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 <afs/param.h>
23 #include "cm_memmap.h"
26 #define TRACE_BUFFER 1
29 extern void afsi_log(char *pattern, ...);
31 /* This module implements the buffer package used by the local transaction
32 * system (cm). It is initialized by calling cm_Init, which calls buf_Init;
33 * it must be initalized before any of its main routines are called.
35 * Each buffer is hashed into a hash table by file ID and offset, and if its
36 * reference count is zero, it is also in a free list.
38 * There are two locks involved in buffer processing. The global lock
39 * buf_globalLock protects all of the global variables defined in this module,
40 * the reference counts and hash pointers in the actual cm_buf_t structures,
41 * and the LRU queue pointers in the buffer structures.
43 * The mutexes in the buffer structures protect the remaining fields in the
44 * buffers, as well the data itself.
46 * The locking hierarchy here is this:
48 * - resv multiple simul. buffers reservation
49 * - lock buffer I/O flags
50 * - lock buffer's mutex
51 * - lock buf_globalLock
55 /* global debugging log */
56 osi_log_t *buf_logp = NULL;
58 /* Global lock protecting hash tables and free lists */
59 osi_rwlock_t buf_globalLock;
61 /* ptr to head of the free list (most recently used) and the
62 * tail (the guy to remove first). We use osi_Q* functions
63 * to put stuff in buf_freeListp, and maintain the end
67 /* a pointer to a list of all buffers, just so that we can find them
68 * easily for debugging, and for the incr syncer. Locked under
72 /* defaults setup; these variables may be manually assigned into
73 * before calling cm_Init, as a way of changing these defaults.
76 /* callouts for reading and writing data, etc */
77 cm_buf_ops_t *cm_buf_opsp;
80 /* for experimental disk caching support in Win95 client */
81 cm_buf_t *buf_diskFreeListp;
82 cm_buf_t *buf_diskFreeListEndp;
83 cm_buf_t *buf_diskAllp;
84 extern int cm_diskCacheEnabled;
85 #endif /* DISKCACHE95 */
87 /* set this to 1 when we are terminating to prevent access attempts */
88 static int buf_ShutdownFlag = 0;
91 void buf_HoldLockedDbg(cm_buf_t *bp, char *file, long line)
93 void buf_HoldLocked(cm_buf_t *bp)
98 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
99 refCount = InterlockedIncrement(&bp->refCount);
100 #ifdef DEBUG_REFCOUNT
101 osi_Log2(afsd_logp,"buf_HoldLocked bp 0x%p ref %d",bp, refCount);
102 afsi_log("%s:%d buf_HoldLocked bp 0x%p, ref %d", file, line, bp, refCount);
106 /* hold a reference to an already held buffer */
107 #ifdef DEBUG_REFCOUNT
108 void buf_HoldDbg(cm_buf_t *bp, char *file, long line)
110 void buf_Hold(cm_buf_t *bp)
115 lock_ObtainRead(&buf_globalLock);
116 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
117 refCount = InterlockedIncrement(&bp->refCount);
118 #ifdef DEBUG_REFCOUNT
119 osi_Log2(afsd_logp,"buf_Hold bp 0x%p ref %d",bp, refCount);
120 afsi_log("%s:%d buf_Hold bp 0x%p, ref %d", file, line, bp, refCount);
122 lock_ReleaseRead(&buf_globalLock);
125 /* code to drop reference count while holding buf_globalLock */
126 #ifdef DEBUG_REFCOUNT
127 void buf_ReleaseLockedDbg(cm_buf_t *bp, afs_uint32 writeLocked, char *file, long line)
129 void buf_ReleaseLocked(cm_buf_t *bp, afs_uint32 writeLocked)
135 lock_AssertWrite(&buf_globalLock);
137 lock_AssertRead(&buf_globalLock);
139 /* ensure that we're in the LRU queue if our ref count is 0 */
140 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
142 refCount = InterlockedDecrement(&bp->refCount);
143 #ifdef DEBUG_REFCOUNT
144 osi_Log3(afsd_logp,"buf_ReleaseLocked %s bp 0x%p ref %d",writeLocked?"write":"read", bp, refCount);
145 afsi_log("%s:%d buf_ReleaseLocked %s bp 0x%p, ref %d", file, line, writeLocked?"write":"read", bp, refCount);
149 osi_panic("buf refcount 0",__FILE__,__LINE__);;
151 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
155 * If we are read locked there could be a race condition
156 * with buf_Find() so we must obtain a write lock and
157 * double check that the refCount is actually zero
158 * before we remove the buffer from the LRU queue.
161 lock_ConvertRToW(&buf_globalLock);
163 if (bp->refCount == 0 &&
164 !(bp->flags & CM_BUF_INLRU)) {
165 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
167 /* watch for transition from empty to one element */
168 if (!cm_data.buf_freeListEndp)
169 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
170 bp->flags |= CM_BUF_INLRU;
174 lock_ConvertWToR(&buf_globalLock);
178 /* release a buffer. Buffer must be referenced, but unlocked. */
179 #ifdef DEBUG_REFCOUNT
180 void buf_ReleaseDbg(cm_buf_t *bp, char *file, long line)
182 void buf_Release(cm_buf_t *bp)
187 /* ensure that we're in the LRU queue if our ref count is 0 */
188 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
190 refCount = InterlockedDecrement(&bp->refCount);
191 #ifdef DEBUG_REFCOUNT
192 osi_Log2(afsd_logp,"buf_Release bp 0x%p ref %d", bp, refCount);
193 afsi_log("%s:%d buf_ReleaseLocked bp 0x%p, ref %d", file, line, bp, refCount);
197 osi_panic("buf refcount 0",__FILE__,__LINE__);;
199 osi_assertx(refCount >= 0, "cm_buf_t refCount == 0");
202 lock_ObtainWrite(&buf_globalLock);
203 if (bp->refCount == 0 &&
204 !(bp->flags & CM_BUF_INLRU)) {
205 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
207 /* watch for transition from empty to one element */
208 if (!cm_data.buf_freeListEndp)
209 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
210 bp->flags |= CM_BUF_INLRU;
212 lock_ReleaseWrite(&buf_globalLock);
217 buf_Sync(int quitOnShutdown)
219 cm_buf_t **bpp, *bp, *prevbp;
220 afs_uint32 wasDirty = 0;
223 /* go through all of the dirty buffers */
224 lock_ObtainRead(&buf_globalLock);
225 for (bpp = &cm_data.buf_dirtyListp, prevbp = NULL; bp = *bpp; ) {
226 if (quitOnShutdown && buf_ShutdownFlag)
229 lock_ReleaseRead(&buf_globalLock);
230 /* all dirty buffers are held when they are added to the
231 * dirty list. No need for an additional hold.
233 lock_ObtainMutex(&bp->mx);
235 if (bp->flags & CM_BUF_DIRTY && !(bp->flags & CM_BUF_REDIR)) {
236 /* start cleaning the buffer; don't touch log pages since
237 * the log code counts on knowing exactly who is writing
238 * a log page at any given instant.
243 req.flags |= CM_REQ_NORETRY;
244 buf_CleanAsyncLocked(bp, &req, &dirty);
248 /* the buffer may or may not have been dirty
249 * and if dirty may or may not have been cleaned
250 * successfully. check the dirty flag again.
252 if (!(bp->flags & CM_BUF_DIRTY)) {
253 /* remove the buffer from the dirty list */
254 lock_ObtainWrite(&buf_globalLock);
255 #ifdef DEBUG_REFCOUNT
256 if (bp->dirtyp == NULL && bp != cm_data.buf_dirtyListEndp) {
257 osi_Log1(afsd_logp,"buf_IncrSyncer bp 0x%p list corruption",bp);
258 afsi_log("buf_IncrSyncer bp 0x%p list corruption", bp);
263 bp->flags &= ~CM_BUF_INDL;
264 if (cm_data.buf_dirtyListp == NULL)
265 cm_data.buf_dirtyListEndp = NULL;
266 else if (cm_data.buf_dirtyListEndp == bp)
267 cm_data.buf_dirtyListEndp = prevbp;
268 buf_ReleaseLocked(bp, TRUE);
269 lock_ConvertWToR(&buf_globalLock);
271 /* advance the pointer so we don't loop forever */
272 lock_ObtainRead(&buf_globalLock);
276 lock_ReleaseMutex(&bp->mx);
277 } /* for loop over a bunch of buffers */
278 lock_ReleaseRead(&buf_globalLock);
283 /* incremental sync daemon. Writes all dirty buffers every 5000 ms */
284 void buf_IncrSyncer(long parm)
289 while (buf_ShutdownFlag == 0) {
292 i = SleepEx(5000, 1);
297 wasDirty = buf_Sync(1);
298 } /* whole daemon's while loop */
302 buf_ValidateBuffers(void)
304 cm_buf_t * bp, *bpf, *bpa, *bpb;
305 afs_uint64 countb = 0, countf = 0, counta = 0;
307 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
308 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
309 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
310 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
314 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
315 if (bp->magic != CM_BUF_MAGIC) {
316 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
317 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
323 if (countb > cm_data.buf_nbuffers) {
324 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
325 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
330 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
331 if (bp->magic != CM_BUF_MAGIC) {
332 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
333 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
339 if (countf > cm_data.buf_nbuffers) {
340 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
341 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
346 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
347 if (bp->magic != CM_BUF_MAGIC) {
348 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
349 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
355 if (counta > cm_data.buf_nbuffers) {
356 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
357 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
362 if (countb != countf) {
363 afsi_log("cm_ValidateBuffers failure: countb != countf");
364 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
368 if (counta != cm_data.buf_nbuffers) {
369 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
370 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
377 void buf_Shutdown(void)
379 /* disable the buf_IncrSyncer() threads */
380 buf_ShutdownFlag = 1;
382 /* then force all dirty buffers to the file servers */
386 /* initialize the buffer package; called with no locks
387 * held during the initialization phase.
389 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
391 static osi_once_t once;
400 cm_data.buf_nbuffers = nbuffers;
402 /* Have to be able to reserve a whole chunk */
403 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
404 return CM_ERROR_TOOFEWBUFS;
407 /* recall for callouts */
410 if (osi_Once(&once)) {
411 /* initialize global locks */
412 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock", LOCK_HIERARCHY_BUF_GLOBAL);
415 /* remember this for those who want to reset it */
416 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
418 /* lower hash size to a prime number */
419 cm_data.buf_hashSize = osi_PrimeLessThan((afs_uint32)(cm_data.buf_nbuffers/7 + 1));
421 /* create hash table */
422 memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
424 /* another hash table */
425 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
427 /* create buffer headers and put in free list */
428 bp = cm_data.bufHeaderBaseAddress;
429 data = cm_data.bufDataBaseAddress;
430 cm_data.buf_allp = NULL;
432 for (i=0; i<cm_data.buf_nbuffers; i++) {
433 osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress,
434 "invalid cm_buf_t address");
435 osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
436 "invalid cm_buf_t data address");
438 /* allocate and zero some storage */
439 memset(bp, 0, sizeof(cm_buf_t));
440 bp->magic = CM_BUF_MAGIC;
441 /* thread on list of all buffers */
442 bp->allp = cm_data.buf_allp;
443 cm_data.buf_allp = bp;
445 osi_QAdd((osi_queue_t **)&cm_data.buf_freeListp, &bp->q);
446 bp->flags |= CM_BUF_INLRU;
447 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
449 /* grab appropriate number of bytes from aligned zone */
452 /* setup last buffer pointer */
454 cm_data.buf_freeListEndp = bp;
458 data += cm_data.buf_blockSize;
461 /* none reserved at first */
462 cm_data.buf_reservedBufs = 0;
464 /* just for safety's sake */
465 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
467 bp = cm_data.bufHeaderBaseAddress;
468 data = cm_data.bufDataBaseAddress;
470 for (i=0; i<cm_data.buf_nbuffers; i++) {
471 lock_InitializeMutex(&bp->mx, "Buffer mutex", LOCK_HIERARCHY_BUFFER);
474 bp->waitRequests = 0;
475 bp->flags &= ~CM_BUF_WAITING;
481 buf_ValidateBufQueues();
485 /* init the buffer trace log */
486 buf_logp = osi_LogCreate("buffer", 1000);
487 osi_LogEnable(buf_logp);
492 /* and create the incr-syncer */
493 phandle = thrd_Create(0, 0,
494 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
497 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
498 CloseHandle(phandle);
502 buf_ValidateBufQueues();
507 /* add nbuffers to the buffer pool, if possible.
508 * Called with no locks held.
510 long buf_AddBuffers(afs_uint64 nbuffers)
512 /* The size of a virtual cache cannot be changed after it has
513 * been created. Subsequent calls to MapViewofFile() with
514 * an existing mapping object name would not allow the
515 * object to be resized. Return failure immediately.
517 * A similar problem now occurs with the persistent cache
518 * given that the memory mapped file now contains a complex
521 afsi_log("request to add %d buffers to the existing cache of size %d denied",
522 nbuffers, cm_data.buf_nbuffers);
524 return CM_ERROR_INVAL;
527 /* interface to set the number of buffers to an exact figure.
528 * Called with no locks held.
530 long buf_SetNBuffers(afs_uint64 nbuffers)
533 return CM_ERROR_INVAL;
534 if (nbuffers == cm_data.buf_nbuffers)
536 else if (nbuffers > cm_data.buf_nbuffers)
537 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
539 return CM_ERROR_INVAL;
542 /* wait for reading or writing to clear; called with write-locked
543 * buffer and unlocked scp and returns with locked buffer.
545 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
550 osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
551 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
554 /* if no IO is happening, we're done */
555 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
558 /* otherwise I/O is happening, but some other thread is waiting for
559 * the I/O already. Wait for that guy to figure out what happened,
560 * and then check again.
562 if ( bp->flags & CM_BUF_WAITING ) {
565 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
567 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
568 bp->flags |= CM_BUF_WAITING;
569 bp->waitCount = bp->waitRequests = 1;
571 osi_SleepM((LONG_PTR)bp, &bp->mx);
573 smb_UpdateServerPriority();
575 lock_ObtainMutex(&bp->mx);
576 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
578 if (bp->waitCount == 0) {
579 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
580 bp->flags &= ~CM_BUF_WAITING;
581 bp->waitRequests = 0;
585 if (scp = cm_FindSCache(&bp->fid))
589 lock_ObtainRead(&scp->rw);
590 if (scp->flags & CM_SCACHEFLAG_WAITING) {
591 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
592 osi_Wakeup((LONG_PTR)&scp->flags);
594 lock_ReleaseRead(&scp->rw);
598 /* if we get here, the IO is done, but we may have to wakeup people waiting for
599 * the I/O to complete. Do so.
601 if (bp->flags & CM_BUF_WAITING) {
602 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
603 osi_Wakeup((LONG_PTR) bp);
605 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
608 cm_ReleaseSCache(scp);
611 /* find a buffer, if any, for a particular file ID and offset. Assumes
612 * that buf_globalLock is write locked when called.
614 cm_buf_t *buf_FindLocked(struct cm_scache *scp, osi_hyper_t *offsetp)
619 i = BUF_HASH(&scp->fid, offsetp);
620 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
621 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
622 && offsetp->LowPart == bp->offset.LowPart
623 && offsetp->HighPart == bp->offset.HighPart) {
629 /* return whatever we found, if anything */
633 /* find a buffer with offset *offsetp for vnode *scp. Called
634 * with no locks held.
636 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
640 lock_ObtainRead(&buf_globalLock);
641 bp = buf_FindLocked(scp, offsetp);
642 lock_ReleaseRead(&buf_globalLock);
647 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
650 * Makes sure that there's only one person writing this block
651 * at any given time, and also ensures that the log is forced sufficiently far,
652 * if this buffer contains logged data.
654 * Returns non-zero if the buffer was dirty.
656 afs_uint32 buf_CleanAsyncLocked(cm_buf_t *bp, cm_req_t *reqp, afs_uint32 *pisdirty)
659 afs_uint32 isdirty = 0;
660 cm_scache_t * scp = NULL;
663 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
665 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
667 lock_ReleaseMutex(&bp->mx);
669 scp = cm_FindSCache(&bp->fid);
671 osi_Log2(buf_logp, "buf_CleanAsyncLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
674 LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
675 code = (*cm_buf_opsp->Writep)(scp, &offset,
677 /* we might as well try to write all of the contiguous
678 * dirty buffers in one RPC
685 osi_Log3(buf_logp, "buf_CleanAsyncLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
687 cm_ReleaseSCache(scp);
690 osi_Log1(buf_logp, "buf_CleanAsyncLocked unable to start I/O - scp not found buf 0x%p", bp);
691 code = CM_ERROR_NOSUCHFILE;
694 lock_ObtainMutex(&bp->mx);
695 /* if the Write routine returns No Such File, clear the dirty flag
696 * because we aren't going to be able to write this data to the file
699 if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD || code == CM_ERROR_NOACCESS ||
700 code == CM_ERROR_QUOTA || code == CM_ERROR_SPACE || code == CM_ERROR_TOOBIG ||
701 code == CM_ERROR_READONLY || code == CM_ERROR_NOSUCHPATH){
702 bp->flags &= ~CM_BUF_DIRTY;
703 bp->flags |= CM_BUF_ERROR;
704 bp->dirty_offset = 0;
705 bp->dirty_length = 0;
707 bp->dataVersion = CM_BUF_VERSION_BAD;
713 /* Disk cache support */
714 /* write buffer to disk cache (synchronous for now) */
715 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
716 #endif /* DISKCACHE95 */
718 /* if we get here and retries are not permitted
719 * then we need to exit this loop regardless of
720 * whether or not we were able to clear the dirty bit
722 if (reqp->flags & CM_REQ_NORETRY)
726 /* if someone was waiting for the I/O that just completed or failed,
729 if (bp->flags & CM_BUF_WAITING) {
730 /* turn off flags and wakeup users */
731 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
732 osi_Wakeup((LONG_PTR) bp);
741 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
742 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
743 * The buffer must already be clean, and no I/O should be happening to it.
745 void buf_Recycle(cm_buf_t *bp)
750 cm_buf_t *prevBp, *nextBp;
752 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
754 /* if we get here, we know that the buffer still has a 0 ref count,
755 * and that it is clean and has no currently pending I/O. This is
756 * the dude to return.
757 * Remember that as long as the ref count is 0, we know that we won't
758 * have any lock conflicts, so we can grab the buffer lock out of
759 * order in the locking hierarchy.
761 osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
762 bp, bp->offset.HighPart, bp->offset.LowPart);
764 osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
765 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
766 "incorrect cm_buf_t flags");
767 lock_AssertWrite(&buf_globalLock);
769 if (bp->flags & CM_BUF_INHASH) {
770 /* Remove from hash */
772 i = BUF_HASH(&bp->fid, &bp->offset);
773 lbpp = &(cm_data.buf_scacheHashTablepp[i]);
774 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
779 /* we better find it */
780 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
782 *lbpp = bp->hashp; /* hash out */
785 /* Remove from file hash */
787 i = BUF_FILEHASH(&bp->fid);
788 prevBp = bp->fileHashBackp;
789 bp->fileHashBackp = NULL;
790 nextBp = bp->fileHashp;
791 bp->fileHashp = NULL;
793 prevBp->fileHashp = nextBp;
795 cm_data.buf_fileHashTablepp[i] = nextBp;
797 nextBp->fileHashBackp = prevBp;
799 bp->flags &= ~CM_BUF_INHASH;
802 /* make the fid unrecognizable */
803 memset(&bp->fid, 0, sizeof(cm_fid_t));
806 /* recycle a buffer, removing it from the free list, hashing in its new identity
807 * and returning it write-locked so that no one can use it. Called without
808 * any locks held, and can return an error if it loses the race condition and
809 * finds that someone else created the desired buffer.
811 * If success is returned, the buffer is returned write-locked.
813 * May be called with null scp and offsetp, if we're just trying to reclaim some
814 * space from the buffer pool. In that case, the buffer will be returned
815 * without being hashed into the hash table.
817 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
819 cm_buf_t *bp; /* buffer we're dealing with */
820 cm_buf_t *nextBp; /* next buffer in file hash chain */
821 afs_uint32 i; /* temp */
824 cm_InitReq(&req); /* just in case */
827 buf_ValidateBufQueues();
832 lock_ObtainRead(&scp->bufCreateLock);
833 lock_ObtainWrite(&buf_globalLock);
834 /* check to see if we lost the race */
836 if (bp = buf_FindLocked(scp, offsetp)) {
837 /* Do not call buf_ReleaseLocked() because we
838 * do not want to allow the buffer to be added
841 afs_int32 refCount = InterlockedDecrement(&bp->refCount);
842 #ifdef DEBUG_REFCOUNT
843 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, refCount);
844 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, refCount);
846 lock_ReleaseWrite(&buf_globalLock);
847 lock_ReleaseRead(&scp->bufCreateLock);
848 return CM_BUF_EXISTS;
852 /* does this fix the problem below? it's a simple solution. */
853 if (!cm_data.buf_freeListEndp)
855 lock_ReleaseWrite(&buf_globalLock);
856 lock_ReleaseRead(&scp->bufCreateLock);
857 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
862 /* for debugging, assert free list isn't empty, although we
863 * really should try waiting for a running tranasction to finish
864 * instead of this; or better, we should have a transaction
865 * throttler prevent us from entering this situation.
867 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
869 /* look at all buffers in free list, some of which may temp.
870 * have high refcounts and which then should be skipped,
871 * starting cleaning I/O for those which are dirty. If we find
872 * a clean buffer, we rehash it, lock it and return it.
874 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
875 /* check to see if it really has zero ref count. This
876 * code can bump refcounts, at least, so it may not be
879 if (bp->refCount > 0)
882 /* we don't have to lock buffer itself, since the ref
883 * count is 0 and we know it will stay zero as long as
884 * we hold the global lock.
887 /* Don't recycle a buffer held by the redirector. */
888 if (bp->flags & CM_BUF_REDIR)
891 /* don't recycle someone in our own chunk */
892 if (!cm_FidCmp(&bp->fid, &scp->fid)
893 && (bp->offset.LowPart & (-cm_chunkSize))
894 == (offsetp->LowPart & (-cm_chunkSize)))
897 /* if this page is being filled (!) or cleaned, see if
898 * the I/O has completed. If not, skip it, otherwise
899 * do the final processing for the I/O.
901 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
902 /* probably shouldn't do this much work while
903 * holding the big lock? Watch for contention
909 if (bp->flags & CM_BUF_DIRTY) {
910 /* if the buffer is dirty, start cleaning it and
911 * move on to the next buffer. We do this with
912 * just the lock required to minimize contention
916 lock_ReleaseWrite(&buf_globalLock);
917 lock_ReleaseRead(&scp->bufCreateLock);
919 /* grab required lock and clean; this only
920 * starts the I/O. By the time we're back,
921 * it'll still be marked dirty, but it will also
922 * have the WRITING flag set, so we won't get
925 buf_CleanAsync(bp, &req, NULL);
927 /* now put it back and go around again */
932 /* if we get here, we know that the buffer still has a 0
933 * ref count, and that it is clean and has no currently
934 * pending I/O. This is the dude to return.
935 * Remember that as long as the ref count is 0, we know
936 * that we won't have any lock conflicts, so we can grab
937 * the buffer lock out of order in the locking hierarchy.
941 /* clean up junk flags */
942 bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
943 bp->dataVersion = CM_BUF_VERSION_BAD; /* unknown so far */
945 /* now hash in as our new buffer, and give it the
946 * appropriate label, if requested.
949 bp->flags |= CM_BUF_INHASH;
954 bp->offset = *offsetp;
955 i = BUF_HASH(&scp->fid, offsetp);
956 bp->hashp = cm_data.buf_scacheHashTablepp[i];
957 cm_data.buf_scacheHashTablepp[i] = bp;
958 i = BUF_FILEHASH(&scp->fid);
959 nextBp = cm_data.buf_fileHashTablepp[i];
960 bp->fileHashp = nextBp;
961 bp->fileHashBackp = NULL;
963 nextBp->fileHashBackp = bp;
964 cm_data.buf_fileHashTablepp[i] = bp;
967 /* we should move it from the lru queue. It better still be there,
968 * since we've held the global (big) lock since we found it there.
970 osi_assertx(bp->flags & CM_BUF_INLRU,
971 "buf_GetNewLocked: LRU screwup");
973 if (cm_data.buf_freeListEndp == bp) {
974 /* we're the last guy in this queue, so maintain it */
975 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
977 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
978 bp->flags &= ~CM_BUF_INLRU;
980 /* prepare to return it. Give it a refcount */
982 #ifdef DEBUG_REFCOUNT
983 osi_Log2(afsd_logp,"buf_GetNewLocked bp 0x%p ref %d", bp, 1);
984 afsi_log("%s:%d buf_GetNewLocked bp 0x%p, ref %d", __FILE__, __LINE__, bp, 1);
986 /* grab the mutex so that people don't use it
987 * before the caller fills it with data. Again, no one
988 * should have been able to get to this dude to lock it.
990 if (!lock_TryMutex(&bp->mx)) {
991 osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked. refCount %d should be 0",
993 osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
996 lock_ReleaseWrite(&buf_globalLock);
997 lock_ReleaseRead(&scp->bufCreateLock);
1002 buf_ValidateBufQueues();
1003 #endif /* TESTING */
1005 } /* for all buffers in lru queue */
1006 lock_ReleaseWrite(&buf_globalLock);
1007 lock_ReleaseRead(&scp->bufCreateLock);
1008 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List has no buffers with a zero refcount - sleeping 100ms");
1009 Sleep(100); /* give some time for a buffer to be freed */
1010 } /* while loop over everything */
1014 /* get a page, returning it held but unlocked. Doesn't fill in the page
1015 * with I/O, since we're going to write the whole thing new.
1017 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
1021 osi_hyper_t pageOffset;
1025 pageOffset.HighPart = offsetp->HighPart;
1026 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
1028 bp = buf_Find(scp, &pageOffset);
1030 /* lock it and break out */
1031 lock_ObtainMutex(&bp->mx);
1035 /* otherwise, we have to create a page */
1036 code = buf_GetNewLocked(scp, &pageOffset, &bp);
1038 /* check if the buffer was created in a race condition branch.
1039 * If so, go around so we can hold a reference to it.
1041 if (code == CM_BUF_EXISTS)
1044 /* something else went wrong */
1048 /* otherwise, we have a locked buffer that we just created */
1051 } /* big while loop */
1053 /* wait for reads */
1054 if (bp->flags & CM_BUF_READING)
1055 buf_WaitIO(scp, bp);
1057 /* once it has been read once, we can unlock it and return it, still
1058 * with its refcount held.
1060 lock_ReleaseMutex(&bp->mx);
1062 osi_Log4(buf_logp, "buf_GetNew returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1063 bp, scp, offsetp->HighPart, offsetp->LowPart);
1067 /* get a page, returning it held but unlocked. Make sure it is complete */
1068 /* The scp must be unlocked when passed to this function */
1069 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
1073 osi_hyper_t pageOffset;
1074 unsigned long tcount;
1078 cm_diskcache_t *dcp;
1079 #endif /* DISKCACHE95 */
1082 pageOffset.HighPart = offsetp->HighPart;
1083 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
1087 buf_ValidateBufQueues();
1088 #endif /* TESTING */
1090 bp = buf_Find(scp, &pageOffset);
1092 /* lock it and break out */
1093 lock_ObtainMutex(&bp->mx);
1096 /* touch disk chunk to update LRU info */
1097 diskcache_Touch(bp->dcp);
1098 #endif /* DISKCACHE95 */
1102 /* otherwise, we have to create a page */
1103 code = buf_GetNewLocked(scp, &pageOffset, &bp);
1104 /* bp->mx is now held */
1106 /* check if the buffer was created in a race condition branch.
1107 * If so, go around so we can hold a reference to it.
1109 if (code == CM_BUF_EXISTS)
1112 /* something else went wrong */
1115 buf_ValidateBufQueues();
1116 #endif /* TESTING */
1120 /* otherwise, we have a locked buffer that we just created */
1123 } /* big while loop */
1125 /* if we get here, we have a locked buffer that may have just been
1126 * created, in which case it needs to be filled with data.
1129 /* load the page; freshly created pages should be idle */
1130 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
1132 /* start the I/O; may drop lock */
1133 bp->flags |= CM_BUF_READING;
1134 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
1137 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
1138 bp->dcp = dcp; /* pointer to disk cache struct. */
1139 #endif /* DISKCACHE95 */
1142 /* failure or queued */
1143 if (code != ERROR_IO_PENDING) {
1145 bp->flags |= CM_BUF_ERROR;
1146 bp->flags &= ~CM_BUF_READING;
1147 if (bp->flags & CM_BUF_WAITING) {
1148 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1149 osi_Wakeup((LONG_PTR) bp);
1151 lock_ReleaseMutex(&bp->mx);
1154 buf_ValidateBufQueues();
1155 #endif /* TESTING */
1159 /* otherwise, I/O completed instantly and we're done, except
1160 * for padding the xfr out with 0s and checking for EOF
1162 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1163 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1165 bp->flags |= CM_BUF_EOF;
1167 bp->flags &= ~CM_BUF_READING;
1168 if (bp->flags & CM_BUF_WAITING) {
1169 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1170 osi_Wakeup((LONG_PTR) bp);
1176 /* wait for reads, either that which we started above, or that someone
1177 * else started. We don't care if we return a buffer being cleaned.
1179 if (bp->flags & CM_BUF_READING)
1180 buf_WaitIO(scp, bp);
1182 /* once it has been read once, we can unlock it and return it, still
1183 * with its refcount held.
1185 lock_ReleaseMutex(&bp->mx);
1188 /* now remove from queue; will be put in at the head (farthest from
1189 * being recycled) when we're done in buf_Release.
1191 lock_ObtainWrite(&buf_globalLock);
1192 if (bp->flags & CM_BUF_INLRU) {
1193 if (cm_data.buf_freeListEndp == bp)
1194 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1195 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1196 bp->flags &= ~CM_BUF_INLRU;
1198 lock_ReleaseWrite(&buf_globalLock);
1200 osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1201 bp, scp, offsetp->HighPart, offsetp->LowPart);
1203 buf_ValidateBufQueues();
1204 #endif /* TESTING */
1208 /* count # of elements in the free list;
1209 * we don't bother doing the proper locking for accessing dataVersion or flags
1210 * since it is a pain, and this is really just an advisory call. If you need
1211 * to do better at some point, rewrite this function.
1213 long buf_CountFreeList(void)
1219 lock_ObtainRead(&buf_globalLock);
1220 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1221 /* if the buffer doesn't have an identity, or if the buffer
1222 * has been invalidate (by having its DV stomped upon), then
1223 * count it as free, since it isn't really being utilized.
1225 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion == CM_BUF_VERSION_BAD)
1228 lock_ReleaseRead(&buf_globalLock);
1232 /* clean a buffer synchronously */
1233 long buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp, afs_uint32 *pisdirty)
1236 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1238 lock_ObtainMutex(&bp->mx);
1239 code = buf_CleanAsyncLocked(bp, reqp, pisdirty);
1240 lock_ReleaseMutex(&bp->mx);
1245 /* wait for a buffer's cleaning to finish */
1246 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp, afs_uint32 locked)
1248 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1251 lock_ObtainMutex(&bp->mx);
1252 if (bp->flags & CM_BUF_WRITING) {
1253 buf_WaitIO(scp, bp);
1256 lock_ReleaseMutex(&bp->mx);
1259 /* set the dirty flag on a buffer, and set associated write-ahead log,
1260 * if there is one. Allow one to be added to a buffer, but not changed.
1262 * The buffer must be locked before calling this routine.
1264 void buf_SetDirty(cm_buf_t *bp, afs_uint32 offset, afs_uint32 length, cm_user_t *userp)
1266 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1267 osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
1269 if (bp->flags & CM_BUF_DIRTY) {
1271 osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
1273 if (bp->dirty_offset <= offset) {
1274 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1275 /* dirty_length remains the same */
1277 bp->dirty_length = offset + length - bp->dirty_offset;
1279 } else /* bp->dirty_offset > offset */ {
1280 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1281 bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
1283 bp->dirty_length = length;
1285 bp->dirty_offset = offset;
1288 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1291 bp->flags |= CM_BUF_DIRTY;
1293 /* and turn off EOF flag, since it has associated data now */
1294 bp->flags &= ~CM_BUF_EOF;
1296 bp->dirty_offset = offset;
1297 bp->dirty_length = length;
1299 /* and add to the dirty list.
1300 * we obtain a hold on the buffer for as long as it remains
1301 * in the list. buffers are only removed from the list by
1302 * the buf_IncrSyncer function regardless of when else the
1303 * dirty flag might be cleared.
1305 * This should never happen but just in case there is a bug
1306 * elsewhere, never add to the dirty list if the buffer is
1309 lock_ObtainWrite(&buf_globalLock);
1310 if (!(bp->flags & CM_BUF_INDL)) {
1312 if (!cm_data.buf_dirtyListp) {
1313 cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
1315 cm_data.buf_dirtyListEndp->dirtyp = bp;
1316 cm_data.buf_dirtyListEndp = bp;
1319 bp->flags |= CM_BUF_INDL;
1321 lock_ReleaseWrite(&buf_globalLock);
1324 /* and record the last writer */
1325 if (bp->userp != userp) {
1328 cm_ReleaseUser(bp->userp);
1333 /* clean all buffers, reset log pointers and invalidate all buffers.
1334 * Called with no locks held, and returns with same.
1336 * This function is guaranteed to clean and remove the log ptr of all the
1337 * buffers that were dirty or had non-zero log ptrs before the call was
1338 * made. That's sufficient to clean up any garbage left around by recovery,
1339 * which is all we're counting on this for; there may be newly created buffers
1340 * added while we're running, but that should be OK.
1342 * In an environment where there are no transactions (artificially imposed, for
1343 * example, when switching the database to raw mode), this function is used to
1344 * make sure that all updates have been written to the disk. In that case, we don't
1345 * really require that we forget the log association between pages and logs, but
1346 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1347 * have to worry about invalidating data in the buffers.
1349 * This function is used at the end of recovery as paranoia to get the recovered
1350 * database out to disk. It removes all references to the recovery log and cleans
1353 long buf_CleanAndReset(void)
1359 lock_ObtainRead(&buf_globalLock);
1360 for(i=0; i<cm_data.buf_hashSize; i++) {
1361 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
1362 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1364 lock_ReleaseRead(&buf_globalLock);
1366 /* now no locks are held; clean buffer and go on */
1368 req.flags |= CM_REQ_NORETRY;
1370 buf_CleanAsync(bp, &req, NULL);
1371 buf_CleanWait(NULL, bp, FALSE);
1373 /* relock and release buffer */
1374 lock_ObtainRead(&buf_globalLock);
1375 buf_ReleaseLocked(bp, FALSE);
1377 } /* over one bucket */
1378 } /* for loop over all hash buckets */
1381 lock_ReleaseRead(&buf_globalLock);
1384 buf_ValidateBufQueues();
1385 #endif /* TESTING */
1387 /* and we're done */
1391 /* called without global lock being held, reserves buffers for callers
1392 * that need more than one held (not locked) at once.
1394 void buf_ReserveBuffers(afs_uint64 nbuffers)
1396 lock_ObtainWrite(&buf_globalLock);
1398 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1399 cm_data.buf_reserveWaiting = 1;
1400 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1401 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1402 lock_ObtainWrite(&buf_globalLock);
1405 cm_data.buf_reservedBufs += nbuffers;
1409 lock_ReleaseWrite(&buf_globalLock);
1412 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1416 lock_ObtainWrite(&buf_globalLock);
1417 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1421 cm_data.buf_reservedBufs += nbuffers;
1424 lock_ReleaseWrite(&buf_globalLock);
1428 /* called without global lock held, releases reservation held by
1429 * buf_ReserveBuffers.
1431 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1433 lock_ObtainWrite(&buf_globalLock);
1434 cm_data.buf_reservedBufs -= nbuffers;
1435 if (cm_data.buf_reserveWaiting) {
1436 cm_data.buf_reserveWaiting = 0;
1437 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1439 lock_ReleaseWrite(&buf_globalLock);
1442 /* truncate the buffers past sizep, zeroing out the page, if we don't
1443 * end on a page boundary.
1445 * Requires cm_bufCreateLock to be write locked.
1447 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1451 cm_buf_t *nbufp; /* next buffer, if didRelease */
1457 /* assert that cm_bufCreateLock is held in write mode */
1458 lock_AssertWrite(&scp->bufCreateLock);
1460 i = BUF_FILEHASH(&scp->fid);
1462 lock_ObtainRead(&buf_globalLock);
1463 bufp = cm_data.buf_fileHashTablepp[i];
1465 lock_ReleaseRead(&buf_globalLock);
1469 buf_HoldLocked(bufp);
1470 lock_ReleaseRead(&buf_globalLock);
1472 lock_ObtainMutex(&bufp->mx);
1474 bufEnd.HighPart = 0;
1475 bufEnd.LowPart = cm_data.buf_blockSize;
1476 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1478 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1479 LargeIntegerLessThan(*sizep, bufEnd)) {
1480 buf_WaitIO(scp, bufp);
1482 lock_ObtainWrite(&scp->rw);
1484 /* make sure we have a callback (so we have the right value for
1485 * the length), and wait for it to be safe to do a truncate.
1487 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1488 CM_SCACHESYNC_NEEDCALLBACK
1489 | CM_SCACHESYNC_GETSTATUS
1490 | CM_SCACHESYNC_SETSIZE
1491 | CM_SCACHESYNC_BUFLOCKED);
1494 /* if we succeeded in our locking, and this applies to the right
1495 * file, and the truncate request overlaps the buffer either
1496 * totally or partially, then do something.
1498 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1499 && LargeIntegerLessThan(*sizep, bufEnd)) {
1502 /* destroy the buffer, turning off its dirty bit, if
1503 * we're truncating the whole buffer. Otherwise, set
1504 * the dirty bit, and clear out the tail of the buffer
1505 * if we just overlap some.
1507 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1508 /* truncating the entire page */
1509 bufp->flags &= ~CM_BUF_DIRTY;
1510 bufp->dirty_offset = 0;
1511 bufp->dirty_length = 0;
1512 bufp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1513 bufp->dirtyCounter++;
1516 /* don't set dirty, since dirty implies
1517 * currently up-to-date. Don't need to do this,
1518 * since we'll update the length anyway.
1520 * Zero out remainder of the page, in case we
1521 * seek and write past EOF, and make this data
1524 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1525 osi_assertx(bufferPos != 0, "non-zero bufferPos");
1526 memset(bufp->datap + bufferPos, 0,
1527 cm_data.buf_blockSize - bufferPos);
1531 cm_SyncOpDone( scp, bufp,
1532 CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
1533 | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
1535 lock_ReleaseWrite(&scp->rw);
1536 lock_ReleaseMutex(&bufp->mx);
1539 nbufp = bufp->fileHashp;
1543 /* This forces the loop to end and the error code
1544 * to be returned. */
1552 buf_ValidateBufQueues();
1553 #endif /* TESTING */
1559 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1562 cm_buf_t *bp; /* buffer we're hacking on */
1567 i = BUF_FILEHASH(&scp->fid);
1570 lock_ObtainRead(&buf_globalLock);
1571 bp = cm_data.buf_fileHashTablepp[i];
1574 lock_ReleaseRead(&buf_globalLock);
1576 for (; bp; bp = nbp) {
1577 didRelease = 0; /* haven't released this buffer yet */
1579 /* clean buffer synchronously */
1580 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1581 lock_ObtainMutex(&bp->mx);
1583 /* start cleaning the buffer, and wait for it to finish */
1584 buf_CleanAsyncLocked(bp, reqp, NULL);
1585 buf_WaitIO(scp, bp);
1586 lock_ReleaseMutex(&bp->mx);
1589 * if the error for the previous buffer was BADFD
1590 * then all buffers for the FID are bad. Do not
1591 * attempt to stabalize.
1593 if (code != CM_ERROR_BADFD) {
1594 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1595 if (code && code != CM_ERROR_BADFD)
1598 if (code == CM_ERROR_BADFD) {
1599 /* if the scp's FID is bad its because we received VNOVNODE
1600 * when attempting to FetchStatus before the write. This
1601 * page therefore contains data that can no longer be stored.
1603 lock_ObtainMutex(&bp->mx);
1604 bp->flags &= ~CM_BUF_DIRTY;
1605 bp->flags |= CM_BUF_ERROR;
1606 bp->error = CM_ERROR_BADFD;
1607 bp->dirty_offset = 0;
1608 bp->dirty_length = 0;
1609 bp->dataVersion = CM_BUF_VERSION_BAD; /* known bad */
1611 lock_ReleaseMutex(&bp->mx);
1614 /* actually, we only know that buffer is clean if ref
1615 * count is 1, since we don't have buffer itself locked.
1617 if (!(bp->flags & CM_BUF_DIRTY)) {
1618 lock_ObtainWrite(&buf_globalLock);
1619 if (bp->refCount == 1) { /* bp is held above */
1620 nbp = bp->fileHashp;
1622 buf_HoldLocked(nbp);
1623 buf_ReleaseLocked(bp, TRUE);
1627 lock_ReleaseWrite(&buf_globalLock);
1631 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1636 lock_ObtainRead(&buf_globalLock);
1637 nbp = bp->fileHashp;
1639 buf_HoldLocked(nbp);
1640 buf_ReleaseLocked(bp, FALSE);
1641 lock_ReleaseRead(&buf_globalLock);
1643 } /* for loop over a bunch of buffers */
1646 buf_ValidateBufQueues();
1647 #endif /* TESTING */
1653 /* Must be called with scp->rw held */
1654 long buf_ForceDataVersion(cm_scache_t * scp, afs_uint64 fromVersion, afs_uint64 toVersion)
1660 lock_AssertAny(&scp->rw);
1662 i = BUF_FILEHASH(&scp->fid);
1664 lock_ObtainRead(&buf_globalLock);
1666 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
1667 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1668 if (bp->dataVersion == fromVersion) {
1669 bp->dataVersion = toVersion;
1674 lock_ReleaseRead(&buf_globalLock);
1682 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1686 cm_buf_t *bp; /* buffer we're hacking on */
1687 cm_buf_t *nbp; /* next one */
1690 i = BUF_FILEHASH(&scp->fid);
1692 lock_ObtainRead(&buf_globalLock);
1693 bp = cm_data.buf_fileHashTablepp[i];
1696 lock_ReleaseRead(&buf_globalLock);
1697 for (; bp; bp = nbp) {
1698 /* clean buffer synchronously */
1699 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1700 lock_ObtainMutex(&bp->mx);
1701 if (bp->flags & CM_BUF_DIRTY) {
1702 if (userp && userp != bp->userp) {
1705 cm_ReleaseUser(bp->userp);
1710 case CM_ERROR_NOSUCHFILE:
1711 case CM_ERROR_BADFD:
1712 case CM_ERROR_NOACCESS:
1713 case CM_ERROR_QUOTA:
1714 case CM_ERROR_SPACE:
1715 case CM_ERROR_TOOBIG:
1716 case CM_ERROR_READONLY:
1717 case CM_ERROR_NOSUCHPATH:
1719 * Apply the previous fatal error to this buffer.
1720 * Do not waste the time attempting to store to
1721 * the file server when we know it will fail.
1723 bp->flags &= ~CM_BUF_DIRTY;
1724 bp->flags |= CM_BUF_ERROR;
1725 bp->dirty_offset = 0;
1726 bp->dirty_length = 0;
1728 bp->dataVersion = CM_BUF_VERSION_BAD;
1732 code = buf_CleanAsyncLocked(bp, reqp, &wasDirty);
1733 if (bp->flags & CM_BUF_ERROR) {
1739 buf_CleanWait(scp, bp, TRUE);
1741 lock_ReleaseMutex(&bp->mx);
1744 lock_ObtainRead(&buf_globalLock);
1745 nbp = bp->fileHashp;
1747 buf_HoldLocked(nbp);
1748 buf_ReleaseLocked(bp, FALSE);
1749 lock_ReleaseRead(&buf_globalLock);
1750 } /* for loop over a bunch of buffers */
1753 buf_ValidateBufQueues();
1754 #endif /* TESTING */
1762 buf_ValidateBufQueues(void)
1764 cm_buf_t * bp, *bpb, *bpf, *bpa;
1765 afs_uint32 countf=0, countb=0, counta=0;
1767 lock_ObtainRead(&buf_globalLock);
1768 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1769 if (bp->magic != CM_BUF_MAGIC)
1770 osi_panic("buf magic error",__FILE__,__LINE__);
1775 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1776 if (bp->magic != CM_BUF_MAGIC)
1777 osi_panic("buf magic error",__FILE__,__LINE__);
1782 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1783 if (bp->magic != CM_BUF_MAGIC)
1784 osi_panic("buf magic error",__FILE__,__LINE__);
1788 lock_ReleaseRead(&buf_globalLock);
1790 if (countb != countf)
1791 osi_panic("buf magic error",__FILE__,__LINE__);
1793 if (counta != cm_data.buf_nbuffers)
1794 osi_panic("buf magic error",__FILE__,__LINE__);
1796 #endif /* TESTING */
1798 /* dump the contents of the buf_scacheHashTablepp. */
1799 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1806 if (cm_data.buf_scacheHashTablepp == NULL)
1810 lock_ObtainRead(&buf_globalLock);
1812 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n",
1813 cookie, cm_data.buf_hashSize);
1814 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1816 for (i = 0; i < cm_data.buf_hashSize; i++)
1818 for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp)
1820 StringCbPrintfA(output, sizeof(output),
1821 "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
1822 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1823 "flags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1824 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1825 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1826 bp->offset.LowPart, bp->dataVersion, bp->flags,
1827 bp->cmFlags, bp->error, bp->refCount);
1828 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1832 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
1833 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1835 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
1836 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1837 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1838 StringCbPrintfA(output, sizeof(output),
1839 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1840 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1841 "flags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1842 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1843 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1844 bp->offset.LowPart, bp->dataVersion, bp->flags,
1845 bp->cmFlags, bp->error, bp->refCount);
1846 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1848 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
1849 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1851 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListp\r\n", cookie);
1852 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1853 for(bp = cm_data.buf_dirtyListp; bp; bp=bp->dirtyp) {
1854 StringCbPrintfA(output, sizeof(output),
1855 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1856 "vnode=%d, unique=%d), offset=%x:%08x, dv=%I64d, "
1857 "flags=0x%x, cmFlags=0x%x, error=0x%x, refCount=%d\r\n",
1858 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1859 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1860 bp->offset.LowPart, bp->dataVersion, bp->flags,
1861 bp->cmFlags, bp->error, bp->refCount);
1862 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1864 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListp.\r\n", cookie);
1865 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1868 lock_ReleaseRead(&buf_globalLock);
1872 void buf_ForceTrace(BOOL flush)
1881 len = GetTempPath(sizeof(buf)-10, buf);
1882 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1883 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1884 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1885 if (handle == INVALID_HANDLE_VALUE) {
1886 osi_panic("Cannot create log file", __FILE__, __LINE__);
1888 osi_LogPrint(buf_logp, handle);
1890 FlushFileBuffers(handle);
1891 CloseHandle(handle);
1894 long buf_DirtyBuffersExist(cm_fid_t *fidp)
1897 afs_uint32 bcount = 0;
1900 i = BUF_FILEHASH(fidp);
1902 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->allp, bcount++) {
1903 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY))
1910 long buf_CleanDirtyBuffers(cm_scache_t *scp)
1913 afs_uint32 bcount = 0;
1914 cm_fid_t * fidp = &scp->fid;
1916 for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
1917 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
1919 lock_ObtainMutex(&bp->mx);
1920 bp->cmFlags &= ~CM_BUF_CMSTORING;
1921 bp->flags &= ~CM_BUF_DIRTY;
1922 bp->dirty_offset = 0;
1923 bp->dirty_length = 0;
1924 bp->flags |= CM_BUF_ERROR;
1925 bp->error = VNOVNODE;
1926 bp->dataVersion = CM_BUF_VERSION_BAD; /* bad */
1928 if (bp->flags & CM_BUF_WAITING) {
1929 osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
1930 osi_Wakeup((long) &bp);
1932 lock_ReleaseMutex(&bp->mx);