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;
90 void buf_HoldLocked(cm_buf_t *bp)
92 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
96 /* hold a reference to an already held buffer */
97 void buf_Hold(cm_buf_t *bp)
99 lock_ObtainWrite(&buf_globalLock);
101 lock_ReleaseWrite(&buf_globalLock);
104 /* code to drop reference count while holding buf_globalLock */
105 void buf_ReleaseLocked(cm_buf_t *bp)
107 /* ensure that we're in the LRU queue if our ref count is 0 */
108 osi_assertx(bp->magic == CM_BUF_MAGIC,"incorrect cm_buf_t magic");
110 if (bp->refCount == 0)
111 osi_panic("buf refcount 0",__FILE__,__LINE__);;
113 osi_assertx(bp->refCount > 0, "cm_buf_t refCount == 0");
115 if (--bp->refCount == 0) {
116 if (!(bp->flags & CM_BUF_INLRU)) {
117 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
119 /* watch for transition from empty to one element */
120 if (!cm_data.buf_freeListEndp)
121 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
122 bp->flags |= CM_BUF_INLRU;
127 /* release a buffer. Buffer must be referenced, but unlocked. */
128 void buf_Release(cm_buf_t *bp)
130 lock_ObtainWrite(&buf_globalLock);
131 buf_ReleaseLocked(bp);
132 lock_ReleaseWrite(&buf_globalLock);
135 /* incremental sync daemon. Writes all dirty buffers every 5000 ms */
136 void buf_IncrSyncer(long parm)
139 long i; /* counter */
143 while (buf_ShutdownFlag == 0) {
145 i = SleepEx(5000, 1);
146 if (i != 0) continue;
151 /* now go through our percentage of the buffers */
152 for (bpp = &cm_data.buf_dirtyListp; bp = *bpp; ) {
154 /* all dirty buffers are held when they are added to the
155 * dirty list. No need for an additional hold.
158 if (bp->flags & CM_BUF_DIRTY) {
159 /* start cleaning the buffer; don't touch log pages since
160 * the log code counts on knowing exactly who is writing
161 * a log page at any given instant.
164 req.flags |= CM_REQ_NORETRY;
165 wasDirty |= buf_CleanAsync(bp, &req);
168 /* the buffer may or may not have been dirty
169 * and if dirty may or may not have been cleaned
170 * successfully. check the dirty flag again.
172 if (!(bp->flags & CM_BUF_DIRTY)) {
173 lock_ObtainMutex(&bp->mx);
174 if (!(bp->flags & CM_BUF_DIRTY)) {
175 /* remove the buffer from the dirty list */
176 lock_ObtainWrite(&buf_globalLock);
179 if (cm_data.buf_dirtyListp == NULL)
180 cm_data.buf_dirtyListEndp = NULL;
181 buf_ReleaseLocked(bp);
182 lock_ReleaseWrite(&buf_globalLock);
184 /* advance the pointer so we don't loop forever */
187 lock_ReleaseMutex(&bp->mx);
189 /* advance the pointer so we don't loop forever */
192 } /* for loop over a bunch of buffers */
193 } /* whole daemon's while loop */
197 buf_ValidateBuffers(void)
199 cm_buf_t * bp, *bpf, *bpa, *bpb;
200 afs_uint64 countb = 0, countf = 0, counta = 0;
202 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
203 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
204 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
205 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
209 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
210 if (bp->magic != CM_BUF_MAGIC) {
211 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
212 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
218 if (countb > cm_data.buf_nbuffers) {
219 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
220 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
225 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
226 if (bp->magic != CM_BUF_MAGIC) {
227 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
228 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
234 if (countf > cm_data.buf_nbuffers) {
235 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
236 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
241 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
242 if (bp->magic != CM_BUF_MAGIC) {
243 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
244 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
250 if (counta > cm_data.buf_nbuffers) {
251 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
252 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
257 if (countb != countf) {
258 afsi_log("cm_ValidateBuffers failure: countb != countf");
259 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
263 if (counta != cm_data.buf_nbuffers) {
264 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
265 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
272 void buf_Shutdown(void)
274 buf_ShutdownFlag = 1;
277 /* initialize the buffer package; called with no locks
278 * held during the initialization phase.
280 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
282 static osi_once_t once;
291 cm_data.buf_nbuffers = nbuffers;
293 /* Have to be able to reserve a whole chunk */
294 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
295 return CM_ERROR_TOOFEWBUFS;
298 /* recall for callouts */
301 if (osi_Once(&once)) {
302 /* initialize global locks */
303 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock");
306 /* remember this for those who want to reset it */
307 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
309 /* lower hash size to a prime number */
310 cm_data.buf_hashSize = osi_PrimeLessThan((afs_uint32)(cm_data.buf_nbuffers/7 + 1));
312 /* create hash table */
313 memset((void *)cm_data.buf_scacheHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
315 /* another hash table */
316 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
318 /* create buffer headers and put in free list */
319 bp = cm_data.bufHeaderBaseAddress;
320 data = cm_data.bufDataBaseAddress;
321 cm_data.buf_allp = NULL;
323 for (i=0; i<cm_data.buf_nbuffers; i++) {
324 osi_assertx(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress,
325 "invalid cm_buf_t address");
326 osi_assertx(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData,
327 "invalid cm_buf_t data address");
329 /* allocate and zero some storage */
330 memset(bp, 0, sizeof(cm_buf_t));
331 bp->magic = CM_BUF_MAGIC;
332 /* thread on list of all buffers */
333 bp->allp = cm_data.buf_allp;
334 cm_data.buf_allp = bp;
336 osi_QAdd((osi_queue_t **)&cm_data.buf_freeListp, &bp->q);
337 bp->flags |= CM_BUF_INLRU;
338 lock_InitializeMutex(&bp->mx, "Buffer mutex");
340 /* grab appropriate number of bytes from aligned zone */
343 /* setup last buffer pointer */
345 cm_data.buf_freeListEndp = bp;
349 data += cm_data.buf_blockSize;
352 /* none reserved at first */
353 cm_data.buf_reservedBufs = 0;
355 /* just for safety's sake */
356 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
358 bp = cm_data.bufHeaderBaseAddress;
359 data = cm_data.bufDataBaseAddress;
361 for (i=0; i<cm_data.buf_nbuffers; i++) {
362 lock_InitializeMutex(&bp->mx, "Buffer mutex");
365 bp->waitRequests = 0;
366 bp->flags &= ~CM_BUF_WAITING;
372 buf_ValidateBufQueues();
376 /* init the buffer trace log */
377 buf_logp = osi_LogCreate("buffer", 1000);
378 osi_LogEnable(buf_logp);
383 /* and create the incr-syncer */
384 phandle = thrd_Create(0, 0,
385 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
388 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
389 CloseHandle(phandle);
393 buf_ValidateBufQueues();
398 /* add nbuffers to the buffer pool, if possible.
399 * Called with no locks held.
401 long buf_AddBuffers(afs_uint64 nbuffers)
403 /* The size of a virtual cache cannot be changed after it has
404 * been created. Subsequent calls to MapViewofFile() with
405 * an existing mapping object name would not allow the
406 * object to be resized. Return failure immediately.
408 * A similar problem now occurs with the persistent cache
409 * given that the memory mapped file now contains a complex
412 afsi_log("request to add %d buffers to the existing cache of size %d denied",
413 nbuffers, cm_data.buf_nbuffers);
415 return CM_ERROR_INVAL;
418 /* interface to set the number of buffers to an exact figure.
419 * Called with no locks held.
421 long buf_SetNBuffers(afs_uint64 nbuffers)
424 return CM_ERROR_INVAL;
425 if (nbuffers == cm_data.buf_nbuffers)
427 else if (nbuffers > cm_data.buf_nbuffers)
428 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
430 return CM_ERROR_INVAL;
433 /* wait for reading or writing to clear; called with write-locked
434 * buffer and unlocked scp and returns with locked buffer.
436 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
441 osi_assertx(scp->magic == CM_SCACHE_MAGIC, "invalid cm_scache_t magic");
442 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
445 /* if no IO is happening, we're done */
446 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
449 /* otherwise I/O is happening, but some other thread is waiting for
450 * the I/O already. Wait for that guy to figure out what happened,
451 * and then check again.
453 if ( bp->flags & CM_BUF_WAITING ) {
456 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
458 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
459 bp->flags |= CM_BUF_WAITING;
460 bp->waitCount = bp->waitRequests = 1;
462 osi_SleepM((LONG_PTR)bp, &bp->mx);
464 smb_UpdateServerPriority();
466 lock_ObtainMutex(&bp->mx);
467 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
469 if (bp->waitCount == 0) {
470 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
471 bp->flags &= ~CM_BUF_WAITING;
472 bp->waitRequests = 0;
476 if (scp = cm_FindSCache(&bp->fid))
480 lock_ObtainMutex(&scp->mx);
481 if (scp->flags & CM_SCACHEFLAG_WAITING) {
482 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%p", scp);
483 osi_Wakeup((LONG_PTR)&scp->flags);
485 lock_ReleaseMutex(&scp->mx);
489 /* if we get here, the IO is done, but we may have to wakeup people waiting for
490 * the I/O to complete. Do so.
492 if (bp->flags & CM_BUF_WAITING) {
493 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
494 osi_Wakeup((LONG_PTR) bp);
496 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%p", bp);
499 cm_ReleaseSCache(scp);
502 /* find a buffer, if any, for a particular file ID and offset. Assumes
503 * that buf_globalLock is write locked when called.
505 cm_buf_t *buf_FindLocked(struct cm_scache *scp, osi_hyper_t *offsetp)
510 i = BUF_HASH(&scp->fid, offsetp);
511 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp) {
512 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
513 && offsetp->LowPart == bp->offset.LowPart
514 && offsetp->HighPart == bp->offset.HighPart) {
520 /* return whatever we found, if anything */
524 /* find a buffer with offset *offsetp for vnode *scp. Called
525 * with no locks held.
527 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
531 lock_ObtainWrite(&buf_globalLock);
532 bp = buf_FindLocked(scp, offsetp);
533 lock_ReleaseWrite(&buf_globalLock);
538 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
541 * Makes sure that there's only one person writing this block
542 * at any given time, and also ensures that the log is forced sufficiently far,
543 * if this buffer contains logged data.
545 * Returns non-zero if the buffer was dirty.
547 long buf_CleanAsyncLocked(cm_buf_t *bp, cm_req_t *reqp)
551 cm_scache_t * scp = NULL;
554 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
556 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
558 lock_ReleaseMutex(&bp->mx);
560 scp = cm_FindSCache(&bp->fid);
562 osi_Log2(buf_logp, "buf_CleanAsyncLocked starts I/O on scp 0x%p buf 0x%p", scp, bp);
565 LargeIntegerAdd(offset, ConvertLongToLargeInteger(bp->dirty_offset));
566 code = (*cm_buf_opsp->Writep)(scp, &offset, bp->dirty_length, 0, bp->userp, reqp);
567 osi_Log3(buf_logp, "buf_CleanAsyncLocked I/O on scp 0x%p buf 0x%p, done=%d", scp, bp, code);
569 cm_ReleaseSCache(scp);
572 osi_Log1(buf_logp, "buf_CleanAsyncLocked unable to start I/O - scp not found buf 0x%p", bp);
573 code = CM_ERROR_NOSUCHFILE;
576 lock_ObtainMutex(&bp->mx);
577 /* if the Write routine returns No Such File, clear the dirty flag
578 * because we aren't going to be able to write this data to the file
581 if (code == CM_ERROR_NOSUCHFILE || code == CM_ERROR_BADFD){
582 bp->flags &= ~CM_BUF_DIRTY;
583 bp->flags |= CM_BUF_ERROR;
584 bp->dirty_offset = 0;
585 bp->dirty_length = 0;
587 bp->dataVersion = -1; /* bad */
592 /* Disk cache support */
593 /* write buffer to disk cache (synchronous for now) */
594 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
595 #endif /* DISKCACHE95 */
597 /* if we get here and retries are not permitted
598 * then we need to exit this loop regardless of
599 * whether or not we were able to clear the dirty bit
601 if (reqp->flags & CM_REQ_NORETRY)
605 if (!(bp->flags & CM_BUF_DIRTY)) {
606 /* remove buffer from dirty buffer queue */
610 /* do logging after call to GetLastError, or else */
612 /* if someone was waiting for the I/O that just completed or failed,
615 if (bp->flags & CM_BUF_WAITING) {
616 /* turn off flags and wakeup users */
617 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
618 osi_Wakeup((LONG_PTR) bp);
623 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
624 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
625 * The buffer must already be clean, and no I/O should be happening to it.
627 void buf_Recycle(cm_buf_t *bp)
632 cm_buf_t *prevBp, *nextBp;
634 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
636 /* if we get here, we know that the buffer still has a 0 ref count,
637 * and that it is clean and has no currently pending I/O. This is
638 * the dude to return.
639 * Remember that as long as the ref count is 0, we know that we won't
640 * have any lock conflicts, so we can grab the buffer lock out of
641 * order in the locking hierarchy.
643 osi_Log3( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x:%08x",
644 bp, bp->offset.HighPart, bp->offset.LowPart);
646 osi_assertx(bp->refCount == 0, "cm_buf_t refcount != 0");
647 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)),
648 "incorrect cm_buf_t flags");
649 lock_AssertWrite(&buf_globalLock);
651 if (bp->flags & CM_BUF_INHASH) {
652 /* Remove from hash */
654 i = BUF_HASH(&bp->fid, &bp->offset);
655 lbpp = &(cm_data.buf_scacheHashTablepp[i]);
656 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
657 if (tbp == bp) break;
660 /* we better find it */
661 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
663 *lbpp = bp->hashp; /* hash out */
665 /* Remove from file hash */
667 i = BUF_FILEHASH(&bp->fid);
668 prevBp = bp->fileHashBackp;
669 nextBp = bp->fileHashp;
671 prevBp->fileHashp = nextBp;
673 cm_data.buf_fileHashTablepp[i] = nextBp;
675 nextBp->fileHashBackp = prevBp;
677 bp->flags &= ~CM_BUF_INHASH;
680 /* bump the soft reference counter now, to invalidate softRefs; no
681 * wakeup is required since people don't sleep waiting for this
686 /* make the fid unrecognizable */
687 memset(&bp->fid, 0, sizeof(cm_fid_t));
690 /* recycle a buffer, removing it from the free list, hashing in its new identity
691 * and returning it write-locked so that no one can use it. Called without
692 * any locks held, and can return an error if it loses the race condition and
693 * finds that someone else created the desired buffer.
695 * If success is returned, the buffer is returned write-locked.
697 * May be called with null scp and offsetp, if we're just trying to reclaim some
698 * space from the buffer pool. In that case, the buffer will be returned
699 * without being hashed into the hash table.
701 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
703 cm_buf_t *bp; /* buffer we're dealing with */
704 cm_buf_t *nextBp; /* next buffer in file hash chain */
708 cm_InitReq(&req); /* just in case */
711 buf_ValidateBufQueues();
716 lock_ObtainWrite(&buf_globalLock);
717 /* check to see if we lost the race */
719 if (bp = buf_FindLocked(scp, offsetp)) {
720 /* Do not call buf_ReleaseLocked() because we
721 * do not want to allow the buffer to be added
725 lock_ReleaseWrite(&buf_globalLock);
726 return CM_BUF_EXISTS;
730 /* does this fix the problem below? it's a simple solution. */
731 if (!cm_data.buf_freeListEndp)
733 lock_ReleaseWrite(&buf_globalLock);
734 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List is empty - sleeping 200ms");
739 /* for debugging, assert free list isn't empty, although we
740 * really should try waiting for a running tranasction to finish
741 * instead of this; or better, we should have a transaction
742 * throttler prevent us from entering this situation.
744 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
746 /* look at all buffers in free list, some of which may temp.
747 * have high refcounts and which then should be skipped,
748 * starting cleaning I/O for those which are dirty. If we find
749 * a clean buffer, we rehash it, lock it and return it.
751 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
752 /* check to see if it really has zero ref count. This
753 * code can bump refcounts, at least, so it may not be
756 if (bp->refCount > 0)
759 /* we don't have to lock buffer itself, since the ref
760 * count is 0 and we know it will stay zero as long as
761 * we hold the global lock.
764 /* don't recycle someone in our own chunk */
765 if (!cm_FidCmp(&bp->fid, &scp->fid)
766 && (bp->offset.LowPart & (-cm_chunkSize))
767 == (offsetp->LowPart & (-cm_chunkSize)))
770 /* if this page is being filled (!) or cleaned, see if
771 * the I/O has completed. If not, skip it, otherwise
772 * do the final processing for the I/O.
774 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
775 /* probably shouldn't do this much work while
776 * holding the big lock? Watch for contention
782 if (bp->flags & CM_BUF_DIRTY) {
783 /* if the buffer is dirty, start cleaning it and
784 * move on to the next buffer. We do this with
785 * just the lock required to minimize contention
789 lock_ReleaseWrite(&buf_globalLock);
791 /* grab required lock and clean; this only
792 * starts the I/O. By the time we're back,
793 * it'll still be marked dirty, but it will also
794 * have the WRITING flag set, so we won't get
797 buf_CleanAsync(bp, &req);
799 /* now put it back and go around again */
804 /* if we get here, we know that the buffer still has a 0
805 * ref count, and that it is clean and has no currently
806 * pending I/O. This is the dude to return.
807 * Remember that as long as the ref count is 0, we know
808 * that we won't have any lock conflicts, so we can grab
809 * the buffer lock out of order in the locking hierarchy.
813 /* clean up junk flags */
814 bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
815 bp->dataVersion = -1; /* unknown so far */
817 /* now hash in as our new buffer, and give it the
818 * appropriate label, if requested.
821 bp->flags |= CM_BUF_INHASH;
826 bp->offset = *offsetp;
827 i = BUF_HASH(&scp->fid, offsetp);
828 bp->hashp = cm_data.buf_scacheHashTablepp[i];
829 cm_data.buf_scacheHashTablepp[i] = bp;
830 i = BUF_FILEHASH(&scp->fid);
831 nextBp = cm_data.buf_fileHashTablepp[i];
832 bp->fileHashp = nextBp;
833 bp->fileHashBackp = NULL;
835 nextBp->fileHashBackp = bp;
836 cm_data.buf_fileHashTablepp[i] = bp;
839 /* we should move it from the lru queue. It better still be there,
840 * since we've held the global (big) lock since we found it there.
842 osi_assertx(bp->flags & CM_BUF_INLRU,
843 "buf_GetNewLocked: LRU screwup");
845 if (cm_data.buf_freeListEndp == bp) {
846 /* we're the last guy in this queue, so maintain it */
847 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
849 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
850 bp->flags &= ~CM_BUF_INLRU;
852 /* grab the mutex so that people don't use it
853 * before the caller fills it with data. Again, no one
854 * should have been able to get to this dude to lock it.
856 if (!lock_TryMutex(&bp->mx)) {
857 osi_Log2(afsd_logp, "buf_GetNewLocked bp 0x%p cannot be mutex locked. refCount %d should be 0",
859 osi_panic("buf_GetNewLocked: TryMutex failed",__FILE__,__LINE__);
862 /* prepare to return it. Give it a refcount */
865 lock_ReleaseWrite(&buf_globalLock);
869 buf_ValidateBufQueues();
872 } /* for all buffers in lru queue */
873 lock_ReleaseWrite(&buf_globalLock);
874 osi_Log0(afsd_logp, "buf_GetNewLocked: Free Buffer List has no buffers with a zero refcount - sleeping 100ms");
875 Sleep(100); /* give some time for a buffer to be freed */
876 } /* while loop over everything */
880 /* get a page, returning it held but unlocked. Doesn't fill in the page
881 * with I/O, since we're going to write the whole thing new.
883 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
887 osi_hyper_t pageOffset;
891 pageOffset.HighPart = offsetp->HighPart;
892 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
894 bp = buf_Find(scp, &pageOffset);
896 /* lock it and break out */
897 lock_ObtainMutex(&bp->mx);
901 /* otherwise, we have to create a page */
902 code = buf_GetNewLocked(scp, &pageOffset, &bp);
904 /* check if the buffer was created in a race condition branch.
905 * If so, go around so we can hold a reference to it.
907 if (code == CM_BUF_EXISTS)
910 /* something else went wrong */
914 /* otherwise, we have a locked buffer that we just created */
917 } /* big while loop */
920 if (bp->flags & CM_BUF_READING)
923 /* once it has been read once, we can unlock it and return it, still
924 * with its refcount held.
926 lock_ReleaseMutex(&bp->mx);
928 osi_Log4(buf_logp, "buf_GetNew returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
929 bp, scp, offsetp->HighPart, offsetp->LowPart);
933 /* get a page, returning it held but unlocked. Make sure it is complete */
934 /* The scp must be unlocked when passed to this function */
935 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
939 osi_hyper_t pageOffset;
940 unsigned long tcount;
945 #endif /* DISKCACHE95 */
948 pageOffset.HighPart = offsetp->HighPart;
949 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
953 buf_ValidateBufQueues();
956 bp = buf_Find(scp, &pageOffset);
958 /* lock it and break out */
959 lock_ObtainMutex(&bp->mx);
962 /* touch disk chunk to update LRU info */
963 diskcache_Touch(bp->dcp);
964 #endif /* DISKCACHE95 */
968 /* otherwise, we have to create a page */
969 code = buf_GetNewLocked(scp, &pageOffset, &bp);
970 /* bp->mx is now held */
972 /* check if the buffer was created in a race condition branch.
973 * If so, go around so we can hold a reference to it.
975 if (code == CM_BUF_EXISTS)
978 /* something else went wrong */
981 buf_ValidateBufQueues();
986 /* otherwise, we have a locked buffer that we just created */
989 } /* big while loop */
991 /* if we get here, we have a locked buffer that may have just been
992 * created, in which case it needs to be filled with data.
995 /* load the page; freshly created pages should be idle */
996 osi_assertx(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)), "incorrect cm_buf_t flags");
998 /* start the I/O; may drop lock */
999 bp->flags |= CM_BUF_READING;
1000 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
1003 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
1004 bp->dcp = dcp; /* pointer to disk cache struct. */
1005 #endif /* DISKCACHE95 */
1008 /* failure or queued */
1009 if (code != ERROR_IO_PENDING) {
1011 bp->flags |= CM_BUF_ERROR;
1012 bp->flags &= ~CM_BUF_READING;
1013 if (bp->flags & CM_BUF_WAITING) {
1014 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1015 osi_Wakeup((LONG_PTR) bp);
1017 lock_ReleaseMutex(&bp->mx);
1020 buf_ValidateBufQueues();
1021 #endif /* TESTING */
1025 /* otherwise, I/O completed instantly and we're done, except
1026 * for padding the xfr out with 0s and checking for EOF
1028 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1029 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1031 bp->flags |= CM_BUF_EOF;
1033 bp->flags &= ~CM_BUF_READING;
1034 if (bp->flags & CM_BUF_WAITING) {
1035 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1036 osi_Wakeup((LONG_PTR) bp);
1042 /* wait for reads, either that which we started above, or that someone
1043 * else started. We don't care if we return a buffer being cleaned.
1045 if (bp->flags & CM_BUF_READING)
1046 buf_WaitIO(scp, bp);
1048 /* once it has been read once, we can unlock it and return it, still
1049 * with its refcount held.
1051 lock_ReleaseMutex(&bp->mx);
1054 /* now remove from queue; will be put in at the head (farthest from
1055 * being recycled) when we're done in buf_Release.
1057 lock_ObtainWrite(&buf_globalLock);
1058 if (bp->flags & CM_BUF_INLRU) {
1059 if (cm_data.buf_freeListEndp == bp)
1060 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1061 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1062 bp->flags &= ~CM_BUF_INLRU;
1064 lock_ReleaseWrite(&buf_globalLock);
1066 osi_Log4(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x:%08x",
1067 bp, scp, offsetp->HighPart, offsetp->LowPart);
1069 buf_ValidateBufQueues();
1070 #endif /* TESTING */
1074 /* count # of elements in the free list;
1075 * we don't bother doing the proper locking for accessing dataVersion or flags
1076 * since it is a pain, and this is really just an advisory call. If you need
1077 * to do better at some point, rewrite this function.
1079 long buf_CountFreeList(void)
1085 lock_ObtainRead(&buf_globalLock);
1086 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1087 /* if the buffer doesn't have an identity, or if the buffer
1088 * has been invalidate (by having its DV stomped upon), then
1089 * count it as free, since it isn't really being utilized.
1091 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion <= 0)
1094 lock_ReleaseRead(&buf_globalLock);
1098 /* clean a buffer synchronously */
1099 long buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
1102 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1104 lock_ObtainMutex(&bp->mx);
1105 code = buf_CleanAsyncLocked(bp, reqp);
1106 lock_ReleaseMutex(&bp->mx);
1111 /* wait for a buffer's cleaning to finish */
1112 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp)
1114 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1116 lock_ObtainMutex(&bp->mx);
1117 if (bp->flags & CM_BUF_WRITING) {
1118 buf_WaitIO(scp, bp);
1120 lock_ReleaseMutex(&bp->mx);
1123 /* set the dirty flag on a buffer, and set associated write-ahead log,
1124 * if there is one. Allow one to be added to a buffer, but not changed.
1126 * The buffer must be locked before calling this routine.
1128 void buf_SetDirty(cm_buf_t *bp, afs_uint32 offset, afs_uint32 length)
1130 osi_assertx(bp->magic == CM_BUF_MAGIC, "invalid cm_buf_t magic");
1131 osi_assertx(bp->refCount > 0, "cm_buf_t refcount 0");
1133 lock_ObtainWrite(&buf_globalLock);
1134 if (bp->flags & CM_BUF_DIRTY) {
1136 osi_Log1(buf_logp, "buf_SetDirty 0x%p already dirty", bp);
1138 if (bp->dirty_offset <= offset) {
1139 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1140 /* dirty_length remains the same */
1142 bp->dirty_length = offset + length - bp->dirty_offset;
1144 } else /* bp->dirty_offset > offset */ {
1145 if (bp->dirty_offset + bp->dirty_length >= offset + length) {
1146 bp->dirty_length = bp->dirty_offset + bp->dirty_length - offset;
1148 bp->dirty_length = length;
1150 bp->dirty_offset = offset;
1153 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1156 bp->flags |= CM_BUF_DIRTY;
1158 /* and turn off EOF flag, since it has associated data now */
1159 bp->flags &= ~CM_BUF_EOF;
1161 bp->dirty_offset = offset;
1162 bp->dirty_length = length;
1164 /* and add to the dirty list.
1165 * we obtain a hold on the buffer for as long as it remains
1166 * in the list. buffers are only removed from the list by
1167 * the buf_IncrSyncer function regardless of when else the
1168 * dirty flag might be cleared.
1170 * This should never happen but just in case there is a bug
1171 * elsewhere, never add to the dirty list if the buffer is
1174 if (bp->dirtyp == NULL && cm_data.buf_dirtyListEndp != bp) {
1176 if (!cm_data.buf_dirtyListp) {
1177 cm_data.buf_dirtyListp = cm_data.buf_dirtyListEndp = bp;
1179 cm_data.buf_dirtyListEndp->dirtyp = bp;
1180 cm_data.buf_dirtyListEndp = bp;
1185 lock_ReleaseWrite(&buf_globalLock);
1188 /* clean all buffers, reset log pointers and invalidate all buffers.
1189 * Called with no locks held, and returns with same.
1191 * This function is guaranteed to clean and remove the log ptr of all the
1192 * buffers that were dirty or had non-zero log ptrs before the call was
1193 * made. That's sufficient to clean up any garbage left around by recovery,
1194 * which is all we're counting on this for; there may be newly created buffers
1195 * added while we're running, but that should be OK.
1197 * In an environment where there are no transactions (artificially imposed, for
1198 * example, when switching the database to raw mode), this function is used to
1199 * make sure that all updates have been written to the disk. In that case, we don't
1200 * really require that we forget the log association between pages and logs, but
1201 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1202 * have to worry about invalidating data in the buffers.
1204 * This function is used at the end of recovery as paranoia to get the recovered
1205 * database out to disk. It removes all references to the recovery log and cleans
1208 long buf_CleanAndReset(void)
1214 lock_ObtainWrite(&buf_globalLock);
1215 for(i=0; i<cm_data.buf_hashSize; i++) {
1216 for(bp = cm_data.buf_scacheHashTablepp[i]; bp; bp = bp->hashp) {
1217 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1219 lock_ReleaseWrite(&buf_globalLock);
1221 /* now no locks are held; clean buffer and go on */
1223 req.flags |= CM_REQ_NORETRY;
1225 buf_CleanAsync(bp, &req);
1226 buf_CleanWait(NULL, bp);
1228 /* relock and release buffer */
1229 lock_ObtainWrite(&buf_globalLock);
1230 buf_ReleaseLocked(bp);
1232 } /* over one bucket */
1233 } /* for loop over all hash buckets */
1236 lock_ReleaseWrite(&buf_globalLock);
1239 buf_ValidateBufQueues();
1240 #endif /* TESTING */
1242 /* and we're done */
1246 /* called without global lock being held, reserves buffers for callers
1247 * that need more than one held (not locked) at once.
1249 void buf_ReserveBuffers(afs_uint64 nbuffers)
1251 lock_ObtainWrite(&buf_globalLock);
1253 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1254 cm_data.buf_reserveWaiting = 1;
1255 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1256 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1257 lock_ObtainWrite(&buf_globalLock);
1260 cm_data.buf_reservedBufs += nbuffers;
1264 lock_ReleaseWrite(&buf_globalLock);
1267 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1271 lock_ObtainWrite(&buf_globalLock);
1272 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1276 cm_data.buf_reservedBufs += nbuffers;
1279 lock_ReleaseWrite(&buf_globalLock);
1283 /* called without global lock held, releases reservation held by
1284 * buf_ReserveBuffers.
1286 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1288 lock_ObtainWrite(&buf_globalLock);
1289 cm_data.buf_reservedBufs -= nbuffers;
1290 if (cm_data.buf_reserveWaiting) {
1291 cm_data.buf_reserveWaiting = 0;
1292 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1294 lock_ReleaseWrite(&buf_globalLock);
1297 /* truncate the buffers past sizep, zeroing out the page, if we don't
1298 * end on a page boundary.
1300 * Requires cm_bufCreateLock to be write locked.
1302 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1306 cm_buf_t *nbufp; /* next buffer, if didRelease */
1312 /* assert that cm_bufCreateLock is held in write mode */
1313 lock_AssertWrite(&scp->bufCreateLock);
1315 i = BUF_FILEHASH(&scp->fid);
1317 lock_ObtainWrite(&buf_globalLock);
1318 bufp = cm_data.buf_fileHashTablepp[i];
1320 lock_ReleaseWrite(&buf_globalLock);
1324 buf_HoldLocked(bufp);
1325 lock_ReleaseWrite(&buf_globalLock);
1327 lock_ObtainMutex(&bufp->mx);
1329 bufEnd.HighPart = 0;
1330 bufEnd.LowPart = cm_data.buf_blockSize;
1331 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1333 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1334 LargeIntegerLessThan(*sizep, bufEnd)) {
1335 buf_WaitIO(scp, bufp);
1337 lock_ObtainMutex(&scp->mx);
1339 /* make sure we have a callback (so we have the right value for
1340 * the length), and wait for it to be safe to do a truncate.
1342 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1343 CM_SCACHESYNC_NEEDCALLBACK
1344 | CM_SCACHESYNC_GETSTATUS
1345 | CM_SCACHESYNC_SETSIZE
1346 | CM_SCACHESYNC_BUFLOCKED);
1349 lock_ObtainWrite(&buf_globalLock);
1350 /* if we succeeded in our locking, and this applies to the right
1351 * file, and the truncate request overlaps the buffer either
1352 * totally or partially, then do something.
1354 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1355 && LargeIntegerLessThan(*sizep, bufEnd)) {
1358 /* destroy the buffer, turning off its dirty bit, if
1359 * we're truncating the whole buffer. Otherwise, set
1360 * the dirty bit, and clear out the tail of the buffer
1361 * if we just overlap some.
1363 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1364 /* truncating the entire page */
1365 bufp->flags &= ~CM_BUF_DIRTY;
1366 bufp->dirty_offset = 0;
1367 bufp->dirty_length = 0;
1368 bufp->dataVersion = -1; /* known bad */
1369 bufp->dirtyCounter++;
1372 /* don't set dirty, since dirty implies
1373 * currently up-to-date. Don't need to do this,
1374 * since we'll update the length anyway.
1376 * Zero out remainder of the page, in case we
1377 * seek and write past EOF, and make this data
1380 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1381 osi_assertx(bufferPos != 0, "non-zero bufferPos");
1382 memset(bufp->datap + bufferPos, 0,
1383 cm_data.buf_blockSize - bufferPos);
1387 cm_SyncOpDone( scp, bufp,
1388 CM_SCACHESYNC_NEEDCALLBACK | CM_SCACHESYNC_GETSTATUS
1389 | CM_SCACHESYNC_SETSIZE | CM_SCACHESYNC_BUFLOCKED);
1391 lock_ReleaseMutex(&scp->mx);
1392 lock_ReleaseMutex(&bufp->mx);
1395 nbufp = bufp->fileHashp;
1397 buf_HoldLocked(nbufp);
1399 /* This forces the loop to end and the error code
1400 * to be returned. */
1403 buf_ReleaseLocked(bufp);
1404 lock_ReleaseWrite(&buf_globalLock);
1409 buf_ValidateBufQueues();
1410 #endif /* TESTING */
1416 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1419 cm_buf_t *bp; /* buffer we're hacking on */
1424 i = BUF_FILEHASH(&scp->fid);
1427 lock_ObtainWrite(&buf_globalLock);
1428 bp = cm_data.buf_fileHashTablepp[i];
1431 lock_ReleaseWrite(&buf_globalLock);
1433 for (; bp; bp = nbp) {
1434 didRelease = 0; /* haven't released this buffer yet */
1436 /* clean buffer synchronously */
1437 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1438 lock_ObtainMutex(&bp->mx);
1440 /* start cleaning the buffer, and wait for it to finish */
1441 buf_CleanAsyncLocked(bp, reqp);
1442 buf_WaitIO(scp, bp);
1443 lock_ReleaseMutex(&bp->mx);
1445 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1446 if (code && code != CM_ERROR_BADFD)
1449 /* if the scp's FID is bad its because we received VNOVNODE
1450 * when attempting to FetchStatus before the write. This
1451 * page therefore contains data that can no longer be stored.
1453 lock_ObtainMutex(&bp->mx);
1454 bp->flags &= ~CM_BUF_DIRTY;
1455 bp->flags |= CM_BUF_ERROR;
1457 bp->dirty_offset = 0;
1458 bp->dirty_length = 0;
1459 bp->dataVersion = -1; /* known bad */
1461 lock_ReleaseMutex(&bp->mx);
1463 lock_ObtainWrite(&buf_globalLock);
1464 /* actually, we only know that buffer is clean if ref
1465 * count is 1, since we don't have buffer itself locked.
1467 if (!(bp->flags & CM_BUF_DIRTY)) {
1468 if (bp->refCount == 1) { /* bp is held above */
1469 nbp = bp->fileHashp;
1471 buf_HoldLocked(nbp);
1472 buf_ReleaseLocked(bp);
1477 lock_ReleaseWrite(&buf_globalLock);
1479 if (code != CM_ERROR_BADFD)
1480 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1485 lock_ObtainWrite(&buf_globalLock);
1486 nbp = bp->fileHashp;
1488 buf_HoldLocked(nbp);
1489 buf_ReleaseLocked(bp);
1490 lock_ReleaseWrite(&buf_globalLock);
1492 } /* for loop over a bunch of buffers */
1495 buf_ValidateBufQueues();
1496 #endif /* TESTING */
1502 /* Must be called with scp->mx held */
1503 long buf_ForceDataVersion(cm_scache_t * scp, afs_uint32 fromVersion, afs_uint32 toVersion)
1510 lock_AssertMutex(&scp->mx);
1512 i = BUF_FILEHASH(&scp->fid);
1514 lock_ObtainWrite(&buf_globalLock);
1516 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp = bp->fileHashp) {
1517 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1518 if (bp->dataVersion == fromVersion) {
1519 bp->dataVersion = toVersion;
1524 lock_ReleaseWrite(&buf_globalLock);
1532 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1536 cm_buf_t *bp; /* buffer we're hacking on */
1537 cm_buf_t *nbp; /* next one */
1540 i = BUF_FILEHASH(&scp->fid);
1542 lock_ObtainWrite(&buf_globalLock);
1543 bp = cm_data.buf_fileHashTablepp[i];
1546 lock_ReleaseWrite(&buf_globalLock);
1547 for (; bp; bp = nbp) {
1548 /* clean buffer synchronously */
1549 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1552 lock_ObtainMutex(&bp->mx);
1554 cm_ReleaseUser(bp->userp);
1556 lock_ReleaseMutex(&bp->mx);
1558 wasDirty = buf_CleanAsync(bp, reqp);
1559 buf_CleanWait(scp, bp);
1560 lock_ObtainMutex(&bp->mx);
1561 if (bp->flags & CM_BUF_ERROR) {
1566 lock_ReleaseMutex(&bp->mx);
1569 lock_ObtainWrite(&buf_globalLock);
1570 nbp = bp->fileHashp;
1572 buf_HoldLocked(nbp);
1573 buf_ReleaseLocked(bp);
1574 lock_ReleaseWrite(&buf_globalLock);
1575 } /* for loop over a bunch of buffers */
1578 buf_ValidateBufQueues();
1579 #endif /* TESTING */
1587 buf_ValidateBufQueues(void)
1589 cm_buf_t * bp, *bpb, *bpf, *bpa;
1590 afs_uint32 countf=0, countb=0, counta=0;
1592 lock_ObtainRead(&buf_globalLock);
1593 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1594 if (bp->magic != CM_BUF_MAGIC)
1595 osi_panic("buf magic error",__FILE__,__LINE__);
1600 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1601 if (bp->magic != CM_BUF_MAGIC)
1602 osi_panic("buf magic error",__FILE__,__LINE__);
1607 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1608 if (bp->magic != CM_BUF_MAGIC)
1609 osi_panic("buf magic error",__FILE__,__LINE__);
1613 lock_ReleaseRead(&buf_globalLock);
1615 if (countb != countf)
1616 osi_panic("buf magic error",__FILE__,__LINE__);
1618 if (counta != cm_data.buf_nbuffers)
1619 osi_panic("buf magic error",__FILE__,__LINE__);
1621 #endif /* TESTING */
1623 /* dump the contents of the buf_scacheHashTablepp. */
1624 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1631 if (cm_data.buf_scacheHashTablepp == NULL)
1635 lock_ObtainRead(&buf_globalLock);
1637 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\r\n",
1638 cookie, cm_data.buf_hashSize);
1639 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1641 for (i = 0; i < cm_data.buf_hashSize; i++)
1643 for (bp = cm_data.buf_scacheHashTablepp[i]; bp; bp=bp->hashp)
1645 StringCbPrintfA(output, sizeof(output),
1646 "%s bp=0x%08X, hash=%d, fid (cell=%d, volume=%d, "
1647 "vnode=%d, unique=%d), offset=%x:%08x, dv=%d, "
1648 "flags=0x%x, cmFlags=0x%x, refCount=%d\r\n",
1649 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1650 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1651 bp->offset.LowPart, bp->dataVersion, bp->flags,
1652 bp->cmFlags, bp->refCount);
1653 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1657 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\r\n", cookie);
1658 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1660 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_freeListEndp\r\n", cookie);
1661 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1662 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1663 StringCbPrintfA(output, sizeof(output),
1664 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1665 "vnode=%d, unique=%d), offset=%x:%08x, dv=%d, "
1666 "flags=0x%x, cmFlags=0x%x, refCount=%d\r\n",
1667 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1668 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1669 bp->offset.LowPart, bp->dataVersion, bp->flags,
1670 bp->cmFlags, bp->refCount);
1671 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1673 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_FreeListEndp.\r\n", cookie);
1674 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1676 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_dirtyListEndp\r\n", cookie);
1677 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1678 for(bp = cm_data.buf_dirtyListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1679 StringCbPrintfA(output, sizeof(output),
1680 "%s bp=0x%08X, fid (cell=%d, volume=%d, "
1681 "vnode=%d, unique=%d), offset=%x:%08x, dv=%d, "
1682 "flags=0x%x, cmFlags=0x%x, refCount=%d\r\n",
1683 cookie, (void *)bp, bp->fid.cell, bp->fid.volume,
1684 bp->fid.vnode, bp->fid.unique, bp->offset.HighPart,
1685 bp->offset.LowPart, bp->dataVersion, bp->flags,
1686 bp->cmFlags, bp->refCount);
1687 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1689 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_dirtyListEndp.\r\n", cookie);
1690 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1693 lock_ReleaseRead(&buf_globalLock);
1697 void buf_ForceTrace(BOOL flush)
1706 len = GetTempPath(sizeof(buf)-10, buf);
1707 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1708 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1709 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1710 if (handle == INVALID_HANDLE_VALUE) {
1711 osi_panic("Cannot create log file", __FILE__, __LINE__);
1713 osi_LogPrint(buf_logp, handle);
1715 FlushFileBuffers(handle);
1716 CloseHandle(handle);
1719 long buf_DirtyBuffersExist(cm_fid_t *fidp)
1722 afs_uint32 bcount = 0;
1725 i = BUF_FILEHASH(fidp);
1727 for (bp = cm_data.buf_fileHashTablepp[i]; bp; bp=bp->allp, bcount++) {
1728 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY))
1735 long buf_CleanDirtyBuffers(cm_scache_t *scp)
1738 afs_uint32 bcount = 0;
1739 cm_fid_t * fidp = &scp->fid;
1741 for (bp = cm_data.buf_allp; bp; bp=bp->allp, bcount++) {
1742 if (!cm_FidCmp(fidp, &bp->fid) && (bp->flags & CM_BUF_DIRTY)) {
1744 lock_ObtainMutex(&bp->mx);
1745 bp->cmFlags &= ~CM_BUF_CMSTORING;
1746 bp->flags &= ~CM_BUF_DIRTY;
1747 bp->dirty_offset = 0;
1748 bp->dirty_length = 0;
1749 bp->flags |= CM_BUF_ERROR;
1750 bp->error = VNOVNODE;
1751 bp->dataVersion = -1; /* bad */
1753 if (bp->flags & CM_BUF_WAITING) {
1754 osi_Log2(buf_logp, "BUF CleanDirtyBuffers Waking [scp 0x%x] bp 0x%x", scp, bp);
1755 osi_Wakeup((long) &bp);
1757 lock_ReleaseMutex(&bp->mx);