume)) {
244 req.flags |= CM_REQ_NORETRY;
245 buf_CleanLocked(NULL, bp, &req, 0, &dirty);
251 /* the buffer may or may not have been dirty
252 * and if dirty may or may not have been cleaned
253 * successfully. check the dirty flag again.
255 if (!(bp->flags & CM_BUF_DIRTY)) {
256 /* remove the buffer from the dirty list */
257 lock_ObtainWrite(&buf_globalLock);
258 #ifdef DEBUG_REFCOUNT
259 if (bp->dirtyp == NULL && bp != cm_data.buf_dirtyListEndp) {
260 osi_Log1(afsd_logp,"buf_Sync bp 0x%p list corruption",bp);
261 afsi_log("buf_Sync bp 0x%p list corruption", bp);
266 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINDL);
267 if (cm_data.buf_dirtyListp == NULL)
268 cm_data.buf_dirtyListEndp = NULL;
269 else if (cm_data.buf_dirtyListEndp == bp)
270 cm_data.buf_dirtyListEndp = prevbp;
271 buf_ReleaseLocked(bp, TRUE);
272 lock_ConvertWToR(&buf_globalLock);
274 if (buf_ShutdownFlag) {
277 char volstr[VL_MAXNAMELEN+12]="";
280 volp = cm_GetVolumeByFID(&bp->fid);
283 if (bp->fid.volume == volp->vol[RWVOL].ID)
285 else if (bp->fid.volume == volp->vol[ROVOL].ID)
287 else if (bp->fid.volume == volp->vol[BACKVOL].ID)
291 snprintf(volstr, sizeof(volstr), "%s%s", volp->namep, ext);
293 cellp = cm_FindCellByID(bp->fid.cell, CM_FLAG_NOPROBE);
294 snprintf(volstr, sizeof(volstr), "%u", bp->fid.volume);
297 LogEvent(EVENTLOG_INFORMATION_TYPE, MSG_DIRTY_BUFFER_AT_SHUTDOWN,
298 cellp->name, volstr, bp->fid.vnode, bp->fid.unique,
299 bp->offset.QuadPart+bp->dirty_offset, bp->dirty_length);
302 /* advance the pointer so we don't loop forever */
303 lock_ObtainRead(&buf_globalLock);
307 lock_ReleaseMutex(&bp->mx);
308 } /* for loop over a bunch of buffers */
309 lock_ReleaseRead(&buf_globalLock);
314 /* incremental sync daemon. Writes all dirty buffers every 5000 ms */
316 buf_IncrSyncer(void * parm)
321 while (buf_ShutdownFlag == 0) {
323 i = SleepEx(5000, 1);
330 wasDirty = buf_Sync(1);
331 } /* whole daemon's while loop */
338 buf_ValidateBuffers(void)
340 cm_buf_t * bp, *bpf, *bpa, *bpb;
341 afs_uint64 countb = 0, countf = 0, counta = 0, countr = 0;
343 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
344 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
345 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
346 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
350 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
351 if (bp->magic != CM_BUF_MAGIC) {
352 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
353 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
359 if (countb > cm_data.buf_nbuffers) {
360 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
361 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
366 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
367 if (bp->magic != CM_BUF_MAGIC) {
368 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
369 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
375 if (countf > cm_data.buf_nbuffers) {
376 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
377 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
382 for ( bp = cm_data.buf_redirListp; bp; bp = (cm_buf_t *) osi_QNext(&bp->q)) {
383 if (!(bp->qFlags & CM_BUF_QREDIR)) {
384 afsi_log("CM_BUF_QREDIR not set on cm_buf_t in buf_redirListp");
385 fprintf(stderr, "CM_BUF_QREDIR not set on cm_buf_t in buf_redirListp");
389 if (countr > cm_data.buf_nbuffers) {
390 afsi_log("cm_ValidateBuffers failure: countr > cm_data.buf_nbuffers");
391 fprintf(stderr, "cm_ValidateBuffers failure: countr > cm_data.buf_nbuffers\n");
396 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
397 if (bp->magic != CM_BUF_MAGIC) {
398 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
399 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
405 if (counta > cm_data.buf_nbuffers) {
406 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
407 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
412 if (countb != countf) {
413 afsi_log("cm_ValidateBuffers failure: countb != countf");
414 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
418 if (counta != cm_data.buf_nbuffers) {
419 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
420 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
427 void buf_Shutdown(void)
429 /* disable the buf_IncrSyncer() threads */
430 buf_ShutdownFlag = 1;
432 /* then force all dirty buffers to the file servers */
436 /* initialize the buffer package; called with no locks
437 * held during the initialization phase.
439 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
441 static osi_once_t once;
444 pthread_attr_t tattr;
451 cm_data.buf_nbuffers = nbuffers;
453 /* Have to be able to reserve a whole chunk */
454 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
455 return CM_ERROR_TOOFEWBUFS;
458 /* recall for callouts */
461 if (osi_Once(&once)) {
462 /* initialize global locks */
463 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock", LOCK_HIERARCHY_BUF_GLOBAL);
464 lock_InitializeMutex(&buf_rdrReleaseExtentsLock, "RDR Release Extents lock", LOCK_HIERARCHY_RDR_EXTENTS);
467 /* remember this for those who want to reset it */
468 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
470 /* lower hash size to a prime number */
471 cm_data.buf_hashSize = cm_NextHighestPowerOf2((afs_uint32)(cm_data.buf_nbuffers/7));
473 /* create hash table */
474 memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
476 /* another hash table */
477 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
479 /* create buffer headers and put in free list */
480 bp = cm_data.bufHeaderBaseAddress;
481 data = cm_data.bufDataBaseAddress;
482 cm_data.buf_allp = NULL;
484 for (i=0; i<cm_data.buf_nbuffers; i++) {
485 osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress,
486 "invalid cm_buf_t address");
487 osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
488 "invalid cm_buf_t data address");
490 /* allocate and zero some storage */
491 memset(bp, 0, sizeof(cm_buf_t));
492 bp->magic = CM_BUF_MAGIC;
493 /* thread on list of all buffers */
494 bp->allp = cm_data.buf_allp;
495 cm_data.buf_allp = bp;
497 osi_QAddH( (osi_queue_t **) &cm_data.buf_freeListp,
498 (osi_queue_t **) &cm_data.buf_freeListEndp,
500 _InterlockedOr(&bp->qFlags, CM_BUF_QINLRU);
501 buf_IncrementFreeCount();
502 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
504 /* grab appropriate number of bytes from aligned zone */
509 data += cm_data.buf_blockSize;
512 /* none reserved at first */
513 cm_data.buf_reservedBufs = 0;
515 /* just for safety's sake */
516 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
518 bp = cm_data.bufHeaderBaseAddress;
519 data = cm_data.bufDataBaseAddress;
521 lock_ObtainWrite(&buf_globalLock);
522 for (i=0; i<cm_data.buf_nbuffers; i++) {
523 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
526 bp->waitRequests = 0;
527 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
529 if (bp->qFlags & CM_BUF_QREDIR) {
531 * extent was not returned by the file system driver.
534 buf_RemoveFromRedirQueue(NULL, bp);
535 bp->dataVersion = CM_BUF_VERSION_BAD;
536 bp->redirq.nextp = bp->redirq.prevp = NULL;
537 bp->redirLastAccess = 0;
538 bp->redirReleaseRequested = 0;
539 buf_ReleaseLocked(bp, TRUE);
545 * There should be nothing left in cm_data.buf_redirListp
546 * but double check just to be sure.
548 for ( bp = cm_data.buf_redirListp;
550 bp = cm_data.buf_redirListp)
553 * extent was not returned by the file system driver.
556 buf_RemoveFromRedirQueue(NULL, bp);
557 bp->dataVersion = CM_BUF_VERSION_BAD;
558 bp->redirq.nextp = bp->redirq.prevp = NULL;
559 bp->redirLastAccess = 0;
560 bp->redirReleaseRequested = 0;
561 buf_ReleaseLocked(bp, TRUE);
563 lock_ReleaseWrite(&buf_globalLock);
567 buf_ValidateBufQueues();
571 /* init the buffer trace log */
572 buf_logp = osi_LogCreate("buffer", 1000);
573 osi_LogEnable(buf_logp);
578 /* and create the incr-syncer */
579 pthread_attr_init(&tattr);
580 pthread_attr_setdetachstate(&tattr, PTHREAD_CREATE_DETACHED);
582 pstatus = pthread_create(&phandle, &tattr, buf_IncrSyncer, 0);
583 osi_assertx(pstatus == 0, "buf: can't create incremental sync proc");
585 pthread_attr_destroy(&tattr);
589 buf_ValidateBufQueues();
594 /* add nbuffers to the buffer pool, if possible.
595 * Called with no locks held.
597 long buf_AddBuffers(afs_uint64 nbuffers)
599 /* The size of a virtual cache cannot be changed after it has
600 * been created. Subsequent calls to MapViewofFile() with
601 * an existing mapping object name would not allow the
602 * object to be resized. Return failure immediately.
604 * A similar problem now occurs with the persistent cache
605 * given that the memory mapped file now contains a complex
608 afsi_log("request to add %d buffers to the existing cache of size %d denied",
609 nbuffers, cm_data.buf_nbuffers);
611 return CM_ERROR_INVAL;
614 /* interface to set the number of buffers to an exact figure.
615 * Called with no locks held.
617 long buf_SetNBuffers(afs_uint64 nbuffers)
620 return CM_ERROR_INVAL;
621 if (nbuffers == cm_data.buf_nbuffers)
623 else if (nbuffers > cm_data.buf_nbuffers)
624 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
626 return CM_ERROR_INVAL;
629 /* wait for reading or writing to clear; called with write-locked
630 * buffer and unlocked scp and returns with locked buffer.
632 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
637 osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
638 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
641 /* if no IO is happening, we're done */
642 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
645 /* otherwise I/O is happening, but some other thread is waiting for
646 * the I/O already. Wait for that guy to figure out what happened,
647 * and then check again.
649 if ( bp->flags & CM_BUF_WAITING ) {
652 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
654 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
655 _InterlockedOr(&bp->flags, CM_BUF_WAITING);
656 bp->waitCount = bp->waitRequests = 1;
658 osi_SleepM((LONG_PTR)bp, &bp->mx);
660 cm_UpdateServerPriority();
662 lock_ObtainMutex(&bp->mx);
663 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
665 if (bp->waitCount == 0) {
666 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
667 _InterlockedAnd(&bp->flags, ~CM_BUF_WAITING);
668 bp->waitRequests = 0;
672 if (scp = cm_FindSCache(&bp->fid))
676 lock_ObtainRead(&scp->rw);
677 if (!osi_QIsEmpty(&scp->waitQueueH)) {
678 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
679 osi_Wakeup((LONG_PTR)&scp->flags);
681 lock_ReleaseRead(&scp->rw);
685 /* if we get here, the IO is done, but we may have to wakeup people waiting for
686 * the I/O to complete. Do so.
688 if (bp->flags & CM_BUF_WAITING) {
689 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
690 osi_Wakeup((LONG_PTR) bp);
692 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
695 cm_ReleaseSCache(scp);
698 /* find a buffer, if any, for a particular file ID and offset. Assumes
699 * that buf_globalLock is write locked when called.
701 cm_buf_t *buf_FindLocked(struct cm_fid *fidp, osi_hyper_t *offsetp)
706 lock_AssertAny(&buf_globalLock);
708 i = BUF_HASH(fidp, offsetp);
709 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
710 if (cm_FidCmp(fidp, &bp->fid) == 0
711 && offsetp->LowPart == bp->offset.LowPart
712 && offsetp->HighPart == bp->offset.HighPart) {
718 /* return whatever we found, if anything */
722 /* find a buffer with offset *offsetp for vnode *scp. Called
723 * with no locks held.
725 cm_buf_t *buf_Find(struct cm_fid *fidp, osi_hyper_t *offsetp)
729 lock_ObtainRead(&buf_globalLock);
730 bp = buf_FindLocked(fidp, offsetp);
731 lock_ReleaseRead(&buf_globalLock);
736 /* find a buffer, if any, for a particular file ID and offset. Assumes
737 * that buf_globalLock is write locked when called. Uses the all buffer
740 cm_buf_t *buf_FindAllLocked(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
745 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
746 if (cm_FidCmp(fidp, &bp->fid) == 0
747 && offsetp->LowPart == bp->offset.LowPart
748 && offsetp->HighPart == bp->offset.HighPart) {
754 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
755 if (cm_FidCmp(fidp, &bp->fid) == 0) {
758 fileOffset = offsetp->QuadPart + cm_data.baseAddress;
759 if (fileOffset == bp->datap) {
766 /* return whatever we found, if anything */
770 /* find a buffer with offset *offsetp for vnode *scp. Called
771 * with no locks held. Use the all buffer list.
773 cm_buf_t *buf_FindAll(struct cm_fid *fidp, osi_hyper_t *offsetp, afs_uint32 flags)
777 lock_ObtainRead(&buf_globalLock);
778 bp = buf_FindAllLocked(fidp, offsetp, flags);
779 lock_ReleaseRead(&buf_globalLock);
784 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
787 * Makes sure that there's only one person writing this block
788 * at any given time, and also ensures that the log is forced sufficiently far,
789 * if this buffer contains logged data.
791 * Returns non-zero if the buffer was dirty.
793 * 'scp' may or may not be NULL. If it is not NULL, the FID for both cm_scache_t
794 * and cm_buf_t must match.
796 afs_uint32 buf_CleanLocked(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp,
797 afs_uint32 flags, afs_uint32 *pisdirty)
800 afs_uint32 isdirty = 0;
804 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
805 osi_assertx(scp == NULL || cm_FidCmp(&scp->fid, &bp->fid) == 0, "scp fid != bp fid");
808 * If the matching cm_scache_t was not provided as a parameter
809 * we must either find one or allocate a new one. It is possible
810 * that the cm_scache_t was recycled out of the cache even though
811 * a cm_buf_t with the same FID is in the cache.
814 cm_GetSCache(&bp->fid, NULL, &scp,
815 bp->userp ? bp->userp : cm_rootUserp,
821 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
823 lock_ReleaseMutex(&bp->mx);
827 * If we didn't find a cm_scache_t object for bp->fid it means
828 * that we no longer have that FID in the cache. It does not
829 * mean that the object does not exist in the cell. That may
830 * in fact be the case but we don't know that until we attempt
831 * a FetchStatus on the FID.
833 osi_Log1(buf_logp, "buf_CleanLocked unable to start I/O - scp not found buf 0x%p", bp);
834 code = CM_ERROR_NOSUCHFILE;
836 osi_Log2(buf_logp, "buf_CleanLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
839 LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
841 * Only specify the dirty length of the current buffer in the call
842 * to cm_BufWrite(). It is the responsibility of cm_BufWrite()
843 * to determine if it is appropriate to fill a full chunk of data
844 * when storing to the file server.
846 code = (*cm_buf_opsp->Writep)(scp, &offset,
848 flags, bp->userp, reqp);
849 osi_Log3(buf_logp, "buf_CleanLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
851 lock_ObtainMutex(&bp->mx);
852 /* if the Write routine returns No Such File, clear the dirty flag
853 * because we aren't going to be able to write this data to the file
856 if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD || code == CM_ERROR_NOACCESS ||
857 code == CM_ERROR_QUOTA || code == CM_ERROR_SPACE || code == CM_ERROR_TOOBIG ||
858 code == CM_ERROR_READONLY || code == CM_ERROR_NOSUCHPATH || code == EIO){
859 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
860 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
861 bp->dirty_length = 0;
863 bp->dataVersion = CM_BUF_VERSION_BAD;
869 /* Disk cache support */
870 /* write buffer to disk cache (synchronous for now) */
871 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
872 #endif /* DISKCACHE95 */
874 /* if we get here and retries are not permitted
875 * then we need to exit this loop regardless of
876 * whether or not we were able to clear the dirty bit
878 if (reqp->flags & CM_REQ_NORETRY)
881 /* Ditto if the hardDeadTimeout or idleTimeout was reached */
882 if (code == CM_ERROR_TIMEDOUT || code == CM_ERROR_ALLDOWN ||
883 code == CM_ERROR_ALLBUSY || code == CM_ERROR_ALLOFFLINE ||
884 code == CM_ERROR_CLOCKSKEW) {
890 cm_ReleaseSCache(scp);
892 /* if someone was waiting for the I/O that just completed or failed,
895 if (bp->flags & CM_BUF_WAITING) {
896 /* turn off flags and wakeup users */
897 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
898 osi_Wakeup((LONG_PTR) bp);
907 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
908 * recycles the buffer, and leaves it ready for reuse with a ref count of 0.
909 * The buffer must already be clean, and no I/O should be happening to it.
911 void buf_Recycle(cm_buf_t *bp)
916 cm_buf_t *prevBp, *nextBp;
918 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
920 osi_assertx(!(bp->qFlags & CM_BUF_QREDIR), "can't recycle redir held buffers");
922 /* if we get here, we know that the buffer still has a 0 ref count,
923 * and that it is clean and has no currently pending I/O. This is
924 * the dude to return.
925 * Remember that as long as the ref count is 0, we know that we won't
926 * have any lock conflicts, so we can grab the buffer lock out of
927 * order in the locking hierarchy.
929 osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
930 bp, bp->offset.HighPart, bp->offset.LowPart);
932 osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
933 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
934 "incorrect cm_buf_t flags");
935 lock_AssertWrite(&buf_globalLock);
937 if (bp->qFlags & CM_BUF_QINHASH) {
938 /* Remove from hash */
940 i = BUF_HASH(&bp->fid, &bp->offset);
941 lbpp = &(cm_data.buf_scacheHashTablepp[i]);
942 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = tbp->hashp) {
947 /* we better find it */
948 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
950 *lbpp = bp->hashp; /* hash out */
953 /* Remove from file hash */
955 i = BUF_FILEHASH(&bp->fid);
956 prevBp = bp->fileHashBackp;
957 bp->fileHashBackp = NULL;
958 nextBp = bp->fileHashp;
959 bp->fileHashp = NULL;
961 prevBp->fileHashp = nextBp;
963 cm_data.buf_fileHashTablepp[i] = nextBp;
965 nextBp->fileHashBackp = prevBp;
967 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINHASH);
970 /* make the fid unrecognizable */
971 memset(&bp->fid, 0, sizeof(cm_fid_t));
973 /* clean up junk flags */
974 _InterlockedAnd(&bp->flags, ~(CM_BUF_EOF | CM_BUF_ERROR));
975 bp->dataVersion = CM_BUF_VERSION_BAD; /* unknown so far */
980 * buf_RDRShakeAnExtentFree
981 * called with buf_globalLock read locked
984 buf_RDRShakeAnExtentFree(cm_buf_t *rbp, cm_req_t *reqp)
987 LARGE_INTEGER heldExtents = {0,0};
988 AFSFileExtentCB extentList[1];
989 DWORD extentCount = 0;
992 if (!(rbp->qFlags & CM_BUF_QREDIR))
995 lock_ReleaseRead(&buf_globalLock);
997 if (!lock_TryMutex(&buf_rdrReleaseExtentsLock)) {
998 osi_Log0(afsd_logp, "Waiting for prior RDR_RequestExtentRelease request to complete");
999 if (reqp->flags & CM_REQ_NORETRY) {
1000 code = CM_ERROR_WOULDBLOCK;
1004 lock_ObtainMutex(&buf_rdrReleaseExtentsLock);
1007 extentList[0].Flags = 0;
1008 extentList[0].Length = cm_data.blockSize;
1009 extentList[0].FileOffset.QuadPart = rbp->offset.QuadPart;
1010 extentList[0].CacheOffset.QuadPart = rbp->datap - cm_data.baseAddress;
1013 code = RDR_RequestExtentRelease(&rbp->fid, heldExtents, extentCount, extentList);
1015 lock_ReleaseMutex(&buf_rdrReleaseExtentsLock);
1018 lock_ObtainRead(&buf_globalLock);
1023 * buf_RDRShakeFileExtentsFree
1024 * requests all extents held by the redirector to be returned for
1025 * the specified cm_scache_t. This function is called with no
1029 buf_RDRShakeFileExtentsFree(cm_scache_t *rscp, cm_req_t *reqp)
1031 afs_uint32 code = 0;
1032 afs_uint64 n_redir = 0;
1034 if (!lock_TryMutex(&buf_rdrReleaseExtentsLock)) {
1035 osi_Log0(afsd_logp, "Waiting for prior RDR_RequestExtentRelease request to complete");
1036 if (reqp->flags & CM_REQ_NORETRY)
1037 return CM_ERROR_WOULDBLOCK;
1039 lock_ObtainMutex(&buf_rdrReleaseExtentsLock);
1042 for ( code = CM_ERROR_RETRY; code == CM_ERROR_RETRY; ) {
1043 LARGE_INTEGER heldExtents = {0,0};
1044 AFSFileExtentCB extentList[1024];
1045 DWORD extentCount = 0;
1049 /* only retry if a call to RDR_RequestExtentRelease says to */
1051 lock_ObtainWrite(&buf_globalLock);
1053 if (rscp->redirBufCount == 0)
1055 lock_ReleaseWrite(&buf_globalLock);
1060 for ( srbp = redirq_to_cm_buf_t(rscp->redirQueueT);
1062 srbp = ((code == 0 && extentCount == 0) ? redirq_to_cm_buf_t(rscp->redirQueueT) :
1063 redirq_to_cm_buf_t(osi_QPrev(&srbp->redirq))))
1065 extentList[extentCount].Flags = 0;
1066 extentList[extentCount].Length = cm_data.blockSize;
1067 extentList[extentCount].FileOffset.QuadPart = srbp->offset.QuadPart;
1068 extentList[extentCount].CacheOffset.QuadPart = srbp->datap - cm_data.baseAddress;
1069 srbp->redirReleaseRequested = now;
1072 if (extentCount == 1024) {
1073 lock_ReleaseWrite(&buf_globalLock);
1074 heldExtents.QuadPart = cm_data.buf_redirCount;
1075 code = RDR_RequestExtentRelease(&rscp->fid, heldExtents, extentCount, extentList);
1077 if (code == CM_ERROR_RETRY) {
1079 * The redirector either is not holding the extents or cannot let them
1080 * go because they are otherwise in use. At the moment, do nothing.
1086 lock_ObtainWrite(&buf_globalLock);
1089 lock_ReleaseWrite(&buf_globalLock);
1091 if (code == 0 && extentCount > 0) {
1092 heldExtents.QuadPart = cm_data.buf_redirCount;
1093 code = RDR_RequestExtentRelease(&rscp->fid, heldExtents, extentCount, extentList);
1096 if ((code == CM_ERROR_RETRY) && (reqp->flags & CM_REQ_NORETRY)) {
1097 code = CM_ERROR_WOULDBLOCK;
1101 lock_ReleaseMutex(&buf_rdrReleaseExtentsLock);
1106 buf_RDRShakeSomeExtentsFree(cm_req_t *reqp, afs_uint32 oneFid, afs_uint32 minage)
1108 afs_uint32 code = 0;
1110 if (!lock_TryMutex(&buf_rdrReleaseExtentsLock)) {
1111 if (reqp->flags & CM_REQ_NORETRY)
1112 return CM_ERROR_WOULDBLOCK;
1114 osi_Log0(afsd_logp, "Waiting for prior RDR_RequestExtentRelease request to complete");
1115 lock_ObtainMutex(&buf_rdrReleaseExtentsLock);
1118 for ( code = CM_ERROR_RETRY; code == CM_ERROR_RETRY; ) {
1119 LARGE_INTEGER heldExtents;
1120 AFSFileExtentCB extentList[1024];
1121 DWORD extentCount = 0;
1122 cm_buf_t *rbp, *srbp;
1125 BOOL locked = FALSE;
1127 /* only retry if a call to RDR_RequestExtentRelease says to */
1129 lock_ObtainWrite(&buf_globalLock);
1132 for ( rbp = cm_data.buf_redirListEndp;
1133 code == 0 && rbp && (!oneFid || extentCount == 0);
1134 rbp = (cm_buf_t *) osi_QPrev(&rbp->q))
1139 if (rbp->redirLastAccess >= rbp->redirReleaseRequested) {
1140 rscp = cm_FindSCache(&rbp->fid);
1145 for ( srbp = redirq_to_cm_buf_t(rscp->redirQueueT);
1146 srbp && extentCount < 1024;
1147 srbp = redirq_to_cm_buf_t(osi_QPrev(&srbp->redirq)))
1150 * Do not request a release if we have already done so
1151 * or if the extent was delivered to windows less than
1152 * 'minage' seconds ago.
1154 if (srbp->redirLastAccess >= srbp->redirReleaseRequested &&
1155 srbp->redirLastAccess < now - minage) {
1156 extentList[extentCount].Flags = 0;
1157 extentList[extentCount].Length = cm_data.blockSize;
1158 extentList[extentCount].FileOffset.QuadPart = srbp->offset.QuadPart;
1159 extentList[extentCount].CacheOffset.QuadPart = srbp->datap - cm_data.baseAddress;
1160 srbp->redirReleaseRequested = now;
1164 cm_ReleaseSCache(rscp);
1167 if ( !oneFid && extentCount > 0) {
1169 lock_ReleaseWrite(&buf_globalLock);
1172 heldExtents.QuadPart = cm_data.buf_redirCount;
1173 code = RDR_RequestExtentRelease(&rbp->fid, heldExtents, extentCount, extentList);
1176 lock_ObtainWrite(&buf_globalLock);
1181 lock_ReleaseWrite(&buf_globalLock);
1184 heldExtents.QuadPart = cm_data.buf_redirCount;
1185 if (rbp && extentCount)
1186 code = RDR_RequestExtentRelease(&rbp->fid, heldExtents, extentCount, extentList);
1188 code = RDR_RequestExtentRelease(NULL, heldExtents, 1024, NULL);
1194 if ((code == CM_ERROR_RETRY) && (reqp->flags & CM_REQ_NORETRY)) {
1195 code = CM_ERROR_WOULDBLOCK;
1199 lock_ReleaseMutex(&buf_rdrReleaseExtentsLock);
1203 /* returns 0 if the buffer does not exist, and non-0 if it does */
1205 buf_ExistsLocked(struct cm_scache *scp, osi_hyper_t *offsetp)
1209 if (bp = buf_FindLocked(&scp->fid, offsetp)) {
1210 /* Do not call buf_ReleaseLocked() because we
1211 * do not want to allow the buffer to be added
1214 afs_int32 refCount = InterlockedDecrement(&bp->refCount);
1215 #ifdef DEBUG_REFCOUNT
1216 osi_Log2(afsd_logp,"buf_ExistsLocked bp 0x%p ref %d", bp, refCount);
1217 afsi_log("%s:%d buf_ExistsLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, refCount);
1219 return CM_BUF_EXISTS;
1225 /* recycle a buffer, removing it from the free list, hashing in its new identity
1226 * and returning it write-locked so that no one can use it. Called without
1227 * any locks held, and can return an error if it loses the race condition and
1228 * finds that someone else created the desired buffer.
1230 * If success is returned, the buffer is returned write-locked.
1232 * May be called with null scp and offsetp, if we're just trying to reclaim some
1233 * space from the buffer pool. In that case, the buffer will be returned
1234 * without being hashed into the hash table.
1236 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
1238 cm_buf_t *bp; /* buffer we're dealing with */
1239 cm_buf_t *nextBp; /* next buffer in file hash chain */
1240 afs_uint32 i; /* temp */
1241 afs_uint64 n_bufs, n_nonzero, n_busy, n_dirty, n_own, n_redir;
1244 buf_ValidateBufQueues();
1245 #endif /* TESTING */
1256 lock_ObtainRead(&scp->bufCreateLock);
1257 lock_ObtainWrite(&buf_globalLock);
1258 /* check to see if we lost the race */
1259 if (buf_ExistsLocked(scp, offsetp)) {
1260 lock_ReleaseWrite(&buf_globalLock);
1261 lock_ReleaseRead(&scp->bufCreateLock);
1262 return CM_BUF_EXISTS;
1265 /* does this fix the problem below? it's a simple solution. */
1266 if (!cm_data.buf_freeListEndp)
1268 lock_ReleaseWrite(&buf_globalLock);
1269 lock_ReleaseRead(&scp->bufCreateLock);
1271 if ( RDR_Initialized )
1274 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
1279 /* for debugging, assert free list isn't empty, although we
1280 * really should try waiting for a running tranasction to finish
1281 * instead of this; or better, we should have a transaction
1282 * throttler prevent us from entering this situation.
1284 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
1286 /* look at all buffers in free list, some of which may temp.
1287 * have high refcounts and which then should be skipped,
1288 * starting cleaning I/O for those which are dirty. If we find
1289 * a clean buffer, we rehash it, lock it and return it.
1291 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1297 /* check to see if it really has zero ref count. This
1298 * code can bump refcounts, at least, so it may not be
1301 if (bp->refCount > 0) {
1306 /* we don't have to lock buffer itself, since the ref
1307 * count is 0 and we know it will stay zero as long as
1308 * we hold the global lock.
1311 /* don't recycle someone in our own chunk */
1312 if (!cm_FidCmp(&bp->fid, &scp->fid) &&
1313 bp->dataVersion >= scp->bufDataVersionLow &&
1314 bp->dataVersion <= scp->dataVersion &&
1315 (bp->offset.LowPart & (-cm_chunkSize)) == (offsetp->LowPart & (-cm_chunkSize))) {
1320 /* if this page is being filled (!) or cleaned, see if
1321 * the I/O has completed. If not, skip it, otherwise
1322 * do the final processing for the I/O.
1324 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
1325 /* probably shouldn't do this much work while
1326 * holding the big lock? Watch for contention
1333 /* leave the buffer alone if held by the redirector */
1334 if (bp->qFlags & CM_BUF_QREDIR) {
1339 if (bp->flags & CM_BUF_DIRTY) {
1342 /* protect against cleaning the same buffer more than once. */
1346 /* if the buffer is dirty, start cleaning it and
1347 * move on to the next buffer. We do this with
1348 * just the lock required to minimize contention
1352 lock_ReleaseWrite(&buf_globalLock);
1353 lock_ReleaseRead(&scp->bufCreateLock);
1356 * grab required lock and clean.
1357 * previously the claim was that the cleaning
1358 * operation was async which it is not. It would
1359 * be a good idea to use an async mechanism here
1360 * but there is none at the moment other than
1361 * the buf_IncrSyncer() thread.
1363 if (cm_FidCmp(&scp->fid, &bp->fid) == 0)
1364 buf_Clean(scp, bp, reqp, 0, NULL);
1366 buf_Clean(NULL, bp, reqp, 0, NULL);
1368 /* now put it back and go around again */
1371 /* but first obtain the locks we gave up
1372 * before the buf_CleanAsync() call */
1373 lock_ObtainRead(&scp->bufCreateLock);
1374 lock_ObtainWrite(&buf_globalLock);
1377 * Since we dropped the locks we need to verify that
1378 * another thread has not allocated the buffer for us.
1380 if (buf_ExistsLocked(scp, offsetp)) {
1381 lock_ReleaseWrite(&buf_globalLock);
1382 lock_ReleaseRead(&scp->bufCreateLock);
1383 return CM_BUF_EXISTS;
1387 * We just cleaned this buffer so we need to
1388 * restart the loop with this buffer so it
1389 * can be retested. Set 'cleaned' so we
1390 * do not attempt another call to buf_Clean()
1391 * if the prior attempt failed.
1397 osi_Log3(afsd_logp, "buf_GetNewLocked: scp 0x%p examined %u buffers before recycling bufp 0x%p",
1399 osi_Log4(afsd_logp, "... nonzero %u; own %u; busy %u; dirty %u", n_nonzero, n_own, n_busy, n_dirty);
1401 /* if we get here, we know that the buffer still has a 0
1402 * ref count, and that it is clean and has no currently
1403 * pending I/O. This is the dude to return.
1404 * Remember that as long as the ref count is 0, we know
1405 * that we won't have any lock conflicts, so we can grab
1406 * the buffer lock out of order in the locking hierarchy.
1410 /* now hash in as our new buffer, and give it the
1411 * appropriate label, if requested.
1414 lock_AssertWrite(&buf_globalLock);
1416 _InterlockedOr(&bp->qFlags, CM_BUF_QINHASH);
1421 bp->offset = *offsetp;
1422 i = BUF_HASH(&scp->fid, offsetp);
1423 bp->hashp = cm_data.buf_scacheHashTablepp[i];
1424 cm_data.buf_scacheHashTablepp[i] = bp;
1425 i = BUF_FILEHASH(&scp->fid);
1426 nextBp = cm_data.buf_fileHashTablepp[i];
1427 bp->fileHashp = nextBp;
1428 bp->fileHashBackp = NULL;
1430 nextBp->fileHashBackp = bp;
1431 cm_data.buf_fileHashTablepp[i] = bp;
1434 /* we should remove it from the lru queue. It better still be there,
1435 * since we've held the global (big) lock since we found it there.
1437 osi_assertx(bp->qFlags & CM_BUF_QINLRU,
1438 "buf_GetNewLocked: LRU screwup");
1440 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
1441 (osi_queue_t **) &cm_data.buf_freeListEndp,
1443 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1444 buf_DecrementFreeCount();
1446 /* prepare to return it. Give it a refcount */
1447 InterlockedIncrement(&bp->refCount);
1448 #ifdef DEBUG_REFCOUNT
1449 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, 1);
1450 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, 1);
1452 /* grab the mutex so that people don't use it
1453 * before the caller fills it with data. Again, no one
1454 * should have been able to get to this dude to lock it.
1456 if (!lock_TryMutex(&bp->mx)) {
1457 osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked. refCount %d should be 0",
1459 osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
1462 lock_ReleaseWrite(&buf_globalLock);
1463 lock_ReleaseRead(&scp->bufCreateLock);
1468 buf_ValidateBufQueues();
1469 #endif /* TESTING */
1471 } /* for all buffers in lru queue */
1472 lock_ReleaseWrite(&buf_globalLock);
1473 lock_ReleaseRead(&scp->bufCreateLock);
1475 osi_Log2(afsd_logp, "buf_GetNewLocked: Free Buffer List has %u buffers none free; redir %u", n_bufs, n_redir);
1476 osi_Log4(afsd_logp, "... nonzero %u; own %u; busy %u; dirty %u", n_nonzero, n_own, n_busy, n_dirty);
1478 if (RDR_Initialized) {
1481 code = buf_RDRShakeSomeExtentsFree(reqp, TRUE, 2 /* seconds */);
1483 case CM_ERROR_RETRY:
1486 case CM_ERROR_WOULDBLOCK:
1487 return CM_ERROR_WOULDBLOCK;
1491 Sleep(100); /* give some time for a buffer to be freed */
1492 } /* while loop over everything */
1497 * get a page, returning it held but unlocked. the page may or may not
1498 * contain valid data.
1500 * The scp must be unlocked when passed in unlocked.
1502 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
1506 osi_hyper_t pageOffset;
1507 unsigned long tcount;
1511 cm_diskcache_t *dcp;
1512 #endif /* DISKCACHE95 */
1515 pageOffset.HighPart = offsetp->HighPart;
1516 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
1520 buf_ValidateBufQueues();
1521 #endif /* TESTING */
1523 bp = buf_Find(&scp->fid, &pageOffset);
1525 /* lock it and break out */
1526 lock_ObtainMutex(&bp->mx);
1529 /* touch disk chunk to update LRU info */
1530 diskcache_Touch(bp->dcp);
1531 #endif /* DISKCACHE95 */
1535 /* otherwise, we have to create a page */
1536 code = buf_GetNewLocked(scp, &pageOffset, reqp, &bp);
1539 /* the requested buffer was created */
1544 * the requested buffer existed by the time the
1545 * scp->bufCreateLock and buf_globalLock could be obtained.
1546 * loop again and permit buf_Find() to obtain a reference.
1551 * the requested buffer could not be created.
1552 * return the error to the caller.
1555 buf_ValidateBufQueues();
1556 #endif /* TESTING */
1559 } /* big while loop */
1561 /* if we get here, we have a locked buffer that may have just been
1562 * created, in which case it needs to be filled with data.
1565 /* load the page; freshly created pages should be idle */
1566 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
1569 * start the I/O; may drop lock. as of this writing, the only
1570 * implementation of Readp is cm_BufRead() which simply sets
1571 * tcount to 0 and returns success.
1573 _InterlockedOr(&bp->flags, CM_BUF_READING);
1574 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
1577 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
1578 bp->dcp = dcp; /* pointer to disk cache struct. */
1579 #endif /* DISKCACHE95 */
1582 /* failure or queued */
1584 /* unless cm_BufRead() is altered, this path cannot be hit */
1585 if (code != ERROR_IO_PENDING) {
1587 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
1588 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1589 if (bp->flags & CM_BUF_WAITING) {
1590 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1591 osi_Wakeup((LONG_PTR) bp);
1593 lock_ReleaseMutex(&bp->mx);
1596 buf_ValidateBufQueues();
1597 #endif /* TESTING */
1602 * otherwise, I/O completed instantly and we're done, except
1603 * for padding the xfr out with 0s and checking for EOF
1605 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1606 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1608 _InterlockedOr(&bp->flags, CM_BUF_EOF);
1610 _InterlockedAnd(&bp->flags, ~CM_BUF_READING);
1611 if (bp->flags & CM_BUF_WAITING) {
1612 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1613 osi_Wakeup((LONG_PTR) bp);
1618 /* wait for reads, either that which we started above, or that someone
1619 * else started. We don't care if we return a buffer being cleaned.
1621 if (bp->flags & CM_BUF_READING)
1622 buf_WaitIO(scp, bp);
1624 /* once it has been read once, we can unlock it and return it, still
1625 * with its refcount held.
1627 lock_ReleaseMutex(&bp->mx);
1630 /* now remove from queue; will be put in at the head (farthest from
1631 * being recycled) when we're done in buf_Release.
1633 lock_ObtainWrite(&buf_globalLock);
1634 if (bp->qFlags & CM_BUF_QINLRU) {
1635 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
1636 (osi_queue_t **) &cm_data.buf_freeListEndp,
1638 _InterlockedAnd(&bp->qFlags, ~CM_BUF_QINLRU);
1639 buf_DecrementFreeCount();
1641 lock_ReleaseWrite(&buf_globalLock);
1643 osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1644 bp, scp, offsetp->HighPart, offsetp->LowPart);
1646 buf_ValidateBufQueues();
1647 #endif /* TESTING */
1651 /* clean a buffer synchronously */
1652 afs_uint32 buf_Clean(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp, afs_uint32 flags, afs_uint32 *pisdirty)
1655 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1656 osi_assertx(!(flags & CM_BUF_WRITE_SCP_LOCKED), "scp->rw must not be held when calling buf_CleanAsync");
1658 lock_ObtainMutex(&bp->mx);
1659 code = buf_CleanLocked(scp, bp, reqp, flags, pisdirty);
1660 lock_ReleaseMutex(&bp->mx);
1665 /* wait for a buffer's cleaning to finish */
1666 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp, afs_uint32 locked)
1668 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1671 lock_ObtainMutex(&bp->mx);
1672 if (bp->flags & CM_BUF_WRITING) {
1673 buf_WaitIO(scp, bp);
1676 lock_ReleaseMutex(&bp->mx);
1679 /* set the dirty flag on a buffer, and set associated write-ahead log,
1680 * if there is one. Allow one to be added to a buffer, but not changed.
1682 * The buffer must be locked before calling this routine.
1684 void buf_SetDirty(cm_buf_t *bp, cm_req_t *reqp, afs_uint32 offset, afs_uint32 length, cm_user_t *userp)
1686 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1687 osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
1688 osi_assertx(userp != NULL, "userp is NULL");
1693 if (bp->flags & CM_BUF_DIRTY) {
1695 osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
1697 if (bp->dirty_offset <= offset) {
1698 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1699 /* dirty_length remains the same */
1701 bp->dirty_length = offset + length - bp->dirty_offset;
1703 } else /* bp->dirty_offset > offset */ {
1704 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1705 bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
1707 bp->dirty_length = length;
1709 bp->dirty_offset = offset;
1712 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1715 _InterlockedOr(&bp->flags, CM_BUF_DIRTY);
1717 /* and turn off EOF flag, since it has associated data now */
1718 _InterlockedAnd(&bp->flags, ~CM_BUF_EOF);
1720 bp->dirty_offset = offset;
1721 bp->dirty_length = length;
1724 * if the request is not from the afs redirector,
1725 * add to the dirty list. The redirector interface ensures
1726 * that a background store operation is queued for each and
1727 * every dirty extent that is released. Therefore, the
1728 * buf_IncrSyncer thread is not required to ensure that
1729 * dirty buffers are written to the file server.
1731 * we obtain a hold on the buffer for as long as it remains
1732 * in the list. buffers are only removed from the list by
1733 * the buf_IncrSyncer function regardless of when else the
1734 * dirty flag might be cleared.
1736 * This should never happen but just in case there is a bug
1737 * elsewhere, never add to the dirty list if the buffer is
1740 if (!(reqp->flags & CM_REQ_SOURCE_REDIR)) {
1741 lock_ObtainWrite(&buf_globalLock);
1742 if (!(bp->qFlags & CM_BUF_QINDL)) {
1744 if (!cm_data.buf_dirtyListp) {
1745 cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
1747 cm_data.buf_dirtyListEndp->dirtyp = bp;
1748 cm_data.buf_dirtyListEndp = bp;
1751 _InterlockedOr(&bp->qFlags, CM_BUF_QINDL);
1753 lock_ReleaseWrite(&buf_globalLock);
1757 /* and record the last writer */
1758 if (bp->userp != userp) {
1761 cm_ReleaseUser(bp->userp);
1766 /* clean all buffers, reset log pointers and invalidate all buffers.
1767 * Called with no locks held, and returns with same.
1769 * This function is guaranteed to clean and remove the log ptr of all the
1770 * buffers that were dirty or had non-zero log ptrs before the call was
1771 * made. That's sufficient to clean up any garbage left around by recovery,
1772 * which is all we're counting on this for; there may be newly created buffers
1773 * added while we're running, but that should be OK.
1775 * In an environment where there are no transactions (artificially imposed, for
1776 * example, when switching the database to raw mode), this function is used to
1777 * make sure that all updates have been written to the disk. In that case, we don't
1778 * really require that we forget the log association between pages and logs, but
1779 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1780 * have to worry about invalidating data in the buffers.
1782 * This function is used at the end of recovery as paranoia to get the recovered
1783 * database out to disk. It removes all references to the recovery log and cleans
1786 long buf_CleanAndReset(void)
1792 lock_ObtainRead(&buf_globalLock);
1793 for(i=0; i<cm_data.buf_hashSize; i++) {
1794 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
1795 if (bp->qFlags & CM_BUF_QREDIR) {
1796 osi_Log1(buf_logp,"buf_CleanAndReset buffer held by redirector bp 0x%p", bp);
1798 /* Request single extent from the redirector */
1799 buf_RDRShakeAnExtentFree(bp, &req);
1802 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1804 lock_ReleaseRead(&buf_globalLock);
1806 /* now no locks are held; clean buffer and go on */
1808 req.flags |= CM_REQ_NORETRY;
1810 buf_Clean(NULL, bp, &req, 0, NULL);
1811 buf_CleanWait(NULL, bp, FALSE);
1813 /* relock and release buffer */
1814 lock_ObtainRead(&buf_globalLock);
1815 buf_ReleaseLocked(bp, FALSE);
1817 } /* over one bucket */
1818 } /* for loop over all hash buckets */
1821 lock_ReleaseRead(&buf_globalLock);
1824 buf_ValidateBufQueues();
1825 #endif /* TESTING */
1827 /* and we're done */
1831 /* called without global lock being held, reserves buffers for callers
1832 * that need more than one held (not locked) at once.
1834 void buf_ReserveBuffers(afs_uint64 nbuffers)
1836 lock_ObtainWrite(&buf_globalLock);
1838 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1839 cm_data.buf_reserveWaiting = 1;
1840 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1841 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1842 lock_ObtainWrite(&buf_globalLock);
1845 cm_data.buf_reservedBufs += nbuffers;
1849 lock_ReleaseWrite(&buf_globalLock);
1852 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1856 lock_ObtainWrite(&buf_globalLock);
1857 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1861 cm_data.buf_reservedBufs += nbuffers;
1864 lock_ReleaseWrite(&buf_globalLock);
1868 /* called without global lock held, releases reservation held by
1869 * buf_ReserveBuffers.
1871 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1873 lock_ObtainWrite(&buf_globalLock);
1874 cm_data.buf_reservedBufs -= nbuffers;
1875 if (cm_data.buf_reserveWaiting) {
1876 cm_data.buf_reserveWaiting = 0;
1877 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1879 lock_ReleaseWrite(&buf_globalLock);
1882 /* truncate the buffers past sizep, zeroing out the page, if we don't
1883 * end on a page boundary.
1885 * Requires cm_bufCreateLock to be write locked.
1887 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1891 cm_buf_t *nbufp; /* next buffer, if didRelease */
1896 afs_uint32 invalidate = 0;
1898 /* assert that cm_bufCreateLock is held in write mode */
1899 lock_AssertWrite(&scp->bufCreateLock);
1901 i = BUF_FILEHASH(&scp->fid);
1903 lock_ObtainRead(&buf_globalLock);
1904 bufp = cm_data.buf_fileHashTablepp[i];
1906 lock_ReleaseRead(&buf_globalLock);
1910 buf_HoldLocked(bufp);
1911 lock_ReleaseRead(&buf_globalLock);
1914 lock_ObtainMutex(&bufp->mx);
1916 bufEnd.HighPart = 0;
1917 bufEnd.LowPart = cm_data.buf_blockSize;
1918 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1920 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1921 LargeIntegerLessThan(*sizep, bufEnd)) {
1922 buf_WaitIO(scp, bufp);
1924 lock_ObtainWrite(&scp->rw);
1926 /* make sure we have a callback (so we have the right value for
1927 * the length), and wait for it to be safe to do a truncate.
1929 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1930 CM_SCACHESYNC_NEEDCALLBACK
1931 | CM_SCACHESYNC_GETSTATUS
1932 | CM_SCACHESYNC_SETSIZE
1933 | CM_SCACHESYNC_BUFLOCKED);
1936 /* if we succeeded in our locking, and this applies to the right
1937 * file, and the truncate request overlaps the buffer either
1938 * totally or partially, then do something.
1940 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1941 && LargeIntegerLessThan(*sizep, bufEnd)) {
1944 /* destroy the buffer, turning off its dirty bit, if
1945 * we're truncating the whole buffer. Otherwise, set
1946 * the dirty bit, and clear out the tail of the buffer
1947 * if we just overlap some.
1949 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1950 /* truncating the entire page */
1951 if (reqp->flags & CM_REQ_SOURCE_REDIR) {
1953 * Implicitly clear the redirector flag
1954 * and release the matching hold.
1956 if (bufp->qFlags & CM_BUF_QREDIR) {
1957 osi_Log4(buf_logp,"buf_Truncate taking from file system bufp 0x%p vno 0x%x foffset 0x%x:%x",
1958 bufp, bufp->fid.vnode, bufp->offset.HighPart, bufp->offset.LowPart);
1959 lock_ObtainWrite(&buf_globalLock);
1960 if (bufp->qFlags & CM_BUF_QREDIR) {
1961 buf_RemoveFromRedirQueue(scp, bufp);
1962 buf_ReleaseLocked(bufp, TRUE);
1964 lock_ReleaseWrite(&buf_globalLock);
1969 _InterlockedAnd(&bufp->flags, ~CM_BUF_DIRTY);
1971 bufp->dirty_length = 0;
1972 bufp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1973 bufp->dirtyCounter++;
1976 /* don't set dirty, since dirty implies
1977 * currently up-to-date. Don't need to do this,
1978 * since we'll update the length anyway.
1980 * Zero out remainder of the page, in case we
1981 * seek and write past EOF, and make this data
1984 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1985 osi_assertx(bufferPos != 0, "non-zero bufferPos");
1986 memset(bufp->datap + bufferPos, 0,
1987 cm_data.buf_blockSize - bufferPos);
1991 cm_SyncOpDone( scp, bufp,
1992 CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
1993 | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
1995 lock_ReleaseWrite(&scp->rw);
1996 lock_ReleaseMutex(&bufp->mx);
1999 nbufp = bufp->fileHashp;
2003 /* This forces the loop to end and the error code
2004 * to be returned. */
2012 buf_ValidateBufQueues();
2013 #endif /* TESTING */
2015 if (invalidate && RDR_Initialized)
2016 RDR_InvalidateObject(scp->fid.cell, scp->fid.volume, scp->fid.vnode,
2017 scp->fid.unique, scp->fid.hash,
2018 scp->fileType, AFS_INVALIDATE_SMB);
2024 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
2027 cm_buf_t *bp; /* buffer we're hacking on */
2031 afs_uint32 stable = 0;
2033 i = BUF_FILEHASH(&scp->fid);
2036 lock_ObtainRead(&buf_globalLock);
2037 bp = cm_data.buf_fileHashTablepp[i];
2040 lock_ReleaseRead(&buf_globalLock);
2042 for (; bp; bp = nbp) {
2043 didRelease = 0; /* haven't released this buffer yet */
2045 /* clean buffer synchronously */
2046 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
2048 if (code == 0 && !stable && (bp->flags & CM_BUF_DIRTY)) {
2050 * we must stabilize the object to ensure that buffer
2051 * changes cannot occur while the flush is performed.
2052 * However, we do not want to Stabilize if we do not
2053 * need to because Stabilize obtains a callback.
2055 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
2056 stable = (code == 0);
2059 if (code == CM_ERROR_BADFD) {
2060 /* if the scp's FID is bad its because we received VNOVNODE
2061 * when attempting to FetchStatus before the write. This
2062 * page therefore contains data that can no longer be stored.
2064 lock_ObtainMutex(&bp->mx);
2065 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2066 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2067 bp->error = CM_ERROR_BADFD;
2068 bp->dirty_length = 0;
2069 bp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
2071 lock_ReleaseMutex(&bp->mx);
2072 } else if (!(scp->flags & CM_SCACHEFLAG_RO)) {
2077 lock_ObtainMutex(&bp->mx);
2079 /* start cleaning the buffer, and wait for it to finish */
2080 buf_CleanLocked(scp, bp, reqp, 0, NULL);
2081 buf_WaitIO(scp, bp);
2083 lock_ReleaseMutex(&bp->mx);
2086 /* actually, we only know that buffer is clean if ref
2087 * count is 1, since we don't have buffer itself locked.
2089 if (!(bp->flags & CM_BUF_DIRTY) && !(bp->qFlags & CM_BUF_QREDIR)) {
2090 lock_ObtainWrite(&buf_globalLock);
2091 if (!(bp->flags & CM_BUF_DIRTY) && !(bp->qFlags & CM_BUF_QREDIR)) {
2092 if (bp->refCount == 1) { /* bp is held above */
2093 nbp = bp->fileHashp;
2095 buf_HoldLocked(nbp);
2096 buf_ReleaseLocked(bp, TRUE);
2101 lock_ReleaseWrite(&buf_globalLock);
2107 lock_ObtainRead(&buf_globalLock);
2108 nbp = bp->fileHashp;
2110 buf_HoldLocked(nbp);
2111 buf_ReleaseLocked(bp, FALSE);
2112 lock_ReleaseRead(&buf_globalLock);
2114 } /* for loop over a bunch of buffers */
2117 (*cm_buf_opsp->Unstabilizep)(scp, userp);
2120 buf_ValidateBufQueues();
2121 #endif /* TESTING */
2127 /* Must be called with scp->rw held */
2128 long buf_InvalidateBuffers(cm_scache_t * scp)
2134 lock_AssertAny(&scp->rw);
2136 i = BUF_FILEHASH(&scp->fid);
2138 lock_ObtainRead(&buf_globalLock);
2140 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
2141 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
2142 bp->dataVersion = CM_BUF_VERSION_BAD;
2146 lock_ReleaseRead(&buf_globalLock);
2154 /* Must be called with scp->rw held */
2155 long buf_ForceDataVersion(cm_scache_t * scp, afs_uint64 fromVersion, afs_uint64 toVersion)
2161 lock_AssertAny(&scp->rw);
2163 i = BUF_FILEHASH(&scp->fid);
2165 lock_ObtainRead(&buf_globalLock);
2167 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
2168 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
2169 if (bp->dataVersion == fromVersion) {
2170 bp->dataVersion = toVersion;
2175 lock_ReleaseRead(&buf_globalLock);
2183 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
2187 cm_buf_t *bp; /* buffer we're hacking on */
2188 cm_buf_t *nbp; /* next one */
2191 if (RDR_Initialized && scp->redirBufCount > 0) {
2192 /* Retrieve all extents for this file from the redirector */
2193 buf_RDRShakeFileExtentsFree(scp, reqp);
2196 i = BUF_FILEHASH(&scp->fid);
2198 lock_ObtainRead(&buf_globalLock);
2199 bp = cm_data.buf_fileHashTablepp[i];
2202 lock_ReleaseRead(&buf_globalLock);
2203 for (; bp; bp = nbp) {
2204 /* clean buffer synchronously */
2205 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
2207 * If the buffer is held by the redirector we must fetch
2208 * it back in order to determine whether or not it is in
2211 lock_ObtainRead(&buf_globalLock);
2212 if (bp->qFlags & CM_BUF_QREDIR) {
2213 osi_Log1(buf_logp,"buf_CleanVnode buffer held by redirector bp 0x%p", bp);
2215 /* Retrieve single extent from the redirector */
2216 buf_RDRShakeAnExtentFree(bp, reqp);
2218 lock_ReleaseRead(&buf_globalLock);
2220 lock_ObtainMutex(&bp->mx);
2221 if ((bp->flags & CM_BUF_DIRTY)) {
2222 if (userp && userp != bp->userp) {
2225 cm_ReleaseUser(bp->userp);
2230 case CM_ERROR_NOSUCHFILE:
2231 case CM_ERROR_BADFD:
2232 case CM_ERROR_NOACCESS:
2233 case CM_ERROR_QUOTA:
2234 case CM_ERROR_SPACE:
2235 case CM_ERROR_TOOBIG:
2236 case CM_ERROR_READONLY:
2237 case CM_ERROR_NOSUCHPATH:
2240 * Apply the previous fatal error to this buffer.
2241 * Do not waste the time attempting to store to
2242 * the file server when we know it will fail.
2244 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2245 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2246 bp->dirty_length = 0;
2248 bp->dataVersion = CM_BUF_VERSION_BAD;
2251 case CM_ERROR_TIMEDOUT:
2252 case CM_ERROR_ALLDOWN:
2253 case CM_ERROR_ALLBUSY:
2254 case CM_ERROR_ALLOFFLINE:
2255 case CM_ERROR_CLOCKSKEW:
2256 /* do not mark the buffer in error state but do
2257 * not attempt to complete the rest either.
2261 code = buf_CleanLocked(scp, bp, reqp, 0, &wasDirty);
2262 if (bp->flags & CM_BUF_ERROR) {
2268 buf_CleanWait(scp, bp, TRUE);
2270 lock_ReleaseMutex(&bp->mx);
2273 lock_ObtainRead(&buf_globalLock);
2274 nbp = bp->fileHashp;
2276 buf_HoldLocked(nbp);
2277 buf_ReleaseLocked(bp, FALSE);
2278 lock_ReleaseRead(&buf_globalLock);
2279 } /* for loop over a bunch of buffers */
2282 buf_ValidateBufQueues();
2283 #endif /* TESTING */
2291 buf_ValidateBufQueues(void)
2293 cm_buf_t * bp, *bpb, *bpf, *bpa;
2294 afs_uint32 countf=0, countb=0, counta=0;
2296 lock_ObtainRead(&buf_globalLock);
2297 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
2298 if (bp->magic != CM_BUF_MAGIC)
2299 osi_panic("buf magic error",__FILE__,__LINE__);
2304 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
2305 if (bp->magic != CM_BUF_MAGIC)
2306 osi_panic("buf magic error",__FILE__,__LINE__);
2311 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
2312 if (bp->magic != CM_BUF_MAGIC)
2313 osi_panic("buf magic error",__FILE__,__LINE__);
2317 lock_ReleaseRead(&buf_globalLock);
2319 if (countb != countf)
2320 osi_panic("buf magic error",__FILE__,__LINE__);
2322 if (counta != cm_data.buf_nbuffers)
2323 osi_panic("buf magic error",__FILE__,__LINE__);
2325 #endif /* TESTING */
2327 /* dump the contents of the buf_scacheHashTablepp. */
2328 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
2335 if (cm_data.buf_scacheHashTablepp == NULL)
2339 lock_ObtainRead(&buf_globalLock);
2341 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n",
2342 cookie, cm_data.buf_hashSize);
2343 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2345 for (i = 0; i < cm_data.buf_hashSize; i++)
2347 for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp)
2349 StringCbPrintfA(output, sizeof(output),
2350 "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
2351 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
2352 "flags=0x%x, qFlags=0x%x cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
2353 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
2354 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
2355 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
2356 bp->cmFlags, bp->error, bp->refCount);
2357 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2361 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
2362 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2364 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
2365 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2366 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
2367 StringCbPrintfA(output, sizeof(output),
2368 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
2369 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
2370 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
2371 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
2372 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
2373 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
2374 bp->cmFlags, bp->error, bp->refCount);
2375 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2377 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
2378 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2380 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListp\r\n", cookie);
2381 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2382 for(bp = cm_data.buf_dirtyListp; bp; bp=bp->dirtyp) {
2383 StringCbPrintfA(output, sizeof(output),
2384 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
2385 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
2386 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
2387 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
2388 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
2389 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
2390 bp->cmFlags, bp->error, bp->refCount);
2391 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2393 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListp.\r\n", cookie);
2394 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
2397 lock_ReleaseRead(&buf_globalLock);
2401 void buf_ForceTrace(BOOL flush)
2410 len = GetTempPath(sizeof(buf)-10, buf);
2411 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
2412 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
2413 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
2414 if (handle == INVALID_HANDLE_VALUE) {
2415 osi_panic("Cannot create log file", __FILE__, __LINE__);
2417 osi_LogPrint(buf_logp, handle);
2419 FlushFileBuffers(handle);
2420 CloseHandle(handle);
2423 long buf_DirtyBuffersExist(cm_fid_t *fidp)
2426 afs_uint32 bcount = 0;
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->flags & CM_BUF_DIRTY)) {
2439 lock_ReleaseRead(&buf_globalLock);
2443 long buf_RDRBuffersExist(cm_fid_t *fidp)
2446 afs_uint32 bcount = 0;
2450 if (!RDR_Initialized)
2453 i = BUF_FILEHASH(fidp);
2455 lock_ObtainRead(&buf_globalLock);
2456 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
2457 if (!cm_FidCmp(fidp, &bp->fid) && (bp->qFlags & CM_BUF_QREDIR)) {
2462 lock_ReleaseRead(&buf_globalLock);
2466 long buf_ClearRDRFlag(cm_scache_t *scp, char *reason)
2468 cm_fid_t *fidp = &scp->fid;
2470 afs_uint32 bcount = 0;
2473 i = BUF_FILEHASH(fidp);
2475 lock_ObtainWrite(&scp->rw);
2476 lock_ObtainRead(&buf_globalLock);
2477 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
2478 if (!cm_FidCmp(fidp, &bp->fid) && (bp->qFlags & CM_BUF_QREDIR)) {
2479 lock_ConvertRToW(&buf_globalLock);
2480 if (bp->qFlags & CM_BUF_QREDIR) {
2481 osi_Log4(buf_logp,"buf_ClearRDRFlag taking from file system bp 0x%p vno 0x%x foffset 0x%x:%x",
2482 bp, bp->fid.vnode, bp->offset.HighPart, bp->offset.LowPart);
2483 buf_RemoveFromRedirQueue(scp, bp);
2484 buf_ReleaseLocked(bp, TRUE);
2486 lock_ConvertWToR(&buf_globalLock);
2490 /* Confirm that there are none left */
2491 lock_ConvertRToW(&buf_globalLock);
2492 for ( bp = redirq_to_cm_buf_t(scp->redirQueueT);
2494 bp = redirq_to_cm_buf_t(scp->redirQueueT))
2496 if (bp->qFlags & CM_BUF_QREDIR) {
2497 osi_Log4(buf_logp,"buf_ClearRDRFlag taking from file system bufp 0x%p vno 0x%x foffset 0x%x:%x",
2498 bp, bp->fid.vnode, bp->offset.HighPart, bp->offset.LowPart);
2499 buf_RemoveFromRedirQueue(scp, bp);
2500 buf_ReleaseLocked(bp, TRUE);
2504 lock_ReleaseWrite(&buf_globalLock);
2505 lock_ReleaseWrite(&scp->rw);
2510 long buf_CleanDirtyBuffers(cm_scache_t *scp)
2513 afs_uint32 bcount = 0;
2514 cm_fid_t * fidp = &scp->fid;
2516 for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
2517 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
2519 lock_ObtainMutex(&bp->mx);
2520 _InterlockedAnd(&bp->cmFlags, ~CM_BUF_CMSTORING);
2521 _InterlockedAnd(&bp->flags, ~CM_BUF_DIRTY);
2522 bp->dirty_length = 0;
2523 _InterlockedOr(&bp->flags, CM_BUF_ERROR);
2524 bp->error = VNOVNODE;
2525 bp->dataVersion = CM_BUF_VERSION_BAD; /* bad */
2527 if (bp->flags & CM_BUF_WAITING) {
2528 osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
2529 osi_Wakeup((long) &bp);
2531 lock_ReleaseMutex(&bp->mx);
2540 * The following routines will not be used on a
2541 * regular basis but are very useful in a variety
2542 * of scenarios when debugging data corruption.
2545 buf_HexCheckSum(cm_buf_t * bp)
2548 static char buf[33];
2549 static char tr[16] = {'0','1','2','3','4','5','6','7','8','9','A','B','C','D','E','F'};
2551 for (i=0;i<16;i++) {
2552 k = bp->md5cksum[i];
2554 buf[i*2] = tr[k / 16];
2555 buf[i*2+1] = tr[k % 16];
2563 buf_ComputeCheckSum(cm_buf_t * bp)
2568 MD5_Update(&md5, bp->datap, cm_data.blockSize);
2569 MD5_Final(bp->md5cksum, &md5);
2571 osi_Log4(buf_logp, "CheckSum bp 0x%p md5 %s, dirty: offset %u length %u",
2572 bp, osi_LogSaveString(buf_logp, buf_HexCheckSum(bp)),
2573 bp->dirty_offset, bp->dirty_length);
2577 buf_ValidateCheckSum(cm_buf_t * bp)
2580 unsigned char tmp[16];
2583 MD5_Update(&md5, bp->datap, cm_data.blockSize);
2584 MD5_Final(tmp, &md5);
2586 if (memcmp(tmp, bp->md5cksum, 16) == 0)
2592 buf_InsertToRedirQueue(cm_scache_t *scp, cm_buf_t *bufp)
2594 lock_AssertWrite(&buf_globalLock);
2597 lock_ObtainMutex(&scp->redirMx);
2600 if (bufp->qFlags & CM_BUF_QINLRU) {
2601 _InterlockedAnd(&bufp->qFlags, ~CM_BUF_QINLRU);
2602 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_freeListp,
2603 (osi_queue_t **) &cm_data.buf_freeListEndp,
2605 buf_DecrementFreeCount();
2607 _InterlockedOr(&bufp->qFlags, CM_BUF_QREDIR);
2608 osi_QAddH( (osi_queue_t **) &cm_data.buf_redirListp,
2609 (osi_queue_t **) &cm_data.buf_redirListEndp,
2611 buf_IncrementRedirCount();
2612 bufp->redirLastAccess = time(NULL);
2614 osi_QAddH( (osi_queue_t **) &scp->redirQueueH,
2615 (osi_queue_t **) &scp->redirQueueT,
2617 scp->redirLastAccess = bufp->redirLastAccess;
2618 InterlockedIncrement(&scp->redirBufCount);
2620 lock_ReleaseMutex(&scp->redirMx);
2625 buf_RemoveFromRedirQueue(cm_scache_t *scp, cm_buf_t *bufp)
2627 lock_AssertWrite(&buf_globalLock);
2629 if (!(bufp->qFlags & CM_BUF_QREDIR))
2633 lock_ObtainMutex(&scp->redirMx);
2636 _InterlockedAnd(&bufp->qFlags, ~CM_BUF_QREDIR);
2637 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_redirListp,
2638 (osi_queue_t **) &cm_data.buf_redirListEndp,
2640 buf_DecrementRedirCount();
2643 osi_QRemoveHT( (osi_queue_t **) &scp->redirQueueH,
2644 (osi_queue_t **) &scp->redirQueueT,
2647 InterlockedDecrement(&scp->redirBufCount);
2648 lock_ReleaseMutex(&scp->redirMx);
2653 buf_MoveToHeadOfRedirQueue(cm_scache_t *scp, cm_buf_t *bufp)
2655 lock_AssertWrite(&buf_globalLock);
2656 if (!(bufp->qFlags & CM_BUF_QREDIR))
2660 lock_ObtainMutex(&scp->redirMx);
2663 osi_QRemoveHT( (osi_queue_t **) &cm_data.buf_redirListp,
2664 (osi_queue_t **) &cm_data.buf_redirListEndp,
2666 osi_QAddH( (osi_queue_t **) &cm_data.buf_redirListp,
2667 (osi_queue_t **) &cm_data.buf_redirListEndp,
2669 bufp->redirLastAccess = time(NULL);
2671 osi_QRemoveHT( (osi_queue_t **) &scp->redirQueueH,
2672 (osi_queue_t **) &scp->redirQueueT,
2674 osi_QAddH( (osi_queue_t **) &scp->redirQueueH,
2675 (osi_queue_t **) &scp->redirQueueT,
2677 scp->redirLastAccess = bufp->redirLastAccess;
2679 lock_ReleaseMutex(&scp->redirMx);