2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
10 /* Copyright (C) 1994 Cazamar Systems, Inc. */
12 #include <afsconfig.h>
13 #include <afs/param.h>
25 #include "cm_memmap.h"
28 #define TRACE_BUFFER 1
31 extern void afsi_log(char *pattern, ...);
33 /* This module implements the buffer package used by the local transaction
34 * system (cm). It is initialized by calling cm_Init, which calls buf_Init;
35 * it must be initalized before any of its main routines are called.
37 * Each buffer is hashed into a hash table by file ID and offset, and if its
38 * reference count is zero, it is also in a free list.
40 * There are two locks involved in buffer processing. The global lock
41 * buf_globalLock protects all of the global variables defined in this module,
42 * the reference counts and hash pointers in the actual cm_buf_t structures,
43 * and the LRU queue pointers in the buffer structures.
45 * The mutexes in the buffer structures protect the remaining fields in the
46 * buffers, as well the data itself.
48 * The locking hierarchy here is this:
50 * - resv multiple simul. buffers reservation
51 * - lock buffer I/O flags
52 * - lock buffer's mutex
53 * - lock buf_globalLock
57 /* global debugging log */
58 osi_log_t *buf_logp = NULL;
60 /* Global lock protecting hash tables and free lists */
61 osi_rwlock_t buf_globalLock;
63 /* ptr to head of the free list (most recently used) and the
64 * tail (the guy to remove first). We use osi_Q* functions
65 * to put stuff in buf_freeListp, and maintain the end
69 /* a pointer to a list of all buffers, just so that we can find them
70 * easily for debugging, and for the incr syncer. Locked under
74 /* defaults setup; these variables may be manually assigned into
75 * before calling cm_Init, as a way of changing these defaults.
78 /* callouts for reading and writing data, etc */
79 cm_buf_ops_t *cm_buf_opsp;
82 /* for experimental disk caching support in Win95 client */
83 cm_buf_t *buf_diskFreeListp;
84 cm_buf_t *buf_diskFreeListEndp;
85 cm_buf_t *buf_diskAllp;
86 extern int cm_diskCacheEnabled;
87 #endif /* DISKCACHE95 */
89 /* set this to 1 when we are terminating to prevent access attempts */
90 static int buf_ShutdownFlag = 0;
93 void buf_HoldLockedDbg(cm_buf_t *bp, char *file, long line)
95 void buf_HoldLocked(cm_buf_t *bp)
100 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
101 refCount = InterlockedIncrement(&bp->refCount);
102 #ifdef DEBUG_REFCOUNT
103 osi_Log2(afsd_logp,"buf_HoldLocked bp 0x%p ref %d",bp, refCount);
104 afsi_log("%s:%d buf_HoldLocked bp 0x%p, ref %d", file, line, bp, refCount);
108 /* hold a reference to an already held buffer */
109 #ifdef DEBUG_REFCOUNT
110 void buf_HoldDbg(cm_buf_t *bp, char *file, long line)
112 void buf_Hold(cm_buf_t *bp)
117 lock_ObtainRead(&buf_globalLock);
118 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
119 refCount = InterlockedIncrement(&bp->refCount);
120 #ifdef DEBUG_REFCOUNT
121 osi_Log2(afsd_logp,"buf_Hold bp 0x%p ref %d",bp, refCount);
122 afsi_log("%s:%d buf_Hold bp 0x%p, ref %d", file, line, bp, refCount);
124 lock_ReleaseRead(&buf_globalLock);
127 /* code to drop reference count while holding buf_globalLock */
128 #ifdef DEBUG_REFCOUNT
129 void buf_ReleaseLockedDbg(cm_buf_t *bp, afs_uint32 writeLocked, char *file, long line)
131 void buf_ReleaseLocked(cm_buf_t *bp, afs_uint32 writeLocked)
137 lock_AssertWrite(&buf_globalLock);
139 lock_AssertRead(&buf_globalLock);
141 /* ensure that we're in the LRU queue if our ref count is 0 */
142 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
144 refCount = InterlockedDecrement(&bp->refCount);
145 #ifdef DEBUG_REFCOUNT
146 osi_Log3(afsd_logp,"buf_ReleaseLocked %s bp 0x%p ref %d",writeLocked?"write":"read", bp, refCount);
147 afsi_log("%s:%d buf_ReleaseLocked %s bp 0x%p, ref %d", file, line, writeLocked?"write":"read", bp, refCount);
151 osi_panic("buf refcount 0",__FILE__,__LINE__);;
153 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
157 * If we are read locked there could be a race condition
158 * with buf_Find() so we must obtain a write lock and
159 * double check that the refCount is actually zero
160 * before we remove the buffer from the LRU queue.
163 lock_ConvertRToW(&buf_globalLock);
165 if (bp->refCount == 0 &&
166 !(bp->qFlags & CM_BUF_QINLRU)) {
167 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
169 /* watch for transition from empty to one element */
170 if (!cm_data.buf_freeListEndp)
171 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
172 bp->qFlags |= CM_BUF_QINLRU;
176 lock_ConvertWToR(&buf_globalLock);
180 /* release a buffer. Buffer must be referenced, but unlocked. */
181 #ifdef DEBUG_REFCOUNT
182 void buf_ReleaseDbg(cm_buf_t *bp, char *file, long line)
184 void buf_Release(cm_buf_t *bp)
189 /* ensure that we're in the LRU queue if our ref count is 0 */
190 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
192 refCount = InterlockedDecrement(&bp->refCount);
193 #ifdef DEBUG_REFCOUNT
194 osi_Log2(afsd_logp,"buf_Release bp 0x%p ref %d", bp, refCount);
195 afsi_log("%s:%d buf_ReleaseLocked bp 0x%p, ref %d", file, line, bp, refCount);
199 osi_panic("buf refcount 0",__FILE__,__LINE__);;
201 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
204 lock_ObtainWrite(&buf_globalLock);
205 if (bp->refCount == 0 &&
206 !(bp->qFlags & CM_BUF_QINLRU)) {
207 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
209 /* watch for transition from empty to one element */
210 if (!cm_data.buf_freeListEndp)
211 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
212 bp->qFlags |= CM_BUF_QINLRU;
214 lock_ReleaseWrite(&buf_globalLock);
219 buf_Sync(int quitOnShutdown)
221 cm_buf_t **bpp, *bp, *prevbp;
222 afs_uint32 wasDirty = 0;
225 /* go through all of the dirty buffers */
226 lock_ObtainRead(&buf_globalLock);
227 for (bpp = &cm_data.buf_dirtyListp, prevbp = NULL; bp = *bpp; ) {
228 if (quitOnShutdown && buf_ShutdownFlag)
231 lock_ReleaseRead(&buf_globalLock);
232 /* all dirty buffers are held when they are added to the
233 * dirty list. No need for an additional hold.
235 lock_ObtainMutex(&bp->mx);
237 if (bp->flags & CM_BUF_DIRTY && !(bp->qFlags & CM_BUF_QREDIR)) {
238 /* start cleaning the buffer; don't touch log pages since
239 * the log code counts on knowing exactly who is writing
240 * a log page at any given instant.
242 * only attempt to write the buffer if the volume might
248 volp = cm_GetVolumeByFID(&bp->fid);
249 switch (cm_GetVolumeStatus(volp, bp->fid.volume)) {
253 req.flags |= CM_REQ_NORETRY;
254 buf_CleanAsyncLocked(NULL, bp, &req, 0, &dirty);
260 /* the buffer may or may not have been dirty
261 * and if dirty may or may not have been cleaned
262 * successfully. check the dirty flag again.
264 if (!(bp->flags & CM_BUF_DIRTY)) {
265 /* remove the buffer from the dirty list */
266 lock_ObtainWrite(&buf_globalLock);
267 #ifdef DEBUG_REFCOUNT
268 if (bp->dirtyp == NULL && bp != cm_data.buf_dirtyListEndp) {
269 osi_Log1(afsd_logp,"buf_IncrSyncer bp 0x%p list corruption",bp);
270 afsi_log("buf_IncrSyncer bp 0x%p list corruption", bp);
275 bp->qFlags &= ~CM_BUF_QINDL;
276 if (cm_data.buf_dirtyListp == NULL)
277 cm_data.buf_dirtyListEndp = NULL;
278 else if (cm_data.buf_dirtyListEndp == bp)
279 cm_data.buf_dirtyListEndp = prevbp;
280 buf_ReleaseLocked(bp, TRUE);
281 lock_ConvertWToR(&buf_globalLock);
283 if (buf_ShutdownFlag) {
286 char volstr[VL_MAXNAMELEN+12]="";
289 volp = cm_GetVolumeByFID(&bp->fid);
292 if (bp->fid.volume == volp->vol[RWVOL].ID)
294 else if (bp->fid.volume == volp->vol[ROVOL].ID)
296 else if (bp->fid.volume == volp->vol[BACKVOL].ID)
300 snprintf(volstr, sizeof(volstr), "%s%s", volp->namep, ext);
302 cellp = cm_FindCellByID(bp->fid.cell, CM_FLAG_NOPROBE);
303 snprintf(volstr, sizeof(volstr), "%u", bp->fid.volume);
306 LogEvent(EVENTLOG_INFORMATION_TYPE, MSG_DIRTY_BUFFER_AT_SHUTDOWN,
307 cellp->name, volstr, bp->fid.vnode, bp->fid.unique,
308 bp->offset.QuadPart+bp->dirty_offset, bp->dirty_length);
311 /* advance the pointer so we don't loop forever */
312 lock_ObtainRead(&buf_globalLock);
316 lock_ReleaseMutex(&bp->mx);
317 } /* for loop over a bunch of buffers */
318 lock_ReleaseRead(&buf_globalLock);
323 /* incremental sync daemon. Writes all dirty buffers every 5000 ms */
324 void buf_IncrSyncer(long parm)
329 while (buf_ShutdownFlag == 0) {
332 i = SleepEx(5000, 1);
339 wasDirty = buf_Sync(1);
340 } /* whole daemon's while loop */
344 buf_ValidateBuffers(void)
346 cm_buf_t * bp, *bpf, *bpa, *bpb;
347 afs_uint64 countb = 0, countf = 0, counta = 0;
349 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
350 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
351 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
352 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
356 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
357 if (bp->magic != CM_BUF_MAGIC) {
358 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
359 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
365 if (countb > cm_data.buf_nbuffers) {
366 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
367 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
372 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
373 if (bp->magic != CM_BUF_MAGIC) {
374 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
375 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
381 if (countf > cm_data.buf_nbuffers) {
382 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
383 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
388 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
389 if (bp->magic != CM_BUF_MAGIC) {
390 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
391 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
397 if (counta > cm_data.buf_nbuffers) {
398 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
399 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
404 if (countb != countf) {
405 afsi_log("cm_ValidateBuffers failure: countb != countf");
406 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
410 if (counta != cm_data.buf_nbuffers) {
411 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
412 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
419 void buf_Shutdown(void)
421 /* disable the buf_IncrSyncer() threads */
422 buf_ShutdownFlag = 1;
424 /* then force all dirty buffers to the file servers */
428 /* initialize the buffer package; called with no locks
429 * held during the initialization phase.
431 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
433 static osi_once_t once;
442 cm_data.buf_nbuffers = nbuffers;
444 /* Have to be able to reserve a whole chunk */
445 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
446 return CM_ERROR_TOOFEWBUFS;
449 /* recall for callouts */
452 if (osi_Once(&once)) {
453 /* initialize global locks */
454 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock", LOCK_HIERARCHY_BUF_GLOBAL);
457 /* remember this for those who want to reset it */
458 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
460 /* lower hash size to a prime number */
461 cm_data.buf_hashSize = osi_PrimeLessThan((afs_uint32)(cm_data.buf_nbuffers/7 + 1));
463 /* create hash table */
464 memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
466 /* another hash table */
467 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
469 /* create buffer headers and put in free list */
470 bp = cm_data.bufHeaderBaseAddress;
471 data = cm_data.bufDataBaseAddress;
472 cm_data.buf_allp = NULL;
474 for (i=0; i<cm_data.buf_nbuffers; i++) {
475 osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress,
476 "invalid cm_buf_t address");
477 osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
478 "invalid cm_buf_t data address");
480 /* allocate and zero some storage */
481 memset(bp, 0, sizeof(cm_buf_t));
482 bp->magic = CM_BUF_MAGIC;
483 /* thread on list of all buffers */
484 bp->allp = cm_data.buf_allp;
485 cm_data.buf_allp = bp;
487 osi_QAdd((osi_queue_t **)&cm_data.buf_freeListp, &bp->q);
488 bp->qFlags |= CM_BUF_QINLRU;
489 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
491 /* grab appropriate number of bytes from aligned zone */
494 /* setup last buffer pointer */
496 cm_data.buf_freeListEndp = bp;
500 data += cm_data.buf_blockSize;
503 /* none reserved at first */
504 cm_data.buf_reservedBufs = 0;
506 /* just for safety's sake */
507 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
509 bp = cm_data.bufHeaderBaseAddress;
510 data = cm_data.bufDataBaseAddress;
512 for (i=0; i<cm_data.buf_nbuffers; i++) {
513 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
516 bp->waitRequests = 0;
517 bp->flags &= ~CM_BUF_WAITING;
523 buf_ValidateBufQueues();
527 /* init the buffer trace log */
528 buf_logp = osi_LogCreate("buffer", 1000);
529 osi_LogEnable(buf_logp);
534 /* and create the incr-syncer */
535 phandle = thrd_Create(0, 0,
536 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
539 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
540 CloseHandle(phandle);
544 buf_ValidateBufQueues();
549 /* add nbuffers to the buffer pool, if possible.
550 * Called with no locks held.
552 long buf_AddBuffers(afs_uint64 nbuffers)
554 /* The size of a virtual cache cannot be changed after it has
555 * been created. Subsequent calls to MapViewofFile() with
556 * an existing mapping object name would not allow the
557 * object to be resized. Return failure immediately.
559 * A similar problem now occurs with the persistent cache
560 * given that the memory mapped file now contains a complex
563 afsi_log("request to add %d buffers to the existing cache of size %d denied",
564 nbuffers, cm_data.buf_nbuffers);
566 return CM_ERROR_INVAL;
569 /* interface to set the number of buffers to an exact figure.
570 * Called with no locks held.
572 long buf_SetNBuffers(afs_uint64 nbuffers)
575 return CM_ERROR_INVAL;
576 if (nbuffers == cm_data.buf_nbuffers)
578 else if (nbuffers > cm_data.buf_nbuffers)
579 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
581 return CM_ERROR_INVAL;
584 /* wait for reading or writing to clear; called with write-locked
585 * buffer and unlocked scp and returns with locked buffer.
587 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
592 osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
593 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
596 /* if no IO is happening, we're done */
597 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
600 /* otherwise I/O is happening, but some other thread is waiting for
601 * the I/O already. Wait for that guy to figure out what happened,
602 * and then check again.
604 if ( bp->flags & CM_BUF_WAITING ) {
607 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
609 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
610 bp->flags |= CM_BUF_WAITING;
611 bp->waitCount = bp->waitRequests = 1;
613 osi_SleepM((LONG_PTR)bp, &bp->mx);
615 cm_UpdateServerPriority();
617 lock_ObtainMutex(&bp->mx);
618 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
620 if (bp->waitCount == 0) {
621 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
622 bp->flags &= ~CM_BUF_WAITING;
623 bp->waitRequests = 0;
627 if (scp = cm_FindSCache(&bp->fid))
631 lock_ObtainRead(&scp->rw);
632 if (scp->flags & CM_SCACHEFLAG_WAITING) {
633 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
634 osi_Wakeup((LONG_PTR)&scp->flags);
636 lock_ReleaseRead(&scp->rw);
640 /* if we get here, the IO is done, but we may have to wakeup people waiting for
641 * the I/O to complete. Do so.
643 if (bp->flags & CM_BUF_WAITING) {
644 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
645 osi_Wakeup((LONG_PTR) bp);
647 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
650 cm_ReleaseSCache(scp);
653 /* find a buffer, if any, for a particular file ID and offset. Assumes
654 * that buf_globalLock is write locked when called.
656 cm_buf_t *buf_FindLocked(struct cm_scache *scp, osi_hyper_t *offsetp)
661 i = BUF_HASH(&scp->fid, offsetp);
662 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
663 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
664 && offsetp->LowPart == bp->offset.LowPart
665 && offsetp->HighPart == bp->offset.HighPart) {
671 /* return whatever we found, if anything */
675 /* find a buffer with offset *offsetp for vnode *scp. Called
676 * with no locks held.
678 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
682 lock_ObtainRead(&buf_globalLock);
683 bp = buf_FindLocked(scp, offsetp);
684 lock_ReleaseRead(&buf_globalLock);
689 /* find a buffer, if any, for a particular file ID and offset. Assumes
690 * that buf_globalLock is write locked when called. Uses the all buffer
693 cm_buf_t *buf_FindAllLocked(struct cm_scache *scp, osi_hyper_t *offsetp, afs_uint32 flags)
698 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
699 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
700 && offsetp->LowPart == bp->offset.LowPart
701 && offsetp->HighPart == bp->offset.HighPart) {
707 for(bp = cm_data.buf_allp; bp; bp=bp->allp) {
708 if (cm_FidCmp(&scp->fid, &bp->fid) == 0) {
711 fileOffset = offsetp->QuadPart + cm_data.baseAddress;
712 if (fileOffset == bp->datap) {
719 /* return whatever we found, if anything */
723 /* find a buffer with offset *offsetp for vnode *scp. Called
724 * with no locks held. Use the all buffer list.
726 cm_buf_t *buf_FindAll(struct cm_scache *scp, osi_hyper_t *offsetp, afs_uint32 flags)
730 lock_ObtainRead(&buf_globalLock);
731 bp = buf_FindAllLocked(scp, offsetp, flags);
732 lock_ReleaseRead(&buf_globalLock);
737 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
740 * Makes sure that there's only one person writing this block
741 * at any given time, and also ensures that the log is forced sufficiently far,
742 * if this buffer contains logged data.
744 * Returns non-zero if the buffer was dirty.
746 * 'scp' may or may not be NULL. If it is not NULL, the FID for both cm_scache_t
747 * and cm_buf_t must match.
749 afs_uint32 buf_CleanAsyncLocked(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp,
750 afs_uint32 flags, afs_uint32 *pisdirty)
753 afs_uint32 isdirty = 0;
757 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
758 osi_assertx(scp == NULL || cm_FidCmp(&scp->fid, &bp->fid) == 0, "scp fid != bp fid");
761 * If the matching cm_scache_t was not provided as a parameter
762 * we must either find one or allocate a new one. It is possible
763 * that the cm_scache_t was recycled out of the cache even though
764 * a cm_buf_t with the same FID is in the cache.
767 if ((scp = cm_FindSCache(&bp->fid)) ||
768 (cm_GetSCache(&bp->fid, &scp,
769 bp->userp ? bp->userp : cm_rootUserp,
775 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
777 lock_ReleaseMutex(&bp->mx);
781 * If we didn't find a cm_scache_t object for bp->fid it means
782 * that we no longer have that FID in the cache. It does not
783 * mean that the object does not exist in the cell. That may
784 * in fact be the case but we don't know that until we attempt
785 * a FetchStatus on the FID.
787 osi_Log1(buf_logp, "buf_CleanAsyncLocked unable to start I/O - scp not found buf 0x%p", bp);
788 code = CM_ERROR_NOSUCHFILE;
790 osi_Log2(buf_logp, "buf_CleanAsyncLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
793 LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
794 code = (*cm_buf_opsp->Writep)(scp, &offset,
796 /* we might as well try to write all of the contiguous
797 * dirty buffers in one RPC
803 flags, bp->userp, reqp);
804 osi_Log3(buf_logp, "buf_CleanAsyncLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
806 lock_ObtainMutex(&bp->mx);
807 /* if the Write routine returns No Such File, clear the dirty flag
808 * because we aren't going to be able to write this data to the file
811 if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD || code == CM_ERROR_NOACCESS ||
812 code == CM_ERROR_QUOTA || code == CM_ERROR_SPACE || code == CM_ERROR_TOOBIG ||
813 code == CM_ERROR_READONLY || code == CM_ERROR_NOSUCHPATH){
814 bp->flags &= ~CM_BUF_DIRTY;
815 bp->flags |= CM_BUF_ERROR;
816 bp->dirty_offset = 0;
817 bp->dirty_length = 0;
819 bp->dataVersion = CM_BUF_VERSION_BAD;
825 /* Disk cache support */
826 /* write buffer to disk cache (synchronous for now) */
827 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
828 #endif /* DISKCACHE95 */
830 /* if we get here and retries are not permitted
831 * then we need to exit this loop regardless of
832 * whether or not we were able to clear the dirty bit
834 if (reqp->flags & CM_REQ_NORETRY)
837 /* Ditto if the hardDeadTimeout or idleTimeout was reached */
838 if (code == CM_ERROR_TIMEDOUT || code == CM_ERROR_ALLDOWN ||
839 code == CM_ERROR_ALLBUSY || code == CM_ERROR_ALLOFFLINE ||
840 code == CM_ERROR_CLOCKSKEW) {
846 cm_ReleaseSCache(scp);
848 /* if someone was waiting for the I/O that just completed or failed,
851 if (bp->flags & CM_BUF_WAITING) {
852 /* turn off flags and wakeup users */
853 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
854 osi_Wakeup((LONG_PTR) bp);
863 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
864 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
865 * The buffer must already be clean, and no I/O should be happening to it.
867 void buf_Recycle(cm_buf_t *bp)
872 cm_buf_t *prevBp, *nextBp;
874 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
876 /* if we get here, we know that the buffer still has a 0 ref count,
877 * and that it is clean and has no currently pending I/O. This is
878 * the dude to return.
879 * Remember that as long as the ref count is 0, we know that we won't
880 * have any lock conflicts, so we can grab the buffer lock out of
881 * order in the locking hierarchy.
883 osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
884 bp, bp->offset.HighPart, bp->offset.LowPart);
886 osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
887 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
888 "incorrect cm_buf_t flags");
889 lock_AssertWrite(&buf_globalLock);
891 if (bp->qFlags & CM_BUF_QINHASH) {
892 /* Remove from hash */
894 i = BUF_HASH(&bp->fid, &bp->offset);
895 lbpp = &(cm_data.buf_scacheHashTablepp[i]);
896 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
901 /* we better find it */
902 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
904 *lbpp = bp->hashp; /* hash out */
907 /* Remove from file hash */
909 i = BUF_FILEHASH(&bp->fid);
910 prevBp = bp->fileHashBackp;
911 bp->fileHashBackp = NULL;
912 nextBp = bp->fileHashp;
913 bp->fileHashp = NULL;
915 prevBp->fileHashp = nextBp;
917 cm_data.buf_fileHashTablepp[i] = nextBp;
919 nextBp->fileHashBackp = prevBp;
921 bp->qFlags &= ~CM_BUF_QINHASH;
924 /* make the fid unrecognizable */
925 memset(&bp->fid, 0, sizeof(cm_fid_t));
927 /* clean up junk flags */
928 bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
929 bp->dataVersion = CM_BUF_VERSION_BAD; /* unknown so far */
932 /* recycle a buffer, removing it from the free list, hashing in its new identity
933 * and returning it write-locked so that no one can use it. Called without
934 * any locks held, and can return an error if it loses the race condition and
935 * finds that someone else created the desired buffer.
937 * If success is returned, the buffer is returned write-locked.
939 * May be called with null scp and offsetp, if we're just trying to reclaim some
940 * space from the buffer pool. In that case, the buffer will be returned
941 * without being hashed into the hash table.
943 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
945 cm_buf_t *bp; /* buffer we're dealing with */
946 cm_buf_t *nextBp; /* next buffer in file hash chain */
947 afs_uint32 i; /* temp */
950 buf_ValidateBufQueues();
955 lock_ObtainRead(&scp->bufCreateLock);
956 lock_ObtainWrite(&buf_globalLock);
957 /* check to see if we lost the race */
959 if (bp = buf_FindLocked(scp, offsetp)) {
960 /* Do not call buf_ReleaseLocked() because we
961 * do not want to allow the buffer to be added
964 afs_int32 refCount = InterlockedDecrement(&bp->refCount);
965 #ifdef DEBUG_REFCOUNT
966 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, refCount);
967 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, refCount);
969 lock_ReleaseWrite(&buf_globalLock);
970 lock_ReleaseRead(&scp->bufCreateLock);
971 return CM_BUF_EXISTS;
975 /* does this fix the problem below? it's a simple solution. */
976 if (!cm_data.buf_freeListEndp)
978 lock_ReleaseWrite(&buf_globalLock);
979 lock_ReleaseRead(&scp->bufCreateLock);
980 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
985 /* for debugging, assert free list isn't empty, although we
986 * really should try waiting for a running tranasction to finish
987 * instead of this; or better, we should have a transaction
988 * throttler prevent us from entering this situation.
990 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
992 /* look at all buffers in free list, some of which may temp.
993 * have high refcounts and which then should be skipped,
994 * starting cleaning I/O for those which are dirty. If we find
995 * a clean buffer, we rehash it, lock it and return it.
997 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
998 /* check to see if it really has zero ref count. This
999 * code can bump refcounts, at least, so it may not be
1002 if (bp->refCount > 0)
1005 /* we don't have to lock buffer itself, since the ref
1006 * count is 0 and we know it will stay zero as long as
1007 * we hold the global lock.
1010 /* Don't recycle a buffer held by the redirector. */
1011 if (bp->qFlags & CM_BUF_QREDIR)
1014 /* don't recycle someone in our own chunk */
1015 if (!cm_FidCmp(&bp->fid, &scp->fid)
1016 && (bp->offset.LowPart & (-cm_chunkSize))
1017 == (offsetp->LowPart & (-cm_chunkSize)))
1020 /* if this page is being filled (!) or cleaned, see if
1021 * the I/O has completed. If not, skip it, otherwise
1022 * do the final processing for the I/O.
1024 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
1025 /* probably shouldn't do this much work while
1026 * holding the big lock? Watch for contention
1032 if (bp->flags & CM_BUF_DIRTY) {
1033 /* if the buffer is dirty, start cleaning it and
1034 * move on to the next buffer. We do this with
1035 * just the lock required to minimize contention
1039 lock_ReleaseWrite(&buf_globalLock);
1040 lock_ReleaseRead(&scp->bufCreateLock);
1042 /* grab required lock and clean; this only
1043 * starts the I/O. By the time we're back,
1044 * it'll still be marked dirty, but it will also
1045 * have the WRITING flag set, so we won't get
1048 if (cm_FidCmp(&scp->fid, &bp->fid) == 0)
1049 buf_CleanAsync(scp, bp, reqp, 0, NULL);
1051 buf_CleanAsync(NULL, bp, reqp, 0, NULL);
1053 /* now put it back and go around again */
1058 /* if we get here, we know that the buffer still has a 0
1059 * ref count, and that it is clean and has no currently
1060 * pending I/O. This is the dude to return.
1061 * Remember that as long as the ref count is 0, we know
1062 * that we won't have any lock conflicts, so we can grab
1063 * the buffer lock out of order in the locking hierarchy.
1067 /* now hash in as our new buffer, and give it the
1068 * appropriate label, if requested.
1071 bp->qFlags |= CM_BUF_QINHASH;
1076 bp->offset = *offsetp;
1077 i = BUF_HASH(&scp->fid, offsetp);
1078 bp->hashp = cm_data.buf_scacheHashTablepp[i];
1079 cm_data.buf_scacheHashTablepp[i] = bp;
1080 i = BUF_FILEHASH(&scp->fid);
1081 nextBp = cm_data.buf_fileHashTablepp[i];
1082 bp->fileHashp = nextBp;
1083 bp->fileHashBackp = NULL;
1085 nextBp->fileHashBackp = bp;
1086 cm_data.buf_fileHashTablepp[i] = bp;
1089 /* we should move it from the lru queue. It better still be there,
1090 * since we've held the global (big) lock since we found it there.
1092 osi_assertx(bp->qFlags & CM_BUF_QINLRU,
1093 "buf_GetNewLocked: LRU screwup");
1095 if (cm_data.buf_freeListEndp == bp) {
1096 /* we're the last guy in this queue, so maintain it */
1097 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1099 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1100 bp->qFlags &= ~CM_BUF_QINLRU;
1102 /* prepare to return it. Give it a refcount */
1104 #ifdef DEBUG_REFCOUNT
1105 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, 1);
1106 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, 1);
1108 /* grab the mutex so that people don't use it
1109 * before the caller fills it with data. Again, no one
1110 * should have been able to get to this dude to lock it.
1112 if (!lock_TryMutex(&bp->mx)) {
1113 osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked. refCount %d should be 0",
1115 osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
1118 lock_ReleaseWrite(&buf_globalLock);
1119 lock_ReleaseRead(&scp->bufCreateLock);
1124 buf_ValidateBufQueues();
1125 #endif /* TESTING */
1127 } /* for all buffers in lru queue */
1128 lock_ReleaseWrite(&buf_globalLock);
1129 lock_ReleaseRead(&scp->bufCreateLock);
1130 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List has no buffers with a zero refcount - sleeping 100ms");
1131 Sleep(100); /* give some time for a buffer to be freed */
1132 } /* while loop over everything */
1137 * get a page, returning it held but unlocked. the page may or may not
1138 * contain valid data.
1140 * The scp must be unlocked when passed in unlocked.
1142 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_req_t *reqp, cm_buf_t **bufpp)
1146 osi_hyper_t pageOffset;
1147 unsigned long tcount;
1151 cm_diskcache_t *dcp;
1152 #endif /* DISKCACHE95 */
1155 pageOffset.HighPart = offsetp->HighPart;
1156 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
1160 buf_ValidateBufQueues();
1161 #endif /* TESTING */
1163 bp = buf_Find(scp, &pageOffset);
1165 /* lock it and break out */
1166 lock_ObtainMutex(&bp->mx);
1169 /* touch disk chunk to update LRU info */
1170 diskcache_Touch(bp->dcp);
1171 #endif /* DISKCACHE95 */
1175 /* otherwise, we have to create a page */
1176 code = buf_GetNewLocked(scp, &pageOffset, reqp, &bp);
1179 /* the requested buffer was created */
1184 * the requested buffer existed by the time the
1185 * scp->bufCreateLock and buf_globalLock could be obtained.
1186 * loop again and permit buf_Find() to obtain a reference.
1191 * the requested buffer could not be created.
1192 * return the error to the caller.
1195 buf_ValidateBufQueues();
1196 #endif /* TESTING */
1199 } /* big while loop */
1201 /* if we get here, we have a locked buffer that may have just been
1202 * created, in which case it needs to be filled with data.
1205 /* load the page; freshly created pages should be idle */
1206 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
1209 * start the I/O; may drop lock. as of this writing, the only
1210 * implementation of Readp is cm_BufRead() which simply sets
1211 * tcount to 0 and returns success.
1213 bp->flags |= CM_BUF_READING;
1214 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
1217 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
1218 bp->dcp = dcp; /* pointer to disk cache struct. */
1219 #endif /* DISKCACHE95 */
1222 /* failure or queued */
1223 if (code != ERROR_IO_PENDING) {
1225 bp->flags |= CM_BUF_ERROR;
1226 bp->flags &= ~CM_BUF_READING;
1227 if (bp->flags & CM_BUF_WAITING) {
1228 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1229 osi_Wakeup((LONG_PTR) bp);
1231 lock_ReleaseMutex(&bp->mx);
1234 buf_ValidateBufQueues();
1235 #endif /* TESTING */
1240 * otherwise, I/O completed instantly and we're done, except
1241 * for padding the xfr out with 0s and checking for EOF
1243 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1244 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1246 bp->flags |= CM_BUF_EOF;
1248 bp->flags &= ~CM_BUF_READING;
1249 if (bp->flags & CM_BUF_WAITING) {
1250 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1251 osi_Wakeup((LONG_PTR) bp);
1256 /* wait for reads, either that which we started above, or that someone
1257 * else started. We don't care if we return a buffer being cleaned.
1259 if (bp->flags & CM_BUF_READING)
1260 buf_WaitIO(scp, bp);
1262 /* once it has been read once, we can unlock it and return it, still
1263 * with its refcount held.
1265 lock_ReleaseMutex(&bp->mx);
1268 /* now remove from queue; will be put in at the head (farthest from
1269 * being recycled) when we're done in buf_Release.
1271 lock_ObtainWrite(&buf_globalLock);
1272 if (bp->qFlags & CM_BUF_QINLRU) {
1273 if (cm_data.buf_freeListEndp == bp)
1274 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1275 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1276 bp->qFlags &= ~CM_BUF_QINLRU;
1278 lock_ReleaseWrite(&buf_globalLock);
1280 osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1281 bp, scp, offsetp->HighPart, offsetp->LowPart);
1283 buf_ValidateBufQueues();
1284 #endif /* TESTING */
1288 /* count # of elements in the free list;
1289 * we don't bother doing the proper locking for accessing dataVersion or flags
1290 * since it is a pain, and this is really just an advisory call. If you need
1291 * to do better at some point, rewrite this function.
1293 long buf_CountFreeList(void)
1299 lock_ObtainRead(&buf_globalLock);
1300 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1301 /* if the buffer doesn't have an identity, or if the buffer
1302 * has been invalidate (by having its DV stomped upon), then
1303 * count it as free, since it isn't really being utilized.
1305 if (!(bufp->qFlags & CM_BUF_QINHASH) || bufp->dataVersion == CM_BUF_VERSION_BAD)
1308 lock_ReleaseRead(&buf_globalLock);
1312 /* clean a buffer synchronously */
1313 afs_uint32 buf_CleanAsync(cm_scache_t *scp, cm_buf_t *bp, cm_req_t *reqp, afs_uint32 flags, afs_uint32 *pisdirty)
1316 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1317 osi_assertx(!(flags & CM_BUF_WRITE_SCP_LOCKED), "scp->rw must not be held when calling buf_CleanAsync");
1319 lock_ObtainMutex(&bp->mx);
1320 code = buf_CleanAsyncLocked(scp, bp, reqp, flags, pisdirty);
1321 lock_ReleaseMutex(&bp->mx);
1326 /* wait for a buffer's cleaning to finish */
1327 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp, afs_uint32 locked)
1329 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1332 lock_ObtainMutex(&bp->mx);
1333 if (bp->flags & CM_BUF_WRITING) {
1334 buf_WaitIO(scp, bp);
1337 lock_ReleaseMutex(&bp->mx);
1340 /* set the dirty flag on a buffer, and set associated write-ahead log,
1341 * if there is one. Allow one to be added to a buffer, but not changed.
1343 * The buffer must be locked before calling this routine.
1345 void buf_SetDirty(cm_buf_t *bp, afs_uint32 offset, afs_uint32 length, cm_user_t *userp)
1347 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1348 osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
1350 if (bp->flags & CM_BUF_DIRTY) {
1352 osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
1354 if (bp->dirty_offset <= offset) {
1355 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1356 /* dirty_length remains the same */
1358 bp->dirty_length = offset + length - bp->dirty_offset;
1360 } else /* bp->dirty_offset > offset */ {
1361 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1362 bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
1364 bp->dirty_length = length;
1366 bp->dirty_offset = offset;
1369 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1372 bp->flags |= CM_BUF_DIRTY;
1374 /* and turn off EOF flag, since it has associated data now */
1375 bp->flags &= ~CM_BUF_EOF;
1377 bp->dirty_offset = offset;
1378 bp->dirty_length = length;
1380 /* and add to the dirty list.
1381 * we obtain a hold on the buffer for as long as it remains
1382 * in the list. buffers are only removed from the list by
1383 * the buf_IncrSyncer function regardless of when else the
1384 * dirty flag might be cleared.
1386 * This should never happen but just in case there is a bug
1387 * elsewhere, never add to the dirty list if the buffer is
1390 lock_ObtainWrite(&buf_globalLock);
1391 if (!(bp->qFlags & CM_BUF_QINDL)) {
1393 if (!cm_data.buf_dirtyListp) {
1394 cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
1396 cm_data.buf_dirtyListEndp->dirtyp = bp;
1397 cm_data.buf_dirtyListEndp = bp;
1400 bp->qFlags |= CM_BUF_QINDL;
1402 lock_ReleaseWrite(&buf_globalLock);
1405 /* and record the last writer */
1406 if (bp->userp != userp) {
1409 cm_ReleaseUser(bp->userp);
1414 /* clean all buffers, reset log pointers and invalidate all buffers.
1415 * Called with no locks held, and returns with same.
1417 * This function is guaranteed to clean and remove the log ptr of all the
1418 * buffers that were dirty or had non-zero log ptrs before the call was
1419 * made. That's sufficient to clean up any garbage left around by recovery,
1420 * which is all we're counting on this for; there may be newly created buffers
1421 * added while we're running, but that should be OK.
1423 * In an environment where there are no transactions (artificially imposed, for
1424 * example, when switching the database to raw mode), this function is used to
1425 * make sure that all updates have been written to the disk. In that case, we don't
1426 * really require that we forget the log association between pages and logs, but
1427 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1428 * have to worry about invalidating data in the buffers.
1430 * This function is used at the end of recovery as paranoia to get the recovered
1431 * database out to disk. It removes all references to the recovery log and cleans
1434 long buf_CleanAndReset(void)
1440 lock_ObtainRead(&buf_globalLock);
1441 for(i=0; i<cm_data.buf_hashSize; i++) {
1442 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
1443 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1445 lock_ReleaseRead(&buf_globalLock);
1447 /* now no locks are held; clean buffer and go on */
1449 req.flags |= CM_REQ_NORETRY;
1451 buf_CleanAsync(NULL, bp, &req, 0, NULL);
1452 buf_CleanWait(NULL, bp, FALSE);
1454 /* relock and release buffer */
1455 lock_ObtainRead(&buf_globalLock);
1456 buf_ReleaseLocked(bp, FALSE);
1458 } /* over one bucket */
1459 } /* for loop over all hash buckets */
1462 lock_ReleaseRead(&buf_globalLock);
1465 buf_ValidateBufQueues();
1466 #endif /* TESTING */
1468 /* and we're done */
1472 /* called without global lock being held, reserves buffers for callers
1473 * that need more than one held (not locked) at once.
1475 void buf_ReserveBuffers(afs_uint64 nbuffers)
1477 lock_ObtainWrite(&buf_globalLock);
1479 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1480 cm_data.buf_reserveWaiting = 1;
1481 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1482 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1483 lock_ObtainWrite(&buf_globalLock);
1486 cm_data.buf_reservedBufs += nbuffers;
1490 lock_ReleaseWrite(&buf_globalLock);
1493 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1497 lock_ObtainWrite(&buf_globalLock);
1498 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1502 cm_data.buf_reservedBufs += nbuffers;
1505 lock_ReleaseWrite(&buf_globalLock);
1509 /* called without global lock held, releases reservation held by
1510 * buf_ReserveBuffers.
1512 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1514 lock_ObtainWrite(&buf_globalLock);
1515 cm_data.buf_reservedBufs -= nbuffers;
1516 if (cm_data.buf_reserveWaiting) {
1517 cm_data.buf_reserveWaiting = 0;
1518 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1520 lock_ReleaseWrite(&buf_globalLock);
1523 /* truncate the buffers past sizep, zeroing out the page, if we don't
1524 * end on a page boundary.
1526 * Requires cm_bufCreateLock to be write locked.
1528 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1532 cm_buf_t *nbufp; /* next buffer, if didRelease */
1538 /* assert that cm_bufCreateLock is held in write mode */
1539 lock_AssertWrite(&scp->bufCreateLock);
1541 i = BUF_FILEHASH(&scp->fid);
1543 lock_ObtainRead(&buf_globalLock);
1544 bufp = cm_data.buf_fileHashTablepp[i];
1546 lock_ReleaseRead(&buf_globalLock);
1550 buf_HoldLocked(bufp);
1551 lock_ReleaseRead(&buf_globalLock);
1553 lock_ObtainMutex(&bufp->mx);
1555 bufEnd.HighPart = 0;
1556 bufEnd.LowPart = cm_data.buf_blockSize;
1557 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1559 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1560 LargeIntegerLessThan(*sizep, bufEnd)) {
1561 buf_WaitIO(scp, bufp);
1563 lock_ObtainWrite(&scp->rw);
1565 /* make sure we have a callback (so we have the right value for
1566 * the length), and wait for it to be safe to do a truncate.
1568 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1569 CM_SCACHESYNC_NEEDCALLBACK
1570 | CM_SCACHESYNC_GETSTATUS
1571 | CM_SCACHESYNC_SETSIZE
1572 | CM_SCACHESYNC_BUFLOCKED);
1575 /* if we succeeded in our locking, and this applies to the right
1576 * file, and the truncate request overlaps the buffer either
1577 * totally or partially, then do something.
1579 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1580 && LargeIntegerLessThan(*sizep, bufEnd)) {
1583 /* destroy the buffer, turning off its dirty bit, if
1584 * we're truncating the whole buffer. Otherwise, set
1585 * the dirty bit, and clear out the tail of the buffer
1586 * if we just overlap some.
1588 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1589 /* truncating the entire page */
1590 bufp->flags &= ~CM_BUF_DIRTY;
1591 bufp->dirty_offset = 0;
1592 bufp->dirty_length = 0;
1593 bufp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1594 bufp->dirtyCounter++;
1597 /* don't set dirty, since dirty implies
1598 * currently up-to-date. Don't need to do this,
1599 * since we'll update the length anyway.
1601 * Zero out remainder of the page, in case we
1602 * seek and write past EOF, and make this data
1605 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1606 osi_assertx(bufferPos != 0, "non-zero bufferPos");
1607 memset(bufp->datap + bufferPos, 0,
1608 cm_data.buf_blockSize - bufferPos);
1612 cm_SyncOpDone( scp, bufp,
1613 CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
1614 | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
1616 lock_ReleaseWrite(&scp->rw);
1617 lock_ReleaseMutex(&bufp->mx);
1620 nbufp = bufp->fileHashp;
1624 /* This forces the loop to end and the error code
1625 * to be returned. */
1633 buf_ValidateBufQueues();
1634 #endif /* TESTING */
1640 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1643 cm_buf_t *bp; /* buffer we're hacking on */
1647 afs_uint32 stable = 0;
1649 i = BUF_FILEHASH(&scp->fid);
1652 lock_ObtainRead(&buf_globalLock);
1653 bp = cm_data.buf_fileHashTablepp[i];
1656 lock_ReleaseRead(&buf_globalLock);
1658 for (; bp; bp = nbp) {
1659 didRelease = 0; /* haven't released this buffer yet */
1661 /* clean buffer synchronously */
1662 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1664 if (code == 0 && !stable && (bp->flags & CM_BUF_DIRTY)) {
1666 * we must stabilize the object to ensure that buffer
1667 * changes cannot occur while the flush is performed.
1668 * However, we do not want to Stabilize if we do not
1669 * need to because Stabilize obtains a callback.
1671 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1672 stable = (code == 0);
1675 if (code == CM_ERROR_BADFD) {
1676 /* if the scp's FID is bad its because we received VNOVNODE
1677 * when attempting to FetchStatus before the write. This
1678 * page therefore contains data that can no longer be stored.
1680 lock_ObtainMutex(&bp->mx);
1681 bp->flags &= ~CM_BUF_DIRTY;
1682 bp->flags |= CM_BUF_ERROR;
1683 bp->error = CM_ERROR_BADFD;
1684 bp->dirty_offset = 0;
1685 bp->dirty_length = 0;
1686 bp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1688 lock_ReleaseMutex(&bp->mx);
1689 } else if (!(scp->flags & CM_SCACHEFLAG_RO)) {
1694 lock_ObtainMutex(&bp->mx);
1696 /* start cleaning the buffer, and wait for it to finish */
1697 buf_CleanAsyncLocked(scp, bp, reqp, 0, NULL);
1698 buf_WaitIO(scp, bp);
1700 lock_ReleaseMutex(&bp->mx);
1703 /* actually, we only know that buffer is clean if ref
1704 * count is 1, since we don't have buffer itself locked.
1706 if (!(bp->flags & CM_BUF_DIRTY)) {
1707 lock_ObtainWrite(&buf_globalLock);
1708 if (bp->refCount == 1) { /* bp is held above */
1709 nbp = bp->fileHashp;
1711 buf_HoldLocked(nbp);
1712 buf_ReleaseLocked(bp, TRUE);
1716 lock_ReleaseWrite(&buf_globalLock);
1722 lock_ObtainRead(&buf_globalLock);
1723 nbp = bp->fileHashp;
1725 buf_HoldLocked(nbp);
1726 buf_ReleaseLocked(bp, FALSE);
1727 lock_ReleaseRead(&buf_globalLock);
1729 } /* for loop over a bunch of buffers */
1732 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1735 buf_ValidateBufQueues();
1736 #endif /* TESTING */
1742 /* Must be called with scp->rw held */
1743 long buf_ForceDataVersion(cm_scache_t * scp, afs_uint64 fromVersion, afs_uint64 toVersion)
1749 lock_AssertAny(&scp->rw);
1751 i = BUF_FILEHASH(&scp->fid);
1753 lock_ObtainRead(&buf_globalLock);
1755 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
1756 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1757 if (bp->dataVersion == fromVersion) {
1758 bp->dataVersion = toVersion;
1763 lock_ReleaseRead(&buf_globalLock);
1771 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1775 cm_buf_t *bp; /* buffer we're hacking on */
1776 cm_buf_t *nbp; /* next one */
1779 i = BUF_FILEHASH(&scp->fid);
1781 lock_ObtainRead(&buf_globalLock);
1782 bp = cm_data.buf_fileHashTablepp[i];
1785 lock_ReleaseRead(&buf_globalLock);
1786 for (; bp; bp = nbp) {
1787 /* clean buffer synchronously */
1788 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1789 lock_ObtainMutex(&bp->mx);
1790 if (bp->flags & CM_BUF_DIRTY) {
1791 if (userp && userp != bp->userp) {
1794 cm_ReleaseUser(bp->userp);
1799 case CM_ERROR_NOSUCHFILE:
1800 case CM_ERROR_BADFD:
1801 case CM_ERROR_NOACCESS:
1802 case CM_ERROR_QUOTA:
1803 case CM_ERROR_SPACE:
1804 case CM_ERROR_TOOBIG:
1805 case CM_ERROR_READONLY:
1806 case CM_ERROR_NOSUCHPATH:
1808 * Apply the previous fatal error to this buffer.
1809 * Do not waste the time attempting to store to
1810 * the file server when we know it will fail.
1812 bp->flags &= ~CM_BUF_DIRTY;
1813 bp->flags |= CM_BUF_ERROR;
1814 bp->dirty_offset = 0;
1815 bp->dirty_length = 0;
1817 bp->dataVersion = CM_BUF_VERSION_BAD;
1820 case CM_ERROR_TIMEDOUT:
1821 case CM_ERROR_ALLDOWN:
1822 case CM_ERROR_ALLBUSY:
1823 case CM_ERROR_ALLOFFLINE:
1824 case CM_ERROR_CLOCKSKEW:
1825 /* do not mark the buffer in error state but do
1826 * not attempt to complete the rest either.
1830 code = buf_CleanAsyncLocked(scp, bp, reqp, 0, &wasDirty);
1831 if (bp->flags & CM_BUF_ERROR) {
1837 buf_CleanWait(scp, bp, TRUE);
1839 lock_ReleaseMutex(&bp->mx);
1842 lock_ObtainRead(&buf_globalLock);
1843 nbp = bp->fileHashp;
1845 buf_HoldLocked(nbp);
1846 buf_ReleaseLocked(bp, FALSE);
1847 lock_ReleaseRead(&buf_globalLock);
1848 } /* for loop over a bunch of buffers */
1851 buf_ValidateBufQueues();
1852 #endif /* TESTING */
1860 buf_ValidateBufQueues(void)
1862 cm_buf_t * bp, *bpb, *bpf, *bpa;
1863 afs_uint32 countf=0, countb=0, counta=0;
1865 lock_ObtainRead(&buf_globalLock);
1866 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1867 if (bp->magic != CM_BUF_MAGIC)
1868 osi_panic("buf magic error",__FILE__,__LINE__);
1873 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1874 if (bp->magic != CM_BUF_MAGIC)
1875 osi_panic("buf magic error",__FILE__,__LINE__);
1880 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1881 if (bp->magic != CM_BUF_MAGIC)
1882 osi_panic("buf magic error",__FILE__,__LINE__);
1886 lock_ReleaseRead(&buf_globalLock);
1888 if (countb != countf)
1889 osi_panic("buf magic error",__FILE__,__LINE__);
1891 if (counta != cm_data.buf_nbuffers)
1892 osi_panic("buf magic error",__FILE__,__LINE__);
1894 #endif /* TESTING */
1896 /* dump the contents of the buf_scacheHashTablepp. */
1897 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1904 if (cm_data.buf_scacheHashTablepp == NULL)
1908 lock_ObtainRead(&buf_globalLock);
1910 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n",
1911 cookie, cm_data.buf_hashSize);
1912 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1914 for (i = 0; i < cm_data.buf_hashSize; i++)
1916 for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp)
1918 StringCbPrintfA(output, sizeof(output),
1919 "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
1920 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1921 "flags=0x%x, qFlags=0x%x cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1922 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1923 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1924 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1925 bp->cmFlags, bp->error, bp->refCount);
1926 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1930 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
1931 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1933 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
1934 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1935 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1936 StringCbPrintfA(output, sizeof(output),
1937 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1938 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1939 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1940 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1941 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1942 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1943 bp->cmFlags, bp->error, bp->refCount);
1944 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1946 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
1947 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1949 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListp\r\n", cookie);
1950 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1951 for(bp = cm_data.buf_dirtyListp; bp; bp=bp->dirtyp) {
1952 StringCbPrintfA(output, sizeof(output),
1953 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1954 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1955 "flags=0x%x, qFlags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1956 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1957 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1958 bp->offset.LowPart, bp->dataVersion, bp->flags, bp->qFlags,
1959 bp->cmFlags, bp->error, bp->refCount);
1960 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1962 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListp.\r\n", cookie);
1963 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1966 lock_ReleaseRead(&buf_globalLock);
1970 void buf_ForceTrace(BOOL flush)
1979 len = GetTempPath(sizeof(buf)-10, buf);
1980 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1981 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1982 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1983 if (handle == INVALID_HANDLE_VALUE) {
1984 osi_panic("Cannot create log file", __FILE__, __LINE__);
1986 osi_LogPrint(buf_logp, handle);
1988 FlushFileBuffers(handle);
1989 CloseHandle(handle);
1992 long buf_DirtyBuffersExist(cm_fid_t *fidp)
1995 afs_uint32 bcount = 0;
1998 i = BUF_FILEHASH(fidp);
2000 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->fileHashp, bcount++) {
2001 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY))
2008 long buf_CleanDirtyBuffers(cm_scache_t *scp)
2011 afs_uint32 bcount = 0;
2012 cm_fid_t * fidp = &scp->fid;
2014 for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
2015 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
2017 lock_ObtainMutex(&bp->mx);
2018 bp->cmFlags &= ~CM_BUF_CMSTORING;
2019 bp->flags &= ~CM_BUF_DIRTY;
2020 bp->dirty_offset = 0;
2021 bp->dirty_length = 0;
2022 bp->flags |= CM_BUF_ERROR;
2023 bp->error = VNOVNODE;
2024 bp->dataVersion = CM_BUF_VERSION_BAD; /* bad */
2026 if (bp->flags & CM_BUF_WAITING) {
2027 osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
2028 osi_Wakeup((long) &bp);
2030 lock_ReleaseMutex(&bp->mx);