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>
24 #include "cm_memmap.h"
27 #define TRACE_BUFFER 1
30 extern void afsi_log(char *pattern, ...);
32 /* This module implements the buffer package used by the local transaction
33 * system (cm). It is initialized by calling cm_Init, which calls buf_Init;
34 * it must be initalized before any of its main routines are called.
36 * Each buffer is hashed into a hash table by file ID and offset, and if its
37 * reference count is zero, it is also in a free list.
39 * There are two locks involved in buffer processing. The global lock
40 * buf_globalLock protects all of the global variables defined in this module,
41 * the reference counts and hash pointers in the actual cm_buf_t structures,
42 * and the LRU queue pointers in the buffer structures.
44 * The mutexes in the buffer structures protect the remaining fields in the
45 * buffers, as well the data itself.
47 * The locking hierarchy here is this:
49 * - resv multiple simul. buffers reservation
50 * - lock buffer I/O flags
51 * - lock buffer's mutex
52 * - lock buf_globalLock
56 /* global debugging log */
57 osi_log_t *buf_logp = NULL;
59 /* Global lock protecting hash tables and free lists */
60 osi_rwlock_t buf_globalLock;
62 /* ptr to head of the free list (most recently used) and the
63 * tail (the guy to remove first). We use osi_Q* functions
64 * to put stuff in buf_freeListp, and maintain the end
68 /* a pointer to a list of all buffers, just so that we can find them
69 * easily for debugging, and for the incr syncer. Locked under
73 /* defaults setup; these variables may be manually assigned into
74 * before calling cm_Init, as a way of changing these defaults.
77 /* callouts for reading and writing data, etc */
78 cm_buf_ops_t *cm_buf_opsp;
81 /* for experimental disk caching support in Win95 client */
82 cm_buf_t *buf_diskFreeListp;
83 cm_buf_t *buf_diskFreeListEndp;
84 cm_buf_t *buf_diskAllp;
85 extern int cm_diskCacheEnabled;
86 #endif /* DISKCACHE95 */
88 /* set this to 1 when we are terminating to prevent access attempts */
89 static int buf_ShutdownFlag = 0;
91 /* hold a reference to an already held buffer */
92 void buf_Hold(cm_buf_t *bp)
94 osi_assert(bp->magic == CM_BUF_MAGIC);
95 lock_ObtainWrite(&buf_globalLock);
97 lock_ReleaseWrite(&buf_globalLock);
100 /* incremental sync daemon. Writes 1/10th of all the buffers every 5000 ms */
101 void buf_IncrSyncer(long parm)
103 cm_buf_t *bp; /* buffer we're hacking on; held */
104 long i; /* counter */
105 long nAtOnce; /* how many to do at once */
108 lock_ObtainWrite(&buf_globalLock);
109 bp = cm_data.buf_allp;
111 lock_ReleaseWrite(&buf_globalLock);
112 nAtOnce = cm_data.buf_nbuffers / 10;
113 while (buf_ShutdownFlag == 0) {
115 i = SleepEx(5000, 1);
116 if (i != 0) continue;
121 if (buf_ShutdownFlag == 1)
124 /* now go through our percentage of the buffers */
125 for (i=0; i<nAtOnce; i++) {
126 /* don't want its identity changing while we're
127 * messing with it, so must do all of this with
131 /* start cleaning the buffer; don't touch log pages since
132 * the log code counts on knowing exactly who is writing
133 * a log page at any given instant.
136 req.flags |= CM_REQ_NORETRY;
137 buf_CleanAsync(bp, &req);
139 /* now advance to the next buffer; the allp chain never changes,
140 * and so can be followed even when holding no locks.
142 lock_ObtainWrite(&buf_globalLock);
143 buf_LockedRelease(bp);
146 bp = cm_data.buf_allp;
148 lock_ReleaseWrite(&buf_globalLock);
149 } /* for loop over a bunch of buffers */
150 } /* whole daemon's while loop */
154 buf_ValidateBuffers(void)
156 cm_buf_t * bp, *bpf, *bpa, *bpb;
157 afs_uint32 countb = 0, countf = 0, counta = 0;
159 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
160 if (bp->magic != CM_BUF_MAGIC) {
161 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
162 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
168 if (countb > cm_data.buf_nbuffers) {
169 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
170 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
175 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
176 if (bp->magic != CM_BUF_MAGIC) {
177 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
178 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
184 if (countf > cm_data.buf_nbuffers) {
185 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
186 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
191 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
192 if (bp->magic != CM_BUF_MAGIC) {
193 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
194 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
200 if (counta > cm_data.buf_nbuffers) {
201 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
202 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
207 if (countb != countf) {
208 afsi_log("cm_ValidateBuffers failure: countb != countf");
209 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
213 if (counta != cm_data.buf_nbuffers) {
214 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
215 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
222 void buf_Shutdown(void)
224 buf_ShutdownFlag = 1;
227 /* initialize the buffer package; called with no locks
228 * held during the initialization phase.
230 long buf_Init(int newFile, cm_buf_ops_t *opsp, long nbuffers)
232 static osi_once_t once;
241 cm_data.buf_nbuffers = nbuffers;
243 /* Have to be able to reserve a whole chunk */
244 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
245 return CM_ERROR_TOOFEWBUFS;
248 /* recall for callouts */
251 if (osi_Once(&once)) {
252 /* initialize global locks */
253 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock");
256 /* remember this for those who want to reset it */
257 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
259 /* lower hash size to a prime number */
260 cm_data.buf_hashSize = osi_PrimeLessThan(CM_BUF_HASHSIZE);
262 /* create hash table */
263 memset((void *)cm_data.buf_hashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
265 /* another hash table */
266 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
268 /* create buffer headers and put in free list */
269 bp = cm_data.bufHeaderBaseAddress;
270 data = cm_data.bufDataBaseAddress;
271 cm_data.buf_allp = NULL;
273 for (i=0; i<cm_data.buf_nbuffers; i++) {
274 osi_assert(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress);
275 osi_assert(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData);
277 /* allocate and zero some storage */
278 memset(bp, 0, sizeof(cm_buf_t));
279 bp->magic = CM_BUF_MAGIC;
280 /* thread on list of all buffers */
281 bp->allp = cm_data.buf_allp;
282 cm_data.buf_allp = bp;
284 osi_QAdd((osi_queue_t **)&cm_data.buf_freeListp, &bp->q);
285 bp->flags |= CM_BUF_INLRU;
286 lock_InitializeMutex(&bp->mx, "Buffer mutex");
288 /* grab appropriate number of bytes from aligned zone */
291 /* setup last buffer pointer */
293 cm_data.buf_freeListEndp = bp;
297 data += cm_data.buf_blockSize;
300 /* none reserved at first */
301 cm_data.buf_reservedBufs = 0;
303 /* just for safety's sake */
304 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
306 bp = cm_data.bufHeaderBaseAddress;
307 data = cm_data.bufDataBaseAddress;
309 for (i=0; i<cm_data.buf_nbuffers; i++) {
310 lock_InitializeMutex(&bp->mx, "Buffer mutex");
317 buf_ValidateBufQueues();
321 /* init the buffer trace log */
322 buf_logp = osi_LogCreate("buffer", 1000);
323 osi_LogEnable(buf_logp);
328 /* and create the incr-syncer */
329 phandle = thrd_Create(0, 0,
330 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
333 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
335 CloseHandle(phandle);
340 buf_ValidateBufQueues();
345 /* add nbuffers to the buffer pool, if possible.
346 * Called with no locks held.
348 long buf_AddBuffers(long nbuffers)
351 /* The size of a virtual cache cannot be changed after it has
352 * been created. Subsequent calls to MapViewofFile() with
353 * an existing mapping object name would not allow the
354 * object to be resized. Return failure immediately.
356 * A similar problem now occurs with the persistent cache
357 * given that the memory mapped file now contains a complex
360 afsi_log("request to add %d buffers to the existing cache of size %d denied",
361 nbuffers, cm_data.buf_nbuffers);
363 return CM_ERROR_INVAL;
369 data = malloc(buf_nbuffers * cm_data.buf_blockSize);
371 /* Create buffer headers and put in free list */
372 bp = malloc(nbuffers * sizeof(*bp));
374 for (i=0; i<nbuffers; i++) {
375 memset(bp, 0, sizeof(*bp));
377 lock_InitializeMutex(&bp->mx, "cm_buf_t");
379 /* grab appropriate number of bytes from aligned zone */
382 bp->flags |= CM_BUF_INLRU;
384 lock_ObtainWrite(&buf_globalLock);
385 /* note that buf_allp chain is covered by buf_globalLock now */
386 bp->allp = cm_data.buf_allp;
387 cm_data.buf_allp = bp;
388 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
389 if (!cm_data.buf_freeListEndp)
390 cm_data.buf_freeListEndp = bp;
391 cm_data.buf_nbuffers++;
392 lock_ReleaseWrite(&buf_globalLock);
395 data += cm_data.buf_blockSize;
397 } /* for loop over all buffers */
403 /* interface to set the number of buffers to an exact figure.
404 * Called with no locks held.
406 long buf_SetNBuffers(long nbuffers)
409 return CM_ERROR_INVAL;
410 if (nbuffers == cm_data.buf_nbuffers)
412 else if (nbuffers > cm_data.buf_nbuffers)
413 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
415 return CM_ERROR_INVAL;
418 /* release a buffer. Buffer must be referenced, but unlocked. */
419 void buf_Release(cm_buf_t *bp)
421 lock_ObtainWrite(&buf_globalLock);
422 buf_LockedRelease(bp);
423 lock_ReleaseWrite(&buf_globalLock);
426 /* wait for reading or writing to clear; called with write-locked
427 * buffer, and returns with locked buffer.
429 void buf_WaitIO(cm_buf_t *bp)
431 osi_assert(bp->magic == CM_BUF_MAGIC);
434 /* if no IO is happening, we're done */
435 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
438 /* otherwise I/O is happening, but some other thread is waiting for
439 * the I/O already. Wait for that guy to figure out what happened,
440 * and then check again.
442 if ( bp->flags & CM_BUF_WAITING )
443 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%x", bp);
445 bp->flags |= CM_BUF_WAITING;
446 osi_SleepM((long) bp, &bp->mx);
447 lock_ObtainMutex(&bp->mx);
448 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%x", bp);
451 /* if we get here, the IO is done, but we may have to wakeup people waiting for
452 * the I/O to complete. Do so.
454 if (bp->flags & CM_BUF_WAITING) {
455 bp->flags &= ~CM_BUF_WAITING;
456 osi_Wakeup((long) bp);
458 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%x", (long) bp);
461 /* code to drop reference count while holding buf_globalLock */
462 void buf_LockedRelease(cm_buf_t *bp)
464 /* ensure that we're in the LRU queue if our ref count is 0 */
465 osi_assert(bp->refCount > 0);
466 if (--bp->refCount == 0) {
467 if (!(bp->flags & CM_BUF_INLRU)) {
468 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
470 /* watch for transition from empty to one element */
471 if (!cm_data.buf_freeListEndp)
472 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
473 bp->flags |= CM_BUF_INLRU;
478 /* find a buffer, if any, for a particular file ID and offset. Assumes
479 * that buf_globalLock is write locked when called.
481 cm_buf_t *buf_LockedFind(struct cm_scache *scp, osi_hyper_t *offsetp)
486 i = BUF_HASH(&scp->fid, offsetp);
487 for(bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp) {
488 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
489 && offsetp->LowPart == bp->offset.LowPart
490 && offsetp->HighPart == bp->offset.HighPart) {
496 /* return whatever we found, if anything */
500 /* find a buffer with offset *offsetp for vnode *scp. Called
501 * with no locks held.
503 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
507 lock_ObtainWrite(&buf_globalLock);
508 bp = buf_LockedFind(scp, offsetp);
509 lock_ReleaseWrite(&buf_globalLock);
514 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
517 * Makes sure that there's only one person writing this block
518 * at any given time, and also ensures that the log is forced sufficiently far,
519 * if this buffer contains logged data.
521 void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
525 osi_assert(bp->magic == CM_BUF_MAGIC);
527 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
528 lock_ReleaseMutex(&bp->mx);
530 code = (*cm_buf_opsp->Writep)(&bp->fid, &bp->offset,
531 cm_data.buf_blockSize, 0, bp->userp,
534 lock_ObtainMutex(&bp->mx);
539 /* Disk cache support */
540 /* write buffer to disk cache (synchronous for now) */
541 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
542 #endif /* DISKCACHE95 */
545 /* do logging after call to GetLastError, or else */
546 osi_Log2(buf_logp, "buf_CleanAsync starts I/O on 0x%x, done=%d", bp, code);
548 /* if someone was waiting for the I/O that just completed or failed,
551 if (bp->flags & CM_BUF_WAITING) {
552 /* turn off flags and wakeup users */
553 bp->flags &= ~CM_BUF_WAITING;
554 osi_Wakeup((long) bp);
558 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
559 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
560 * The buffer must already be clean, and no I/O should be happening to it.
562 void buf_Recycle(cm_buf_t *bp)
567 cm_buf_t *prevBp, *nextBp;
569 osi_assert(bp->magic == CM_BUF_MAGIC);
571 /* if we get here, we know that the buffer still has a 0 ref count,
572 * and that it is clean and has no currently pending I/O. This is
573 * the dude to return.
574 * Remember that as long as the ref count is 0, we know that we won't
575 * have any lock conflicts, so we can grab the buffer lock out of
576 * order in the locking hierarchy.
578 osi_Log2( buf_logp, "buf_Recycle recycles 0x%x, off 0x%x",
579 bp, bp->offset.LowPart);
581 osi_assert(bp->refCount == 0);
582 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)));
583 lock_AssertWrite(&buf_globalLock);
585 if (bp->flags & CM_BUF_INHASH) {
586 /* Remove from hash */
588 i = BUF_HASH(&bp->fid, &bp->offset);
589 lbpp = &(cm_data.buf_hashTablepp[i]);
590 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
591 if (tbp == bp) break;
594 /* we better find it */
595 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
597 *lbpp = bp->hashp; /* hash out */
599 /* Remove from file hash */
601 i = BUF_FILEHASH(&bp->fid);
602 prevBp = bp->fileHashBackp;
603 nextBp = bp->fileHashp;
605 prevBp->fileHashp = nextBp;
607 cm_data.buf_fileHashTablepp[i] = nextBp;
609 nextBp->fileHashBackp = prevBp;
611 bp->flags &= ~CM_BUF_INHASH;
614 /* bump the soft reference counter now, to invalidate softRefs; no
615 * wakeup is required since people don't sleep waiting for this
620 /* make the fid unrecognizable */
621 memset(&bp->fid, 0, sizeof(cm_fid_t));
624 /* recycle a buffer, removing it from the free list, hashing in its new identity
625 * and returning it write-locked so that no one can use it. Called without
626 * any locks held, and can return an error if it loses the race condition and
627 * finds that someone else created the desired buffer.
629 * If success is returned, the buffer is returned write-locked.
631 * May be called with null scp and offsetp, if we're just trying to reclaim some
632 * space from the buffer pool. In that case, the buffer will be returned
633 * without being hashed into the hash table.
635 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
637 cm_buf_t *bp; /* buffer we're dealing with */
638 cm_buf_t *nextBp; /* next buffer in file hash chain */
642 cm_InitReq(&req); /* just in case */
645 buf_ValidateBufQueues();
650 lock_ObtainWrite(&buf_globalLock);
651 /* check to see if we lost the race */
653 if (bp = buf_LockedFind(scp, offsetp)) {
655 lock_ReleaseWrite(&buf_globalLock);
656 return CM_BUF_EXISTS;
660 /* for debugging, assert free list isn't empty, although we
661 * really should try waiting for a running tranasction to finish
662 * instead of this; or better, we should have a transaction
663 * throttler prevent us from entering this situation.
665 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
667 /* look at all buffers in free list, some of which may temp.
668 * have high refcounts and which then should be skipped,
669 * starting cleaning I/O for those which are dirty. If we find
670 * a clean buffer, we rehash it, lock it and return it.
672 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
673 /* check to see if it really has zero ref count. This
674 * code can bump refcounts, at least, so it may not be
677 if (bp->refCount > 0)
680 /* we don't have to lock buffer itself, since the ref
681 * count is 0 and we know it will stay zero as long as
682 * we hold the global lock.
685 /* don't recycle someone in our own chunk */
686 if (!cm_FidCmp(&bp->fid, &scp->fid)
687 && (bp->offset.LowPart & (-cm_chunkSize))
688 == (offsetp->LowPart & (-cm_chunkSize)))
691 /* if this page is being filled (!) or cleaned, see if
692 * the I/O has completed. If not, skip it, otherwise
693 * do the final processing for the I/O.
695 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
696 /* probably shouldn't do this much work while
697 * holding the big lock? Watch for contention
703 if (bp->flags & CM_BUF_DIRTY) {
704 /* if the buffer is dirty, start cleaning it and
705 * move on to the next buffer. We do this with
706 * just the lock required to minimize contention
710 lock_ReleaseWrite(&buf_globalLock);
712 /* grab required lock and clean; this only
713 * starts the I/O. By the time we're back,
714 * it'll still be marked dirty, but it will also
715 * have the WRITING flag set, so we won't get
718 buf_CleanAsync(bp, &req);
720 /* now put it back and go around again */
725 /* if we get here, we know that the buffer still has a 0
726 * ref count, and that it is clean and has no currently
727 * pending I/O. This is the dude to return.
728 * Remember that as long as the ref count is 0, we know
729 * that we won't have any lock conflicts, so we can grab
730 * the buffer lock out of order in the locking hierarchy.
734 /* clean up junk flags */
735 bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
736 bp->dataVersion = -1; /* unknown so far */
738 /* now hash in as our new buffer, and give it the
739 * appropriate label, if requested.
742 bp->flags |= CM_BUF_INHASH;
744 bp->offset = *offsetp;
745 i = BUF_HASH(&scp->fid, offsetp);
746 bp->hashp = cm_data.buf_hashTablepp[i];
747 cm_data.buf_hashTablepp[i] = bp;
748 i = BUF_FILEHASH(&scp->fid);
749 nextBp = cm_data.buf_fileHashTablepp[i];
750 bp->fileHashp = nextBp;
751 bp->fileHashBackp = NULL;
753 nextBp->fileHashBackp = bp;
754 cm_data.buf_fileHashTablepp[i] = bp;
757 /* prepare to return it. Start by giving it a good
761 /* and since it has a non-zero ref count, we should move
762 * it from the lru queue. It better be still there,
763 * since we've held the global (big) lock since we found
766 osi_assertx(bp->flags & CM_BUF_INLRU,
767 "buf_GetNewLocked: LRU screwup");
768 if (cm_data.buf_freeListEndp == bp) {
769 /* we're the last guy in this queue, so maintain it */
770 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
772 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
773 bp->flags &= ~CM_BUF_INLRU;
775 /* finally, grab the mutex so that people don't use it
776 * before the caller fills it with data. Again, no one
777 * should have been able to get to this dude to lock it.
779 osi_assertx(lock_TryMutex(&bp->mx),
780 "buf_GetNewLocked: TryMutex failed");
782 lock_ReleaseWrite(&buf_globalLock);
786 buf_ValidateBufQueues();
789 } /* for all buffers in lru queue */
790 lock_ReleaseWrite(&buf_globalLock);
791 } /* while loop over everything */
795 /* get a page, returning it held but unlocked. Doesn't fill in the page
796 * with I/O, since we're going to write the whole thing new.
798 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
802 osi_hyper_t pageOffset;
806 pageOffset.HighPart = offsetp->HighPart;
807 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
809 bp = buf_Find(scp, &pageOffset);
811 /* lock it and break out */
812 lock_ObtainMutex(&bp->mx);
816 /* otherwise, we have to create a page */
817 code = buf_GetNewLocked(scp, &pageOffset, &bp);
819 /* check if the buffer was created in a race condition branch.
820 * If so, go around so we can hold a reference to it.
822 if (code == CM_BUF_EXISTS)
825 /* something else went wrong */
829 /* otherwise, we have a locked buffer that we just created */
832 } /* big while loop */
835 if (bp->flags & CM_BUF_READING)
838 /* once it has been read once, we can unlock it and return it, still
839 * with its refcount held.
841 lock_ReleaseMutex(&bp->mx);
843 osi_Log3(buf_logp, "buf_GetNew returning bp 0x%x for file 0x%x, offset 0x%x",
844 bp, (long) scp, offsetp->LowPart);
848 /* get a page, returning it held but unlocked. Make sure it is complete */
849 /* The scp must be unlocked when passed to this function */
850 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
854 osi_hyper_t pageOffset;
855 unsigned long tcount;
860 #endif /* DISKCACHE95 */
863 pageOffset.HighPart = offsetp->HighPart;
864 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
868 buf_ValidateBufQueues();
871 bp = buf_Find(scp, &pageOffset);
873 /* lock it and break out */
874 lock_ObtainMutex(&bp->mx);
878 /* touch disk chunk to update LRU info */
879 diskcache_Touch(bp->dcp);
880 #endif /* DISKCACHE95 */
883 /* otherwise, we have to create a page */
884 code = buf_GetNewLocked(scp, &pageOffset, &bp);
886 /* check if the buffer was created in a race condition branch.
887 * If so, go around so we can hold a reference to it.
889 if (code == CM_BUF_EXISTS)
892 /* something else went wrong */
895 buf_ValidateBufQueues();
900 /* otherwise, we have a locked buffer that we just created */
903 } /* big while loop */
905 /* if we get here, we have a locked buffer that may have just been
906 * created, in which case it needs to be filled with data.
909 /* load the page; freshly created pages should be idle */
910 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)));
912 /* setup offset, event */
913 #ifndef DJGPP /* doesn't seem to be used */
914 bp->over.Offset = bp->offset.LowPart;
915 bp->over.OffsetHigh = bp->offset.HighPart;
918 /* start the I/O; may drop lock */
919 bp->flags |= CM_BUF_READING;
920 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
923 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
924 bp->dcp = dcp; /* pointer to disk cache struct. */
925 #endif /* DISKCACHE95 */
928 /* failure or queued */
929 #ifndef DJGPP /* cm_bufRead always returns 0 */
930 if (code != ERROR_IO_PENDING) {
933 bp->flags |= CM_BUF_ERROR;
934 bp->flags &= ~CM_BUF_READING;
935 if (bp->flags & CM_BUF_WAITING) {
936 bp->flags &= ~CM_BUF_WAITING;
937 osi_Wakeup((long) bp);
939 lock_ReleaseMutex(&bp->mx);
942 buf_ValidateBufQueues();
949 /* otherwise, I/O completed instantly and we're done, except
950 * for padding the xfr out with 0s and checking for EOF
952 if (tcount < (unsigned long) cm_data.buf_blockSize) {
953 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
955 bp->flags |= CM_BUF_EOF;
957 bp->flags &= ~CM_BUF_READING;
958 if (bp->flags & CM_BUF_WAITING) {
959 bp->flags &= ~CM_BUF_WAITING;
960 osi_Wakeup((long) bp);
966 /* wait for reads, either that which we started above, or that someone
967 * else started. We don't care if we return a buffer being cleaned.
969 if (bp->flags & CM_BUF_READING)
972 /* once it has been read once, we can unlock it and return it, still
973 * with its refcount held.
975 lock_ReleaseMutex(&bp->mx);
978 /* now remove from queue; will be put in at the head (farthest from
979 * being recycled) when we're done in buf_Release.
981 lock_ObtainWrite(&buf_globalLock);
982 if (bp->flags & CM_BUF_INLRU) {
983 if (cm_data.buf_freeListEndp == bp)
984 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
985 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
986 bp->flags &= ~CM_BUF_INLRU;
988 lock_ReleaseWrite(&buf_globalLock);
990 osi_Log3(buf_logp, "buf_Get returning bp 0x%x for file 0x%x, offset 0x%x",
991 bp, (long) scp, offsetp->LowPart);
993 buf_ValidateBufQueues();
998 /* count # of elements in the free list;
999 * we don't bother doing the proper locking for accessing dataVersion or flags
1000 * since it is a pain, and this is really just an advisory call. If you need
1001 * to do better at some point, rewrite this function.
1003 long buf_CountFreeList(void)
1009 lock_ObtainRead(&buf_globalLock);
1010 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1011 /* if the buffer doesn't have an identity, or if the buffer
1012 * has been invalidate (by having its DV stomped upon), then
1013 * count it as free, since it isn't really being utilized.
1015 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion <= 0)
1018 lock_ReleaseRead(&buf_globalLock);
1022 /* clean a buffer synchronously */
1023 void buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
1025 osi_assert(bp->magic == CM_BUF_MAGIC);
1027 lock_ObtainMutex(&bp->mx);
1028 buf_LockedCleanAsync(bp, reqp);
1029 lock_ReleaseMutex(&bp->mx);
1032 /* wait for a buffer's cleaning to finish */
1033 void buf_CleanWait(cm_buf_t *bp)
1035 osi_assert(bp->magic == CM_BUF_MAGIC);
1037 lock_ObtainMutex(&bp->mx);
1038 if (bp->flags & CM_BUF_WRITING) {
1041 lock_ReleaseMutex(&bp->mx);
1044 /* set the dirty flag on a buffer, and set associated write-ahead log,
1045 * if there is one. Allow one to be added to a buffer, but not changed.
1047 * The buffer must be locked before calling this routine.
1049 void buf_SetDirty(cm_buf_t *bp)
1051 osi_assert(bp->magic == CM_BUF_MAGIC);
1052 osi_assert(bp->refCount > 0);
1054 osi_Log1(buf_logp, "buf_SetDirty 0x%x", bp);
1057 bp->flags |= CM_BUF_DIRTY;
1059 /* and turn off EOF flag, since it has associated data now */
1060 bp->flags &= ~CM_BUF_EOF;
1063 /* clean all buffers, reset log pointers and invalidate all buffers.
1064 * Called with no locks held, and returns with same.
1066 * This function is guaranteed to clean and remove the log ptr of all the
1067 * buffers that were dirty or had non-zero log ptrs before the call was
1068 * made. That's sufficient to clean up any garbage left around by recovery,
1069 * which is all we're counting on this for; there may be newly created buffers
1070 * added while we're running, but that should be OK.
1072 * In an environment where there are no transactions (artificially imposed, for
1073 * example, when switching the database to raw mode), this function is used to
1074 * make sure that all updates have been written to the disk. In that case, we don't
1075 * really require that we forget the log association between pages and logs, but
1076 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1077 * have to worry about invalidating data in the buffers.
1079 * This function is used at the end of recovery as paranoia to get the recovered
1080 * database out to disk. It removes all references to the recovery log and cleans
1083 long buf_CleanAndReset(void)
1089 lock_ObtainWrite(&buf_globalLock);
1090 for(i=0; i<cm_data.buf_hashSize; i++) {
1091 for(bp = cm_data.buf_hashTablepp[i]; bp; bp = bp->hashp) {
1092 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1094 lock_ReleaseWrite(&buf_globalLock);
1096 /* now no locks are held; clean buffer and go on */
1098 buf_CleanAsync(bp, &req);
1101 /* relock and release buffer */
1102 lock_ObtainWrite(&buf_globalLock);
1103 buf_LockedRelease(bp);
1105 } /* over one bucket */
1106 } /* for loop over all hash buckets */
1109 lock_ReleaseWrite(&buf_globalLock);
1112 buf_ValidateBufQueues();
1113 #endif /* TESTING */
1115 /* and we're done */
1119 /* called without global lock being held, reserves buffers for callers
1120 * that need more than one held (not locked) at once.
1122 void buf_ReserveBuffers(long nbuffers)
1124 lock_ObtainWrite(&buf_globalLock);
1126 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1127 cm_data.buf_reserveWaiting = 1;
1128 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1129 osi_SleepW((long) &cm_data.buf_reservedBufs, &buf_globalLock);
1130 lock_ObtainWrite(&buf_globalLock);
1133 cm_data.buf_reservedBufs += nbuffers;
1137 lock_ReleaseWrite(&buf_globalLock);
1140 int buf_TryReserveBuffers(long nbuffers)
1144 lock_ObtainWrite(&buf_globalLock);
1145 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1149 cm_data.buf_reservedBufs += nbuffers;
1152 lock_ReleaseWrite(&buf_globalLock);
1156 /* called without global lock held, releases reservation held by
1157 * buf_ReserveBuffers.
1159 void buf_UnreserveBuffers(long nbuffers)
1161 lock_ObtainWrite(&buf_globalLock);
1162 cm_data.buf_reservedBufs -= nbuffers;
1163 if (cm_data.buf_reserveWaiting) {
1164 cm_data.buf_reserveWaiting = 0;
1165 osi_Wakeup((long) &cm_data.buf_reservedBufs);
1167 lock_ReleaseWrite(&buf_globalLock);
1170 /* truncate the buffers past sizep, zeroing out the page, if we don't
1171 * end on a page boundary.
1173 * Requires cm_bufCreateLock to be write locked.
1175 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1179 cm_buf_t *nbufp; /* next buffer, if didRelease */
1186 /* assert that cm_bufCreateLock is held in write mode */
1187 lock_AssertWrite(&scp->bufCreateLock);
1189 i = BUF_FILEHASH(&scp->fid);
1191 lock_ObtainWrite(&buf_globalLock);
1192 bufp = cm_data.buf_fileHashTablepp[i];
1194 lock_ReleaseWrite(&buf_globalLock);
1199 lock_ReleaseWrite(&buf_globalLock);
1200 for(; bufp; bufp = nbufp) {
1202 lock_ObtainMutex(&bufp->mx);
1204 bufEnd.HighPart = 0;
1205 bufEnd.LowPart = cm_data.buf_blockSize;
1206 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1208 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1209 LargeIntegerLessThan(*sizep, bufEnd)) {
1212 lock_ObtainMutex(&scp->mx);
1214 /* make sure we have a callback (so we have the right value for
1215 * the length), and wait for it to be safe to do a truncate.
1217 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1218 CM_SCACHESYNC_NEEDCALLBACK
1219 | CM_SCACHESYNC_GETSTATUS
1220 | CM_SCACHESYNC_SETSIZE
1221 | CM_SCACHESYNC_BUFLOCKED);
1222 /* if we succeeded in our locking, and this applies to the right
1223 * file, and the truncate request overlaps the buffer either
1224 * totally or partially, then do something.
1226 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1227 && LargeIntegerLessThan(*sizep, bufEnd)) {
1229 lock_ObtainWrite(&buf_globalLock);
1231 /* destroy the buffer, turning off its dirty bit, if
1232 * we're truncating the whole buffer. Otherwise, set
1233 * the dirty bit, and clear out the tail of the buffer
1234 * if we just overlap some.
1236 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1237 /* truncating the entire page */
1238 bufp->flags &= ~CM_BUF_DIRTY;
1239 bufp->dataVersion = -1; /* known bad */
1240 bufp->dirtyCounter++;
1243 /* don't set dirty, since dirty implies
1244 * currently up-to-date. Don't need to do this,
1245 * since we'll update the length anyway.
1247 * Zero out remainder of the page, in case we
1248 * seek and write past EOF, and make this data
1251 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1252 osi_assert(bufferPos != 0);
1253 memset(bufp->datap + bufferPos, 0,
1254 cm_data.buf_blockSize - bufferPos);
1257 lock_ReleaseWrite(&buf_globalLock);
1260 lock_ReleaseMutex(&scp->mx);
1261 lock_ReleaseMutex(&bufp->mx);
1263 lock_ObtainWrite(&buf_globalLock);
1264 nbufp = bufp->fileHashp;
1265 if (nbufp) nbufp->refCount++;
1266 buf_LockedRelease(bufp);
1267 lock_ReleaseWrite(&buf_globalLock);
1270 /* bail out early if we fail */
1272 /* at this point, nbufp is held; bufp has already been
1279 buf_ValidateBufQueues();
1280 #endif /* TESTING */
1287 buf_ValidateBufQueues();
1288 #endif /* TESTING */
1294 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1297 cm_buf_t *bp; /* buffer we're hacking on */
1302 i = BUF_FILEHASH(&scp->fid);
1305 lock_ObtainWrite(&buf_globalLock);
1306 bp = cm_data.buf_fileHashTablepp[i];
1309 lock_ReleaseWrite(&buf_globalLock);
1310 for (; bp; bp = nbp) {
1311 didRelease = 0; /* haven't released this buffer yet */
1313 /* clean buffer synchronously */
1314 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1315 lock_ObtainMutex(&bp->mx);
1317 /* start cleaning the buffer, and wait for it to finish */
1318 buf_LockedCleanAsync(bp, reqp);
1320 lock_ReleaseMutex(&bp->mx);
1322 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1326 lock_ObtainWrite(&buf_globalLock);
1327 /* actually, we only know that buffer is clean if ref
1328 * count is 1, since we don't have buffer itself locked.
1330 if (!(bp->flags & CM_BUF_DIRTY)) {
1331 if (bp->refCount == 1) { /* bp is held above */
1332 buf_LockedRelease(bp);
1333 nbp = bp->fileHashp;
1340 lock_ReleaseWrite(&buf_globalLock);
1342 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1347 lock_ObtainWrite(&buf_globalLock);
1348 if (nbp = bp->fileHashp)
1350 buf_LockedRelease(bp);
1351 lock_ReleaseWrite(&buf_globalLock);
1353 } /* for loop over a bunch of buffers */
1356 buf_ValidateBufQueues();
1357 #endif /* TESTING */
1363 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1366 cm_buf_t *bp; /* buffer we're hacking on */
1367 cm_buf_t *nbp; /* next one */
1370 i = BUF_FILEHASH(&scp->fid);
1373 lock_ObtainWrite(&buf_globalLock);
1374 bp = cm_data.buf_fileHashTablepp[i];
1377 lock_ReleaseWrite(&buf_globalLock);
1378 for (; bp; bp = nbp) {
1379 /* clean buffer synchronously */
1380 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1383 lock_ObtainMutex(&bp->mx);
1385 cm_ReleaseUser(bp->userp);
1387 lock_ReleaseMutex(&bp->mx);
1389 buf_CleanAsync(bp, reqp);
1391 lock_ObtainMutex(&bp->mx);
1392 if (bp->flags & CM_BUF_ERROR) {
1393 if (code == 0 || code == -1)
1398 lock_ReleaseMutex(&bp->mx);
1401 lock_ObtainWrite(&buf_globalLock);
1402 buf_LockedRelease(bp);
1403 nbp = bp->fileHashp;
1406 lock_ReleaseWrite(&buf_globalLock);
1407 } /* for loop over a bunch of buffers */
1410 buf_ValidateBufQueues();
1411 #endif /* TESTING */
1419 buf_ValidateBufQueues(void)
1421 cm_buf_t * bp, *bpb, *bpf, *bpa;
1422 afs_uint32 countf=0, countb=0, counta=0;
1424 lock_ObtainRead(&buf_globalLock);
1425 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1426 if (bp->magic != CM_BUF_MAGIC)
1432 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1433 if (bp->magic != CM_BUF_MAGIC)
1439 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1440 if (bp->magic != CM_BUF_MAGIC)
1445 lock_ReleaseRead(&buf_globalLock);
1447 if (countb != countf)
1450 if (counta != cm_data.buf_nbuffers)
1453 #endif /* TESTING */
1455 /* dump the contents of the buf_hashTablepp. */
1456 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1463 if (cm_data.buf_hashTablepp == NULL)
1467 lock_ObtainRead(&buf_globalLock);
1469 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\n",
1470 cookie, cm_data.buf_hashSize);
1471 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1473 for (i = 0; i < cm_data.buf_hashSize; i++)
1475 for (bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp)
1479 StringCbPrintfA(output, sizeof(output), "vnode=%d, unique=%d), size=%d refCount=%d\n",
1480 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1481 bp->fid.vnode, bp->fid.unique, bp->size, bp->refCount);
1482 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1487 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\n", cookie);
1488 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1491 lock_ReleaseRead(&buf_globalLock);
1495 void buf_ForceTrace(BOOL flush)
1504 len = GetTempPath(sizeof(buf)-10, buf);
1505 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1506 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1507 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1508 if (handle == INVALID_HANDLE_VALUE) {
1509 osi_panic("Cannot create log file", __FILE__, __LINE__);
1511 osi_LogPrint(buf_logp, handle);
1513 FlushFileBuffers(handle);
1514 CloseHandle(handle);