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 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
160 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
161 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
162 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
166 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
167 if (bp->magic != CM_BUF_MAGIC) {
168 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
169 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
175 if (countb > cm_data.buf_nbuffers) {
176 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
177 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
182 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
183 if (bp->magic != CM_BUF_MAGIC) {
184 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
185 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
191 if (countf > cm_data.buf_nbuffers) {
192 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
193 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
198 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
199 if (bp->magic != CM_BUF_MAGIC) {
200 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
201 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
207 if (counta > cm_data.buf_nbuffers) {
208 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
209 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
214 if (countb != countf) {
215 afsi_log("cm_ValidateBuffers failure: countb != countf");
216 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
220 if (counta != cm_data.buf_nbuffers) {
221 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
222 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
229 void buf_Shutdown(void)
231 buf_ShutdownFlag = 1;
234 /* initialize the buffer package; called with no locks
235 * held during the initialization phase.
237 long buf_Init(int newFile, cm_buf_ops_t *opsp, long nbuffers)
239 static osi_once_t once;
248 cm_data.buf_nbuffers = nbuffers;
250 /* Have to be able to reserve a whole chunk */
251 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
252 return CM_ERROR_TOOFEWBUFS;
255 /* recall for callouts */
258 if (osi_Once(&once)) {
259 /* initialize global locks */
260 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock");
263 /* remember this for those who want to reset it */
264 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
266 /* lower hash size to a prime number */
267 cm_data.buf_hashSize = osi_PrimeLessThan(CM_BUF_HASHSIZE);
269 /* create hash table */
270 memset((void *)cm_data.buf_hashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
272 /* another hash table */
273 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
275 /* create buffer headers and put in free list */
276 bp = cm_data.bufHeaderBaseAddress;
277 data = cm_data.bufDataBaseAddress;
278 cm_data.buf_allp = NULL;
280 for (i=0; i<cm_data.buf_nbuffers; i++) {
281 osi_assert(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress);
282 osi_assert(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData);
284 /* allocate and zero some storage */
285 memset(bp, 0, sizeof(cm_buf_t));
286 bp->magic = CM_BUF_MAGIC;
287 /* thread on list of all buffers */
288 bp->allp = cm_data.buf_allp;
289 cm_data.buf_allp = bp;
291 osi_QAdd((osi_queue_t **)&cm_data.buf_freeListp, &bp->q);
292 bp->flags |= CM_BUF_INLRU;
293 lock_InitializeMutex(&bp->mx, "Buffer mutex");
295 /* grab appropriate number of bytes from aligned zone */
298 /* setup last buffer pointer */
300 cm_data.buf_freeListEndp = bp;
304 data += cm_data.buf_blockSize;
307 /* none reserved at first */
308 cm_data.buf_reservedBufs = 0;
310 /* just for safety's sake */
311 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
313 bp = cm_data.bufHeaderBaseAddress;
314 data = cm_data.bufDataBaseAddress;
316 for (i=0; i<cm_data.buf_nbuffers; i++) {
317 lock_InitializeMutex(&bp->mx, "Buffer mutex");
320 bp->waitRequests = 0;
321 bp->flags &= ~CM_BUF_WAITING;
327 buf_ValidateBufQueues();
331 /* init the buffer trace log */
332 buf_logp = osi_LogCreate("buffer", 1000);
333 osi_LogEnable(buf_logp);
338 /* and create the incr-syncer */
339 phandle = thrd_Create(0, 0,
340 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
343 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
345 CloseHandle(phandle);
350 buf_ValidateBufQueues();
355 /* add nbuffers to the buffer pool, if possible.
356 * Called with no locks held.
358 long buf_AddBuffers(long nbuffers)
361 /* The size of a virtual cache cannot be changed after it has
362 * been created. Subsequent calls to MapViewofFile() with
363 * an existing mapping object name would not allow the
364 * object to be resized. Return failure immediately.
366 * A similar problem now occurs with the persistent cache
367 * given that the memory mapped file now contains a complex
370 afsi_log("request to add %d buffers to the existing cache of size %d denied",
371 nbuffers, cm_data.buf_nbuffers);
373 return CM_ERROR_INVAL;
379 data = malloc(buf_nbuffers * cm_data.buf_blockSize);
381 /* Create buffer headers and put in free list */
382 bp = malloc(nbuffers * sizeof(*bp));
384 for (i=0; i<nbuffers; i++) {
385 memset(bp, 0, sizeof(*bp));
387 lock_InitializeMutex(&bp->mx, "cm_buf_t");
389 /* grab appropriate number of bytes from aligned zone */
392 bp->flags |= CM_BUF_INLRU;
394 lock_ObtainWrite(&buf_globalLock);
395 /* note that buf_allp chain is covered by buf_globalLock now */
396 bp->allp = cm_data.buf_allp;
397 cm_data.buf_allp = bp;
398 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
399 if (!cm_data.buf_freeListEndp)
400 cm_data.buf_freeListEndp = bp;
401 cm_data.buf_nbuffers++;
402 lock_ReleaseWrite(&buf_globalLock);
405 data += cm_data.buf_blockSize;
407 } /* for loop over all buffers */
413 /* interface to set the number of buffers to an exact figure.
414 * Called with no locks held.
416 long buf_SetNBuffers(long nbuffers)
419 return CM_ERROR_INVAL;
420 if (nbuffers == cm_data.buf_nbuffers)
422 else if (nbuffers > cm_data.buf_nbuffers)
423 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
425 return CM_ERROR_INVAL;
428 /* release a buffer. Buffer must be referenced, but unlocked. */
429 void buf_Release(cm_buf_t *bp)
431 lock_ObtainWrite(&buf_globalLock);
432 buf_LockedRelease(bp);
433 lock_ReleaseWrite(&buf_globalLock);
436 /* wait for reading or writing to clear; called with write-locked
437 * buffer, and returns with locked buffer.
439 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
442 osi_assert(scp->magic == CM_SCACHE_MAGIC);
443 osi_assert(bp->magic == CM_BUF_MAGIC);
446 /* if no IO is happening, we're done */
447 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
450 /* otherwise I/O is happening, but some other thread is waiting for
451 * the I/O already. Wait for that guy to figure out what happened,
452 * and then check again.
454 if ( bp->flags & CM_BUF_WAITING ) {
457 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%x", bp);
459 osi_Log1(buf_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%x", bp);
460 bp->flags |= CM_BUF_WAITING;
461 bp->waitCount = bp->waitRequests = 1;
463 osi_SleepM((long) bp, &bp->mx);
464 lock_ObtainMutex(&bp->mx);
465 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%x", bp);
467 if (bp->waitCount == 0) {
468 osi_Log1(afsd_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%x", bp);
469 bp->flags &= ~CM_BUF_WAITING;
470 bp->waitRequests = 0;
474 scp = cm_FindSCache(&bp->fid);
477 lock_ObtainMutex(&scp->mx);
478 if (scp->flags & CM_SCACHEFLAG_WAITING) {
479 osi_Log1(buf_logp, "buf_WaitIO waking scp 0x%x", scp);
480 osi_Wakeup((long)&scp->flags);
481 lock_ReleaseMutex(&scp->mx);
486 /* if we get here, the IO is done, but we may have to wakeup people waiting for
487 * the I/O to complete. Do so.
489 if (bp->flags & CM_BUF_WAITING) {
490 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%x", bp);
491 osi_Wakeup((long) bp);
493 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%x", (long) bp);
496 /* code to drop reference count while holding buf_globalLock */
497 void buf_LockedRelease(cm_buf_t *bp)
499 /* ensure that we're in the LRU queue if our ref count is 0 */
500 osi_assert(bp->refCount > 0);
501 if (--bp->refCount == 0) {
502 if (!(bp->flags & CM_BUF_INLRU)) {
503 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
505 /* watch for transition from empty to one element */
506 if (!cm_data.buf_freeListEndp)
507 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
508 bp->flags |= CM_BUF_INLRU;
513 /* find a buffer, if any, for a particular file ID and offset. Assumes
514 * that buf_globalLock is write locked when called.
516 cm_buf_t *buf_LockedFind(struct cm_scache *scp, osi_hyper_t *offsetp)
521 i = BUF_HASH(&scp->fid, offsetp);
522 for(bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp) {
523 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
524 && offsetp->LowPart == bp->offset.LowPart
525 && offsetp->HighPart == bp->offset.HighPart) {
531 /* return whatever we found, if anything */
535 /* find a buffer with offset *offsetp for vnode *scp. Called
536 * with no locks held.
538 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
542 lock_ObtainWrite(&buf_globalLock);
543 bp = buf_LockedFind(scp, offsetp);
544 lock_ReleaseWrite(&buf_globalLock);
549 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
552 * Makes sure that there's only one person writing this block
553 * at any given time, and also ensures that the log is forced sufficiently far,
554 * if this buffer contains logged data.
556 void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
560 osi_assert(bp->magic == CM_BUF_MAGIC);
562 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
563 lock_ReleaseMutex(&bp->mx);
565 code = (*cm_buf_opsp->Writep)(&bp->fid, &bp->offset,
566 cm_data.buf_blockSize, 0, bp->userp,
569 lock_ObtainMutex(&bp->mx);
574 /* Disk cache support */
575 /* write buffer to disk cache (synchronous for now) */
576 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
577 #endif /* DISKCACHE95 */
580 /* do logging after call to GetLastError, or else */
581 osi_Log2(buf_logp, "buf_CleanAsync starts I/O on 0x%x, done=%d", bp, code);
583 /* if someone was waiting for the I/O that just completed or failed,
586 if (bp->flags & CM_BUF_WAITING) {
587 /* turn off flags and wakeup users */
588 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%x", bp);
589 osi_Wakeup((long) bp);
593 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
594 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
595 * The buffer must already be clean, and no I/O should be happening to it.
597 void buf_Recycle(cm_buf_t *bp)
602 cm_buf_t *prevBp, *nextBp;
604 osi_assert(bp->magic == CM_BUF_MAGIC);
606 /* if we get here, we know that the buffer still has a 0 ref count,
607 * and that it is clean and has no currently pending I/O. This is
608 * the dude to return.
609 * Remember that as long as the ref count is 0, we know that we won't
610 * have any lock conflicts, so we can grab the buffer lock out of
611 * order in the locking hierarchy.
613 osi_Log2( buf_logp, "buf_Recycle recycles 0x%x, off 0x%x",
614 bp, bp->offset.LowPart);
616 osi_assert(bp->refCount == 0);
617 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)));
618 lock_AssertWrite(&buf_globalLock);
620 if (bp->flags & CM_BUF_INHASH) {
621 /* Remove from hash */
623 i = BUF_HASH(&bp->fid, &bp->offset);
624 lbpp = &(cm_data.buf_hashTablepp[i]);
625 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
626 if (tbp == bp) break;
629 /* we better find it */
630 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
632 *lbpp = bp->hashp; /* hash out */
634 /* Remove from file hash */
636 i = BUF_FILEHASH(&bp->fid);
637 prevBp = bp->fileHashBackp;
638 nextBp = bp->fileHashp;
640 prevBp->fileHashp = nextBp;
642 cm_data.buf_fileHashTablepp[i] = nextBp;
644 nextBp->fileHashBackp = prevBp;
646 bp->flags &= ~CM_BUF_INHASH;
649 /* bump the soft reference counter now, to invalidate softRefs; no
650 * wakeup is required since people don't sleep waiting for this
655 /* make the fid unrecognizable */
656 memset(&bp->fid, 0, sizeof(cm_fid_t));
659 /* recycle a buffer, removing it from the free list, hashing in its new identity
660 * and returning it write-locked so that no one can use it. Called without
661 * any locks held, and can return an error if it loses the race condition and
662 * finds that someone else created the desired buffer.
664 * If success is returned, the buffer is returned write-locked.
666 * May be called with null scp and offsetp, if we're just trying to reclaim some
667 * space from the buffer pool. In that case, the buffer will be returned
668 * without being hashed into the hash table.
670 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
672 cm_buf_t *bp; /* buffer we're dealing with */
673 cm_buf_t *nextBp; /* next buffer in file hash chain */
677 cm_InitReq(&req); /* just in case */
680 buf_ValidateBufQueues();
685 lock_ObtainWrite(&buf_globalLock);
686 /* check to see if we lost the race */
688 if (bp = buf_LockedFind(scp, offsetp)) {
690 lock_ReleaseWrite(&buf_globalLock);
691 return CM_BUF_EXISTS;
695 /* does this fix the problem below? it's a simple solution. */
696 if (!cm_data.buf_freeListEndp)
698 lock_ReleaseWrite(&buf_globalLock);
703 /* for debugging, assert free list isn't empty, although we
704 * really should try waiting for a running tranasction to finish
705 * instead of this; or better, we should have a transaction
706 * throttler prevent us from entering this situation.
708 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
710 /* look at all buffers in free list, some of which may temp.
711 * have high refcounts and which then should be skipped,
712 * starting cleaning I/O for those which are dirty. If we find
713 * a clean buffer, we rehash it, lock it and return it.
715 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
716 /* check to see if it really has zero ref count. This
717 * code can bump refcounts, at least, so it may not be
720 if (bp->refCount > 0)
723 /* we don't have to lock buffer itself, since the ref
724 * count is 0 and we know it will stay zero as long as
725 * we hold the global lock.
728 /* don't recycle someone in our own chunk */
729 if (!cm_FidCmp(&bp->fid, &scp->fid)
730 && (bp->offset.LowPart & (-cm_chunkSize))
731 == (offsetp->LowPart & (-cm_chunkSize)))
734 /* if this page is being filled (!) or cleaned, see if
735 * the I/O has completed. If not, skip it, otherwise
736 * do the final processing for the I/O.
738 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
739 /* probably shouldn't do this much work while
740 * holding the big lock? Watch for contention
746 if (bp->flags & CM_BUF_DIRTY) {
747 /* if the buffer is dirty, start cleaning it and
748 * move on to the next buffer. We do this with
749 * just the lock required to minimize contention
753 lock_ReleaseWrite(&buf_globalLock);
755 /* grab required lock and clean; this only
756 * starts the I/O. By the time we're back,
757 * it'll still be marked dirty, but it will also
758 * have the WRITING flag set, so we won't get
761 buf_CleanAsync(bp, &req);
763 /* now put it back and go around again */
768 /* if we get here, we know that the buffer still has a 0
769 * ref count, and that it is clean and has no currently
770 * pending I/O. This is the dude to return.
771 * Remember that as long as the ref count is 0, we know
772 * that we won't have any lock conflicts, so we can grab
773 * the buffer lock out of order in the locking hierarchy.
777 /* clean up junk flags */
778 bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
779 bp->dataVersion = -1; /* unknown so far */
781 /* now hash in as our new buffer, and give it the
782 * appropriate label, if requested.
785 bp->flags |= CM_BUF_INHASH;
787 bp->offset = *offsetp;
788 i = BUF_HASH(&scp->fid, offsetp);
789 bp->hashp = cm_data.buf_hashTablepp[i];
790 cm_data.buf_hashTablepp[i] = bp;
791 i = BUF_FILEHASH(&scp->fid);
792 nextBp = cm_data.buf_fileHashTablepp[i];
793 bp->fileHashp = nextBp;
794 bp->fileHashBackp = NULL;
796 nextBp->fileHashBackp = bp;
797 cm_data.buf_fileHashTablepp[i] = bp;
800 /* prepare to return it. Start by giving it a good
804 /* and since it has a non-zero ref count, we should move
805 * it from the lru queue. It better be still there,
806 * since we've held the global (big) lock since we found
809 osi_assertx(bp->flags & CM_BUF_INLRU,
810 "buf_GetNewLocked: LRU screwup");
811 if (cm_data.buf_freeListEndp == bp) {
812 /* we're the last guy in this queue, so maintain it */
813 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
815 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
816 bp->flags &= ~CM_BUF_INLRU;
818 /* finally, grab the mutex so that people don't use it
819 * before the caller fills it with data. Again, no one
820 * should have been able to get to this dude to lock it.
822 osi_assertx(lock_TryMutex(&bp->mx),
823 "buf_GetNewLocked: TryMutex failed");
825 lock_ReleaseWrite(&buf_globalLock);
829 buf_ValidateBufQueues();
832 } /* for all buffers in lru queue */
833 lock_ReleaseWrite(&buf_globalLock);
834 } /* while loop over everything */
838 /* get a page, returning it held but unlocked. Doesn't fill in the page
839 * with I/O, since we're going to write the whole thing new.
841 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
845 osi_hyper_t pageOffset;
849 pageOffset.HighPart = offsetp->HighPart;
850 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
852 bp = buf_Find(scp, &pageOffset);
854 /* lock it and break out */
855 lock_ObtainMutex(&bp->mx);
859 /* otherwise, we have to create a page */
860 code = buf_GetNewLocked(scp, &pageOffset, &bp);
862 /* check if the buffer was created in a race condition branch.
863 * If so, go around so we can hold a reference to it.
865 if (code == CM_BUF_EXISTS)
868 /* something else went wrong */
872 /* otherwise, we have a locked buffer that we just created */
875 } /* big while loop */
878 if (bp->flags & CM_BUF_READING)
881 /* once it has been read once, we can unlock it and return it, still
882 * with its refcount held.
884 lock_ReleaseMutex(&bp->mx);
886 osi_Log3(buf_logp, "buf_GetNew returning bp 0x%x for file 0x%x, offset 0x%x",
887 bp, (long) scp, offsetp->LowPart);
891 /* get a page, returning it held but unlocked. Make sure it is complete */
892 /* The scp must be unlocked when passed to this function */
893 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
897 osi_hyper_t pageOffset;
898 unsigned long tcount;
903 #endif /* DISKCACHE95 */
906 pageOffset.HighPart = offsetp->HighPart;
907 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
911 buf_ValidateBufQueues();
914 bp = buf_Find(scp, &pageOffset);
916 /* lock it and break out */
917 lock_ObtainMutex(&bp->mx);
921 /* touch disk chunk to update LRU info */
922 diskcache_Touch(bp->dcp);
923 #endif /* DISKCACHE95 */
926 /* otherwise, we have to create a page */
927 code = buf_GetNewLocked(scp, &pageOffset, &bp);
929 /* check if the buffer was created in a race condition branch.
930 * If so, go around so we can hold a reference to it.
932 if (code == CM_BUF_EXISTS)
935 /* something else went wrong */
938 buf_ValidateBufQueues();
943 /* otherwise, we have a locked buffer that we just created */
946 } /* big while loop */
948 /* if we get here, we have a locked buffer that may have just been
949 * created, in which case it needs to be filled with data.
952 /* load the page; freshly created pages should be idle */
953 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)));
955 /* setup offset, event */
956 #ifndef DJGPP /* doesn't seem to be used */
957 bp->over.Offset = bp->offset.LowPart;
958 bp->over.OffsetHigh = bp->offset.HighPart;
961 /* start the I/O; may drop lock */
962 bp->flags |= CM_BUF_READING;
963 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
966 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
967 bp->dcp = dcp; /* pointer to disk cache struct. */
968 #endif /* DISKCACHE95 */
971 /* failure or queued */
972 #ifndef DJGPP /* cm_bufRead always returns 0 */
973 if (code != ERROR_IO_PENDING) {
976 bp->flags |= CM_BUF_ERROR;
977 bp->flags &= ~CM_BUF_READING;
978 if (bp->flags & CM_BUF_WAITING) {
979 osi_Log1(buf_logp, "buf_Get Waking bp 0x%x", bp);
980 osi_Wakeup((long) bp);
982 lock_ReleaseMutex(&bp->mx);
985 buf_ValidateBufQueues();
992 /* otherwise, I/O completed instantly and we're done, except
993 * for padding the xfr out with 0s and checking for EOF
995 if (tcount < (unsigned long) cm_data.buf_blockSize) {
996 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
998 bp->flags |= CM_BUF_EOF;
1000 bp->flags &= ~CM_BUF_READING;
1001 if (bp->flags & CM_BUF_WAITING) {
1002 osi_Log1(buf_logp, "buf_Get Waking bp 0x%x", bp);
1003 osi_Wakeup((long) bp);
1009 /* wait for reads, either that which we started above, or that someone
1010 * else started. We don't care if we return a buffer being cleaned.
1012 if (bp->flags & CM_BUF_READING)
1013 buf_WaitIO(scp, bp);
1015 /* once it has been read once, we can unlock it and return it, still
1016 * with its refcount held.
1018 lock_ReleaseMutex(&bp->mx);
1021 /* now remove from queue; will be put in at the head (farthest from
1022 * being recycled) when we're done in buf_Release.
1024 lock_ObtainWrite(&buf_globalLock);
1025 if (bp->flags & CM_BUF_INLRU) {
1026 if (cm_data.buf_freeListEndp == bp)
1027 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1028 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1029 bp->flags &= ~CM_BUF_INLRU;
1031 lock_ReleaseWrite(&buf_globalLock);
1033 osi_Log3(buf_logp, "buf_Get returning bp 0x%x for file 0x%x, offset 0x%x",
1034 bp, (long) scp, offsetp->LowPart);
1036 buf_ValidateBufQueues();
1037 #endif /* TESTING */
1041 /* count # of elements in the free list;
1042 * we don't bother doing the proper locking for accessing dataVersion or flags
1043 * since it is a pain, and this is really just an advisory call. If you need
1044 * to do better at some point, rewrite this function.
1046 long buf_CountFreeList(void)
1052 lock_ObtainRead(&buf_globalLock);
1053 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1054 /* if the buffer doesn't have an identity, or if the buffer
1055 * has been invalidate (by having its DV stomped upon), then
1056 * count it as free, since it isn't really being utilized.
1058 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion <= 0)
1061 lock_ReleaseRead(&buf_globalLock);
1065 /* clean a buffer synchronously */
1066 void buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
1068 osi_assert(bp->magic == CM_BUF_MAGIC);
1070 lock_ObtainMutex(&bp->mx);
1071 buf_LockedCleanAsync(bp, reqp);
1072 lock_ReleaseMutex(&bp->mx);
1075 /* wait for a buffer's cleaning to finish */
1076 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp)
1078 osi_assert(bp->magic == CM_BUF_MAGIC);
1080 lock_ObtainMutex(&bp->mx);
1081 if (bp->flags & CM_BUF_WRITING) {
1082 buf_WaitIO(scp, bp);
1084 lock_ReleaseMutex(&bp->mx);
1087 /* set the dirty flag on a buffer, and set associated write-ahead log,
1088 * if there is one. Allow one to be added to a buffer, but not changed.
1090 * The buffer must be locked before calling this routine.
1092 void buf_SetDirty(cm_buf_t *bp)
1094 osi_assert(bp->magic == CM_BUF_MAGIC);
1095 osi_assert(bp->refCount > 0);
1097 osi_Log1(buf_logp, "buf_SetDirty 0x%x", bp);
1100 bp->flags |= CM_BUF_DIRTY;
1102 /* and turn off EOF flag, since it has associated data now */
1103 bp->flags &= ~CM_BUF_EOF;
1106 /* clean all buffers, reset log pointers and invalidate all buffers.
1107 * Called with no locks held, and returns with same.
1109 * This function is guaranteed to clean and remove the log ptr of all the
1110 * buffers that were dirty or had non-zero log ptrs before the call was
1111 * made. That's sufficient to clean up any garbage left around by recovery,
1112 * which is all we're counting on this for; there may be newly created buffers
1113 * added while we're running, but that should be OK.
1115 * In an environment where there are no transactions (artificially imposed, for
1116 * example, when switching the database to raw mode), this function is used to
1117 * make sure that all updates have been written to the disk. In that case, we don't
1118 * really require that we forget the log association between pages and logs, but
1119 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1120 * have to worry about invalidating data in the buffers.
1122 * This function is used at the end of recovery as paranoia to get the recovered
1123 * database out to disk. It removes all references to the recovery log and cleans
1126 long buf_CleanAndReset(void)
1132 lock_ObtainWrite(&buf_globalLock);
1133 for(i=0; i<cm_data.buf_hashSize; i++) {
1134 for(bp = cm_data.buf_hashTablepp[i]; bp; bp = bp->hashp) {
1135 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1137 lock_ReleaseWrite(&buf_globalLock);
1139 /* now no locks are held; clean buffer and go on */
1141 buf_CleanAsync(bp, &req);
1142 buf_CleanWait(NULL, bp);
1144 /* relock and release buffer */
1145 lock_ObtainWrite(&buf_globalLock);
1146 buf_LockedRelease(bp);
1148 } /* over one bucket */
1149 } /* for loop over all hash buckets */
1152 lock_ReleaseWrite(&buf_globalLock);
1155 buf_ValidateBufQueues();
1156 #endif /* TESTING */
1158 /* and we're done */
1162 /* called without global lock being held, reserves buffers for callers
1163 * that need more than one held (not locked) at once.
1165 void buf_ReserveBuffers(long nbuffers)
1167 lock_ObtainWrite(&buf_globalLock);
1169 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1170 cm_data.buf_reserveWaiting = 1;
1171 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1172 osi_SleepW((long) &cm_data.buf_reservedBufs, &buf_globalLock);
1173 lock_ObtainWrite(&buf_globalLock);
1176 cm_data.buf_reservedBufs += nbuffers;
1180 lock_ReleaseWrite(&buf_globalLock);
1183 int buf_TryReserveBuffers(long nbuffers)
1187 lock_ObtainWrite(&buf_globalLock);
1188 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1192 cm_data.buf_reservedBufs += nbuffers;
1195 lock_ReleaseWrite(&buf_globalLock);
1199 /* called without global lock held, releases reservation held by
1200 * buf_ReserveBuffers.
1202 void buf_UnreserveBuffers(long nbuffers)
1204 lock_ObtainWrite(&buf_globalLock);
1205 cm_data.buf_reservedBufs -= nbuffers;
1206 if (cm_data.buf_reserveWaiting) {
1207 cm_data.buf_reserveWaiting = 0;
1208 osi_Wakeup((long) &cm_data.buf_reservedBufs);
1210 lock_ReleaseWrite(&buf_globalLock);
1213 /* truncate the buffers past sizep, zeroing out the page, if we don't
1214 * end on a page boundary.
1216 * Requires cm_bufCreateLock to be write locked.
1218 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1222 cm_buf_t *nbufp; /* next buffer, if didRelease */
1229 /* assert that cm_bufCreateLock is held in write mode */
1230 lock_AssertWrite(&scp->bufCreateLock);
1232 i = BUF_FILEHASH(&scp->fid);
1234 lock_ObtainWrite(&buf_globalLock);
1235 bufp = cm_data.buf_fileHashTablepp[i];
1237 lock_ReleaseWrite(&buf_globalLock);
1242 lock_ReleaseWrite(&buf_globalLock);
1243 for(; bufp; bufp = nbufp) {
1245 lock_ObtainMutex(&bufp->mx);
1247 bufEnd.HighPart = 0;
1248 bufEnd.LowPart = cm_data.buf_blockSize;
1249 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1251 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1252 LargeIntegerLessThan(*sizep, bufEnd)) {
1253 buf_WaitIO(scp, bufp);
1255 lock_ObtainMutex(&scp->mx);
1257 /* make sure we have a callback (so we have the right value for
1258 * the length), and wait for it to be safe to do a truncate.
1260 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1261 CM_SCACHESYNC_NEEDCALLBACK
1262 | CM_SCACHESYNC_GETSTATUS
1263 | CM_SCACHESYNC_SETSIZE
1264 | CM_SCACHESYNC_BUFLOCKED);
1265 /* if we succeeded in our locking, and this applies to the right
1266 * file, and the truncate request overlaps the buffer either
1267 * totally or partially, then do something.
1269 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1270 && LargeIntegerLessThan(*sizep, bufEnd)) {
1272 lock_ObtainWrite(&buf_globalLock);
1274 /* destroy the buffer, turning off its dirty bit, if
1275 * we're truncating the whole buffer. Otherwise, set
1276 * the dirty bit, and clear out the tail of the buffer
1277 * if we just overlap some.
1279 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1280 /* truncating the entire page */
1281 bufp->flags &= ~CM_BUF_DIRTY;
1282 bufp->dataVersion = -1; /* known bad */
1283 bufp->dirtyCounter++;
1286 /* don't set dirty, since dirty implies
1287 * currently up-to-date. Don't need to do this,
1288 * since we'll update the length anyway.
1290 * Zero out remainder of the page, in case we
1291 * seek and write past EOF, and make this data
1294 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1295 osi_assert(bufferPos != 0);
1296 memset(bufp->datap + bufferPos, 0,
1297 cm_data.buf_blockSize - bufferPos);
1300 lock_ReleaseWrite(&buf_globalLock);
1303 lock_ReleaseMutex(&scp->mx);
1304 lock_ReleaseMutex(&bufp->mx);
1306 lock_ObtainWrite(&buf_globalLock);
1307 nbufp = bufp->fileHashp;
1308 if (nbufp) nbufp->refCount++;
1309 buf_LockedRelease(bufp);
1310 lock_ReleaseWrite(&buf_globalLock);
1313 /* bail out early if we fail */
1315 /* at this point, nbufp is held; bufp has already been
1322 buf_ValidateBufQueues();
1323 #endif /* TESTING */
1330 buf_ValidateBufQueues();
1331 #endif /* TESTING */
1337 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1340 cm_buf_t *bp; /* buffer we're hacking on */
1345 i = BUF_FILEHASH(&scp->fid);
1348 lock_ObtainWrite(&buf_globalLock);
1349 bp = cm_data.buf_fileHashTablepp[i];
1352 lock_ReleaseWrite(&buf_globalLock);
1353 for (; bp; bp = nbp) {
1354 didRelease = 0; /* haven't released this buffer yet */
1356 /* clean buffer synchronously */
1357 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1358 lock_ObtainMutex(&bp->mx);
1360 /* start cleaning the buffer, and wait for it to finish */
1361 buf_LockedCleanAsync(bp, reqp);
1362 buf_WaitIO(scp, bp);
1363 lock_ReleaseMutex(&bp->mx);
1365 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1369 lock_ObtainWrite(&buf_globalLock);
1370 /* actually, we only know that buffer is clean if ref
1371 * count is 1, since we don't have buffer itself locked.
1373 if (!(bp->flags & CM_BUF_DIRTY)) {
1374 if (bp->refCount == 1) { /* bp is held above */
1375 buf_LockedRelease(bp);
1376 nbp = bp->fileHashp;
1383 lock_ReleaseWrite(&buf_globalLock);
1385 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1390 lock_ObtainWrite(&buf_globalLock);
1391 if (nbp = bp->fileHashp)
1393 buf_LockedRelease(bp);
1394 lock_ReleaseWrite(&buf_globalLock);
1396 } /* for loop over a bunch of buffers */
1399 buf_ValidateBufQueues();
1400 #endif /* TESTING */
1406 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1409 cm_buf_t *bp; /* buffer we're hacking on */
1410 cm_buf_t *nbp; /* next one */
1413 i = BUF_FILEHASH(&scp->fid);
1416 lock_ObtainWrite(&buf_globalLock);
1417 bp = cm_data.buf_fileHashTablepp[i];
1420 lock_ReleaseWrite(&buf_globalLock);
1421 for (; bp; bp = nbp) {
1422 /* clean buffer synchronously */
1423 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1426 lock_ObtainMutex(&bp->mx);
1428 cm_ReleaseUser(bp->userp);
1430 lock_ReleaseMutex(&bp->mx);
1432 buf_CleanAsync(bp, reqp);
1433 buf_CleanWait(scp, bp);
1434 lock_ObtainMutex(&bp->mx);
1435 if (bp->flags & CM_BUF_ERROR) {
1436 if (code == 0 || code == -1)
1441 lock_ReleaseMutex(&bp->mx);
1444 lock_ObtainWrite(&buf_globalLock);
1445 buf_LockedRelease(bp);
1446 nbp = bp->fileHashp;
1449 lock_ReleaseWrite(&buf_globalLock);
1450 } /* for loop over a bunch of buffers */
1453 buf_ValidateBufQueues();
1454 #endif /* TESTING */
1462 buf_ValidateBufQueues(void)
1464 cm_buf_t * bp, *bpb, *bpf, *bpa;
1465 afs_uint32 countf=0, countb=0, counta=0;
1467 lock_ObtainRead(&buf_globalLock);
1468 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1469 if (bp->magic != CM_BUF_MAGIC)
1475 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1476 if (bp->magic != CM_BUF_MAGIC)
1482 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1483 if (bp->magic != CM_BUF_MAGIC)
1488 lock_ReleaseRead(&buf_globalLock);
1490 if (countb != countf)
1493 if (counta != cm_data.buf_nbuffers)
1496 #endif /* TESTING */
1498 /* dump the contents of the buf_hashTablepp. */
1499 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1506 if (cm_data.buf_hashTablepp == NULL)
1510 lock_ObtainRead(&buf_globalLock);
1512 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\n",
1513 cookie, cm_data.buf_hashSize);
1514 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1516 for (i = 0; i < cm_data.buf_hashSize; i++)
1518 for (bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp)
1522 StringCbPrintfA(output, sizeof(output), "vnode=%d, unique=%d), size=%d refCount=%d\n",
1523 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1524 bp->fid.vnode, bp->fid.unique, bp->size, bp->refCount);
1525 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1530 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\n", cookie);
1531 WriteFile(outputFile, output, strlen(output), &zilch, NULL);
1534 lock_ReleaseRead(&buf_globalLock);
1538 void buf_ForceTrace(BOOL flush)
1547 len = GetTempPath(sizeof(buf)-10, buf);
1548 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1549 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1550 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1551 if (handle == INVALID_HANDLE_VALUE) {
1552 osi_panic("Cannot create log file", __FILE__, __LINE__);
1554 osi_LogPrint(buf_logp, handle);
1556 FlushFileBuffers(handle);
1557 CloseHandle(handle);