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>
25 #include "cm_memmap.h"
28 #define TRACE_BUFFER 1
31 extern void afsi_log(char *pattern, ...);
33 /* This module implements the buffer package used by the local transaction
34 * system (cm). It is initialized by calling cm_Init, which calls buf_Init;
35 * it must be initalized before any of its main routines are called.
37 * Each buffer is hashed into a hash table by file ID and offset, and if its
38 * reference count is zero, it is also in a free list.
40 * There are two locks involved in buffer processing. The global lock
41 * buf_globalLock protects all of the global variables defined in this module,
42 * the reference counts and hash pointers in the actual cm_buf_t structures,
43 * and the LRU queue pointers in the buffer structures.
45 * The mutexes in the buffer structures protect the remaining fields in the
46 * buffers, as well the data itself.
48 * The locking hierarchy here is this:
50 * - resv multiple simul. buffers reservation
51 * - lock buffer I/O flags
52 * - lock buffer's mutex
53 * - lock buf_globalLock
57 /* global debugging log */
58 osi_log_t *buf_logp = NULL;
60 /* Global lock protecting hash tables and free lists */
61 osi_rwlock_t buf_globalLock;
63 /* ptr to head of the free list (most recently used) and the
64 * tail (the guy to remove first). We use osi_Q* functions
65 * to put stuff in buf_freeListp, and maintain the end
69 /* a pointer to a list of all buffers, just so that we can find them
70 * easily for debugging, and for the incr syncer. Locked under
74 /* defaults setup; these variables may be manually assigned into
75 * before calling cm_Init, as a way of changing these defaults.
78 /* callouts for reading and writing data, etc */
79 cm_buf_ops_t *cm_buf_opsp;
82 /* for experimental disk caching support in Win95 client */
83 cm_buf_t *buf_diskFreeListp;
84 cm_buf_t *buf_diskFreeListEndp;
85 cm_buf_t *buf_diskAllp;
86 extern int cm_diskCacheEnabled;
87 #endif /* DISKCACHE95 */
89 /* set this to 1 when we are terminating to prevent access attempts */
90 static int buf_ShutdownFlag = 0;
92 /* hold a reference to an already held buffer */
93 void buf_Hold(cm_buf_t *bp)
95 osi_assert(bp->magic == CM_BUF_MAGIC);
96 lock_ObtainWrite(&buf_globalLock);
98 lock_ReleaseWrite(&buf_globalLock);
101 /* incremental sync daemon. Writes 1/10th of all the buffers every 5000 ms */
102 void buf_IncrSyncer(long parm)
104 cm_buf_t *bp; /* buffer we're hacking on; held */
105 long i; /* counter */
106 long nAtOnce; /* how many to do at once */
109 lock_ObtainWrite(&buf_globalLock);
110 bp = cm_data.buf_allp;
112 lock_ReleaseWrite(&buf_globalLock);
113 nAtOnce = (long)sqrt((double)cm_data.buf_nbuffers);
114 while (buf_ShutdownFlag == 0) {
116 i = SleepEx(5000, 1);
117 if (i != 0) continue;
122 if (buf_ShutdownFlag == 1)
125 /* now go through our percentage of the buffers */
126 for (i=0; i<nAtOnce; i++) {
127 /* don't want its identity changing while we're
128 * messing with it, so must do all of this with
132 /* start cleaning the buffer; don't touch log pages since
133 * the log code counts on knowing exactly who is writing
134 * a log page at any given instant.
137 req.flags |= CM_REQ_NORETRY;
138 buf_CleanAsync(bp, &req);
140 /* now advance to the next buffer; the allp chain never changes,
141 * and so can be followed even when holding no locks.
143 lock_ObtainWrite(&buf_globalLock);
144 buf_LockedRelease(bp);
147 bp = cm_data.buf_allp;
149 lock_ReleaseWrite(&buf_globalLock);
150 } /* for loop over a bunch of buffers */
151 } /* whole daemon's while loop */
155 buf_ValidateBuffers(void)
157 cm_buf_t * bp, *bpf, *bpa, *bpb;
158 afs_uint64 countb = 0, countf = 0, counta = 0;
160 if (cm_data.buf_freeListp == NULL && cm_data.buf_freeListEndp != NULL ||
161 cm_data.buf_freeListp != NULL && cm_data.buf_freeListEndp == NULL) {
162 afsi_log("cm_ValidateBuffers failure: inconsistent free list pointers");
163 fprintf(stderr, "cm_ValidateBuffers failure: inconsistent free list pointers\n");
167 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
168 if (bp->magic != CM_BUF_MAGIC) {
169 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
170 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
176 if (countb > cm_data.buf_nbuffers) {
177 afsi_log("cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers");
178 fprintf(stderr, "cm_ValidateBuffers failure: countb > cm_data.buf_nbuffers\n");
183 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
184 if (bp->magic != CM_BUF_MAGIC) {
185 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
186 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
192 if (countf > cm_data.buf_nbuffers) {
193 afsi_log("cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers");
194 fprintf(stderr, "cm_ValidateBuffers failure: countf > cm_data.buf_nbuffers\n");
199 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
200 if (bp->magic != CM_BUF_MAGIC) {
201 afsi_log("cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC");
202 fprintf(stderr, "cm_ValidateBuffers failure: bp->magic != CM_BUF_MAGIC\n");
208 if (counta > cm_data.buf_nbuffers) {
209 afsi_log("cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers");
210 fprintf(stderr, "cm_ValidateBuffers failure: counta > cm_data.buf_nbuffers\n");
215 if (countb != countf) {
216 afsi_log("cm_ValidateBuffers failure: countb != countf");
217 fprintf(stderr, "cm_ValidateBuffers failure: countb != countf\n");
221 if (counta != cm_data.buf_nbuffers) {
222 afsi_log("cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers");
223 fprintf(stderr, "cm_ValidateBuffers failure: counta != cm_data.buf_nbuffers\n");
230 void buf_Shutdown(void)
232 buf_ShutdownFlag = 1;
235 /* initialize the buffer package; called with no locks
236 * held during the initialization phase.
238 long buf_Init(int newFile, cm_buf_ops_t *opsp, afs_uint64 nbuffers)
240 static osi_once_t once;
249 cm_data.buf_nbuffers = nbuffers;
251 /* Have to be able to reserve a whole chunk */
252 if (((cm_data.buf_nbuffers - 3) * cm_data.buf_blockSize) < cm_chunkSize)
253 return CM_ERROR_TOOFEWBUFS;
256 /* recall for callouts */
259 if (osi_Once(&once)) {
260 /* initialize global locks */
261 lock_InitializeRWLock(&buf_globalLock, "Global buffer lock");
264 /* remember this for those who want to reset it */
265 cm_data.buf_nOrigBuffers = cm_data.buf_nbuffers;
267 /* lower hash size to a prime number */
268 cm_data.buf_hashSize = osi_PrimeLessThan(CM_BUF_HASHSIZE);
270 /* create hash table */
271 memset((void *)cm_data.buf_hashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
273 /* another hash table */
274 memset((void *)cm_data.buf_fileHashTablepp, 0, cm_data.buf_hashSize * sizeof(cm_buf_t *));
276 /* create buffer headers and put in free list */
277 bp = cm_data.bufHeaderBaseAddress;
278 data = cm_data.bufDataBaseAddress;
279 cm_data.buf_allp = NULL;
281 for (i=0; i<cm_data.buf_nbuffers; i++) {
282 osi_assert(bp >= cm_data.bufHeaderBaseAddress && bp < (cm_buf_t *)cm_data.bufDataBaseAddress);
283 osi_assert(data >= cm_data.bufDataBaseAddress && data < cm_data.bufEndOfData);
285 /* allocate and zero some storage */
286 memset(bp, 0, sizeof(cm_buf_t));
287 bp->magic = CM_BUF_MAGIC;
288 /* thread on list of all buffers */
289 bp->allp = cm_data.buf_allp;
290 cm_data.buf_allp = bp;
292 osi_QAdd((osi_queue_t **)&cm_data.buf_freeListp, &bp->q);
293 bp->flags |= CM_BUF_INLRU;
294 lock_InitializeMutex(&bp->mx, "Buffer mutex");
296 /* grab appropriate number of bytes from aligned zone */
299 /* setup last buffer pointer */
301 cm_data.buf_freeListEndp = bp;
305 data += cm_data.buf_blockSize;
308 /* none reserved at first */
309 cm_data.buf_reservedBufs = 0;
311 /* just for safety's sake */
312 cm_data.buf_maxReservedBufs = cm_data.buf_nbuffers - 3;
314 bp = cm_data.bufHeaderBaseAddress;
315 data = cm_data.bufDataBaseAddress;
317 for (i=0; i<cm_data.buf_nbuffers; i++) {
318 lock_InitializeMutex(&bp->mx, "Buffer mutex");
321 bp->waitRequests = 0;
322 bp->flags &= ~CM_BUF_WAITING;
328 buf_ValidateBufQueues();
332 /* init the buffer trace log */
333 buf_logp = osi_LogCreate("buffer", 1000);
334 osi_LogEnable(buf_logp);
339 /* and create the incr-syncer */
340 phandle = thrd_Create(0, 0,
341 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
344 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
346 CloseHandle(phandle);
351 buf_ValidateBufQueues();
356 /* add nbuffers to the buffer pool, if possible.
357 * Called with no locks held.
359 long buf_AddBuffers(afs_uint64 nbuffers)
362 /* The size of a virtual cache cannot be changed after it has
363 * been created. Subsequent calls to MapViewofFile() with
364 * an existing mapping object name would not allow the
365 * object to be resized. Return failure immediately.
367 * A similar problem now occurs with the persistent cache
368 * given that the memory mapped file now contains a complex
371 afsi_log("request to add %d buffers to the existing cache of size %d denied",
372 nbuffers, cm_data.buf_nbuffers);
374 return CM_ERROR_INVAL;
380 data = malloc(buf_nbuffers * cm_data.buf_blockSize);
382 /* Create buffer headers and put in free list */
383 bp = malloc(nbuffers * sizeof(*bp));
385 for (i=0; i<nbuffers; i++) {
386 memset(bp, 0, sizeof(*bp));
388 lock_InitializeMutex(&bp->mx, "cm_buf_t");
390 /* grab appropriate number of bytes from aligned zone */
393 bp->flags |= CM_BUF_INLRU;
395 lock_ObtainWrite(&buf_globalLock);
396 /* note that buf_allp chain is covered by buf_globalLock now */
397 bp->allp = cm_data.buf_allp;
398 cm_data.buf_allp = bp;
399 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
400 if (!cm_data.buf_freeListEndp)
401 cm_data.buf_freeListEndp = bp;
402 cm_data.buf_nbuffers++;
403 lock_ReleaseWrite(&buf_globalLock);
406 data += cm_data.buf_blockSize;
408 } /* for loop over all buffers */
414 /* interface to set the number of buffers to an exact figure.
415 * Called with no locks held.
417 long buf_SetNBuffers(afs_uint64 nbuffers)
420 return CM_ERROR_INVAL;
421 if (nbuffers == cm_data.buf_nbuffers)
423 else if (nbuffers > cm_data.buf_nbuffers)
424 return buf_AddBuffers(nbuffers - cm_data.buf_nbuffers);
426 return CM_ERROR_INVAL;
429 /* release a buffer. Buffer must be referenced, but unlocked. */
430 void buf_Release(cm_buf_t *bp)
432 lock_ObtainWrite(&buf_globalLock);
433 buf_LockedRelease(bp);
434 lock_ReleaseWrite(&buf_globalLock);
437 /* wait for reading or writing to clear; called with write-locked
438 * buffer and unlocked scp and returns with locked buffer.
440 void buf_WaitIO(cm_scache_t * scp, cm_buf_t *bp)
443 osi_assert(scp->magic == CM_SCACHE_MAGIC);
444 osi_assert(bp->magic == CM_BUF_MAGIC);
447 /* if no IO is happening, we're done */
448 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
451 /* otherwise I/O is happening, but some other thread is waiting for
452 * the I/O already. Wait for that guy to figure out what happened,
453 * and then check again.
455 if ( bp->flags & CM_BUF_WAITING ) {
458 osi_Log1(afsd_logp, "buf_WaitIO CM_BUF_WAITING already set for 0x%p", bp);
460 osi_Log1(afsd_logp, "buf_WaitIO CM_BUF_WAITING set for 0x%p", bp);
461 bp->flags |= CM_BUF_WAITING;
462 bp->waitCount = bp->waitRequests = 1;
464 osi_SleepM((LONG_PTR)bp, &bp->mx);
466 smb_UpdateServerPriority();
468 lock_ObtainMutex(&bp->mx);
469 osi_Log1(afsd_logp, "buf_WaitIO conflict wait done for 0x%p", bp);
471 if (bp->waitCount == 0) {
472 osi_Log1(afsd_logp, "buf_WaitIO CM_BUF_WAITING reset for 0x%p", bp);
473 bp->flags &= ~CM_BUF_WAITING;
474 bp->waitRequests = 0;
478 scp = cm_FindSCache(&bp->fid);
481 lock_ObtainMutex(&scp->mx);
482 if (scp->flags & CM_SCACHEFLAG_WAITING) {
483 osi_Log1(afsd_logp, "buf_WaitIO waking scp 0x%p", scp);
484 osi_Wakeup((LONG_PTR)&scp->flags);
486 lock_ReleaseMutex(&scp->mx);
490 /* if we get here, the IO is done, but we may have to wakeup people waiting for
491 * the I/O to complete. Do so.
493 if (bp->flags & CM_BUF_WAITING) {
494 osi_Log1(afsd_logp, "buf_WaitIO Waking bp 0x%p", bp);
495 osi_Wakeup((LONG_PTR) bp);
497 osi_Log1(afsd_logp, "WaitIO finished wait for bp 0x%p", bp);
500 /* code to drop reference count while holding buf_globalLock */
501 void buf_LockedRelease(cm_buf_t *bp)
503 /* ensure that we're in the LRU queue if our ref count is 0 */
504 osi_assert(bp->refCount > 0);
505 if (--bp->refCount == 0) {
506 if (!(bp->flags & CM_BUF_INLRU)) {
507 osi_QAdd((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
509 /* watch for transition from empty to one element */
510 if (!cm_data.buf_freeListEndp)
511 cm_data.buf_freeListEndp = cm_data.buf_freeListp;
512 bp->flags |= CM_BUF_INLRU;
517 /* find a buffer, if any, for a particular file ID and offset. Assumes
518 * that buf_globalLock is write locked when called.
520 cm_buf_t *buf_LockedFind(struct cm_scache *scp, osi_hyper_t *offsetp)
525 i = BUF_HASH(&scp->fid, offsetp);
526 for(bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp) {
527 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
528 && offsetp->LowPart == bp->offset.LowPart
529 && offsetp->HighPart == bp->offset.HighPart) {
535 /* return whatever we found, if anything */
539 /* find a buffer with offset *offsetp for vnode *scp. Called
540 * with no locks held.
542 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
546 lock_ObtainWrite(&buf_globalLock);
547 bp = buf_LockedFind(scp, offsetp);
548 lock_ReleaseWrite(&buf_globalLock);
553 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
556 * Makes sure that there's only one person writing this block
557 * at any given time, and also ensures that the log is forced sufficiently far,
558 * if this buffer contains logged data.
560 void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
564 osi_assert(bp->magic == CM_BUF_MAGIC);
566 while ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
567 lock_ReleaseMutex(&bp->mx);
569 osi_Log1(afsd_logp, "buf_LockedCleanAsync starts I/O on 0x%p", bp);
570 code = (*cm_buf_opsp->Writep)(&bp->fid, &bp->offset,
571 cm_data.buf_blockSize, 0, bp->userp,
573 osi_Log2(afsd_logp, "buf_LockedCleanAsync I/O on 0x%p, done=%d", bp, code);
575 lock_ObtainMutex(&bp->mx);
580 /* Disk cache support */
581 /* write buffer to disk cache (synchronous for now) */
582 diskcache_Update(bp->dcp, bp->datap, cm_data.buf_blockSize, bp->dataVersion);
583 #endif /* DISKCACHE95 */
586 /* do logging after call to GetLastError, or else */
588 /* if someone was waiting for the I/O that just completed or failed,
591 if (bp->flags & CM_BUF_WAITING) {
592 /* turn off flags and wakeup users */
593 osi_Log1(buf_logp, "buf_WaitIO Waking bp 0x%p", bp);
594 osi_Wakeup((LONG_PTR) bp);
598 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
599 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
600 * The buffer must already be clean, and no I/O should be happening to it.
602 void buf_Recycle(cm_buf_t *bp)
607 cm_buf_t *prevBp, *nextBp;
609 osi_assert(bp->magic == CM_BUF_MAGIC);
611 /* if we get here, we know that the buffer still has a 0 ref count,
612 * and that it is clean and has no currently pending I/O. This is
613 * the dude to return.
614 * Remember that as long as the ref count is 0, we know that we won't
615 * have any lock conflicts, so we can grab the buffer lock out of
616 * order in the locking hierarchy.
618 osi_Log2( buf_logp, "buf_Recycle recycles 0x%p, off 0x%x",
619 bp, bp->offset.LowPart);
621 osi_assert(bp->refCount == 0);
622 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)));
623 lock_AssertWrite(&buf_globalLock);
625 if (bp->flags & CM_BUF_INHASH) {
626 /* Remove from hash */
628 i = BUF_HASH(&bp->fid, &bp->offset);
629 lbpp = &(cm_data.buf_hashTablepp[i]);
630 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
631 if (tbp == bp) break;
634 /* we better find it */
635 osi_assertx(tbp != NULL, "buf_Recycle: hash table screwup");
637 *lbpp = bp->hashp; /* hash out */
639 /* Remove from file hash */
641 i = BUF_FILEHASH(&bp->fid);
642 prevBp = bp->fileHashBackp;
643 nextBp = bp->fileHashp;
645 prevBp->fileHashp = nextBp;
647 cm_data.buf_fileHashTablepp[i] = nextBp;
649 nextBp->fileHashBackp = prevBp;
651 bp->flags &= ~CM_BUF_INHASH;
654 /* bump the soft reference counter now, to invalidate softRefs; no
655 * wakeup is required since people don't sleep waiting for this
660 /* make the fid unrecognizable */
661 memset(&bp->fid, 0, sizeof(cm_fid_t));
664 /* recycle a buffer, removing it from the free list, hashing in its new identity
665 * and returning it write-locked so that no one can use it. Called without
666 * any locks held, and can return an error if it loses the race condition and
667 * finds that someone else created the desired buffer.
669 * If success is returned, the buffer is returned write-locked.
671 * May be called with null scp and offsetp, if we're just trying to reclaim some
672 * space from the buffer pool. In that case, the buffer will be returned
673 * without being hashed into the hash table.
675 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
677 cm_buf_t *bp; /* buffer we're dealing with */
678 cm_buf_t *nextBp; /* next buffer in file hash chain */
682 cm_InitReq(&req); /* just in case */
685 buf_ValidateBufQueues();
690 lock_ObtainWrite(&buf_globalLock);
691 /* check to see if we lost the race */
693 if (bp = buf_LockedFind(scp, offsetp)) {
695 lock_ReleaseWrite(&buf_globalLock);
696 return CM_BUF_EXISTS;
700 /* does this fix the problem below? it's a simple solution. */
701 if (!cm_data.buf_freeListEndp)
703 lock_ReleaseWrite(&buf_globalLock);
708 /* for debugging, assert free list isn't empty, although we
709 * really should try waiting for a running tranasction to finish
710 * instead of this; or better, we should have a transaction
711 * throttler prevent us from entering this situation.
713 osi_assertx(cm_data.buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
715 /* look at all buffers in free list, some of which may temp.
716 * have high refcounts and which then should be skipped,
717 * starting cleaning I/O for those which are dirty. If we find
718 * a clean buffer, we rehash it, lock it and return it.
720 for(bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
721 /* check to see if it really has zero ref count. This
722 * code can bump refcounts, at least, so it may not be
725 if (bp->refCount > 0)
728 /* we don't have to lock buffer itself, since the ref
729 * count is 0 and we know it will stay zero as long as
730 * we hold the global lock.
733 /* don't recycle someone in our own chunk */
734 if (!cm_FidCmp(&bp->fid, &scp->fid)
735 && (bp->offset.LowPart & (-cm_chunkSize))
736 == (offsetp->LowPart & (-cm_chunkSize)))
739 /* if this page is being filled (!) or cleaned, see if
740 * the I/O has completed. If not, skip it, otherwise
741 * do the final processing for the I/O.
743 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
744 /* probably shouldn't do this much work while
745 * holding the big lock? Watch for contention
751 if (bp->flags & CM_BUF_DIRTY) {
752 /* if the buffer is dirty, start cleaning it and
753 * move on to the next buffer. We do this with
754 * just the lock required to minimize contention
758 lock_ReleaseWrite(&buf_globalLock);
760 /* grab required lock and clean; this only
761 * starts the I/O. By the time we're back,
762 * it'll still be marked dirty, but it will also
763 * have the WRITING flag set, so we won't get
766 buf_CleanAsync(bp, &req);
768 /* now put it back and go around again */
773 /* if we get here, we know that the buffer still has a 0
774 * ref count, and that it is clean and has no currently
775 * pending I/O. This is the dude to return.
776 * Remember that as long as the ref count is 0, we know
777 * that we won't have any lock conflicts, so we can grab
778 * the buffer lock out of order in the locking hierarchy.
782 /* clean up junk flags */
783 bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
784 bp->dataVersion = -1; /* unknown so far */
786 /* now hash in as our new buffer, and give it the
787 * appropriate label, if requested.
790 bp->flags |= CM_BUF_INHASH;
792 bp->offset = *offsetp;
793 i = BUF_HASH(&scp->fid, offsetp);
794 bp->hashp = cm_data.buf_hashTablepp[i];
795 cm_data.buf_hashTablepp[i] = bp;
796 i = BUF_FILEHASH(&scp->fid);
797 nextBp = cm_data.buf_fileHashTablepp[i];
798 bp->fileHashp = nextBp;
799 bp->fileHashBackp = NULL;
801 nextBp->fileHashBackp = bp;
802 cm_data.buf_fileHashTablepp[i] = bp;
805 /* prepare to return it. Start by giving it a good
809 /* and since it has a non-zero ref count, we should move
810 * it from the lru queue. It better be still there,
811 * since we've held the global (big) lock since we found
814 osi_assertx(bp->flags & CM_BUF_INLRU,
815 "buf_GetNewLocked: LRU screwup");
816 if (cm_data.buf_freeListEndp == bp) {
817 /* we're the last guy in this queue, so maintain it */
818 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
820 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
821 bp->flags &= ~CM_BUF_INLRU;
823 /* finally, grab the mutex so that people don't use it
824 * before the caller fills it with data. Again, no one
825 * should have been able to get to this dude to lock it.
827 osi_assertx(lock_TryMutex(&bp->mx),
828 "buf_GetNewLocked: TryMutex failed");
830 lock_ReleaseWrite(&buf_globalLock);
834 buf_ValidateBufQueues();
837 } /* for all buffers in lru queue */
838 lock_ReleaseWrite(&buf_globalLock);
839 } /* while loop over everything */
843 /* get a page, returning it held but unlocked. Doesn't fill in the page
844 * with I/O, since we're going to write the whole thing new.
846 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
850 osi_hyper_t pageOffset;
854 pageOffset.HighPart = offsetp->HighPart;
855 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
857 bp = buf_Find(scp, &pageOffset);
859 /* lock it and break out */
860 lock_ObtainMutex(&bp->mx);
864 /* otherwise, we have to create a page */
865 code = buf_GetNewLocked(scp, &pageOffset, &bp);
867 /* check if the buffer was created in a race condition branch.
868 * If so, go around so we can hold a reference to it.
870 if (code == CM_BUF_EXISTS)
873 /* something else went wrong */
877 /* otherwise, we have a locked buffer that we just created */
880 } /* big while loop */
883 if (bp->flags & CM_BUF_READING)
886 /* once it has been read once, we can unlock it and return it, still
887 * with its refcount held.
889 lock_ReleaseMutex(&bp->mx);
891 osi_Log3(buf_logp, "buf_GetNew returning bp 0x%p for scp 0x%p, offset 0x%x",
892 bp, scp, offsetp->LowPart);
896 /* get a page, returning it held but unlocked. Make sure it is complete */
897 /* The scp must be unlocked when passed to this function */
898 long buf_Get(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
902 osi_hyper_t pageOffset;
903 unsigned long tcount;
908 #endif /* DISKCACHE95 */
911 pageOffset.HighPart = offsetp->HighPart;
912 pageOffset.LowPart = offsetp->LowPart & ~(cm_data.buf_blockSize-1);
916 buf_ValidateBufQueues();
919 bp = buf_Find(scp, &pageOffset);
921 /* lock it and break out */
922 lock_ObtainMutex(&bp->mx);
926 /* touch disk chunk to update LRU info */
927 diskcache_Touch(bp->dcp);
928 #endif /* DISKCACHE95 */
931 /* otherwise, we have to create a page */
932 code = buf_GetNewLocked(scp, &pageOffset, &bp);
934 /* check if the buffer was created in a race condition branch.
935 * If so, go around so we can hold a reference to it.
937 if (code == CM_BUF_EXISTS)
940 /* something else went wrong */
943 buf_ValidateBufQueues();
948 /* otherwise, we have a locked buffer that we just created */
951 } /* big while loop */
953 /* if we get here, we have a locked buffer that may have just been
954 * created, in which case it needs to be filled with data.
957 /* load the page; freshly created pages should be idle */
958 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)));
960 /* setup offset, event */
961 #ifndef DJGPP /* doesn't seem to be used */
962 bp->over.Offset = bp->offset.LowPart;
963 bp->over.OffsetHigh = bp->offset.HighPart;
966 /* start the I/O; may drop lock */
967 bp->flags |= CM_BUF_READING;
968 code = (*cm_buf_opsp->Readp)(bp, cm_data.buf_blockSize, &tcount, NULL);
971 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, cm_data.buf_blockSize, &bp->dataVersion, &tcount, &dcp);
972 bp->dcp = dcp; /* pointer to disk cache struct. */
973 #endif /* DISKCACHE95 */
976 /* failure or queued */
977 #ifndef DJGPP /* cm_bufRead always returns 0 */
978 if (code != ERROR_IO_PENDING) {
981 bp->flags |= CM_BUF_ERROR;
982 bp->flags &= ~CM_BUF_READING;
983 if (bp->flags & CM_BUF_WAITING) {
984 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
985 osi_Wakeup((LONG_PTR) bp);
987 lock_ReleaseMutex(&bp->mx);
990 buf_ValidateBufQueues();
997 /* otherwise, I/O completed instantly and we're done, except
998 * for padding the xfr out with 0s and checking for EOF
1000 if (tcount < (unsigned long) cm_data.buf_blockSize) {
1001 memset(bp->datap+tcount, 0, cm_data.buf_blockSize - tcount);
1003 bp->flags |= CM_BUF_EOF;
1005 bp->flags &= ~CM_BUF_READING;
1006 if (bp->flags & CM_BUF_WAITING) {
1007 osi_Log1(buf_logp, "buf_Get Waking bp 0x%p", bp);
1008 osi_Wakeup((LONG_PTR) bp);
1014 /* wait for reads, either that which we started above, or that someone
1015 * else started. We don't care if we return a buffer being cleaned.
1017 if (bp->flags & CM_BUF_READING)
1018 buf_WaitIO(scp, bp);
1020 /* once it has been read once, we can unlock it and return it, still
1021 * with its refcount held.
1023 lock_ReleaseMutex(&bp->mx);
1026 /* now remove from queue; will be put in at the head (farthest from
1027 * being recycled) when we're done in buf_Release.
1029 lock_ObtainWrite(&buf_globalLock);
1030 if (bp->flags & CM_BUF_INLRU) {
1031 if (cm_data.buf_freeListEndp == bp)
1032 cm_data.buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1033 osi_QRemove((osi_queue_t **) &cm_data.buf_freeListp, &bp->q);
1034 bp->flags &= ~CM_BUF_INLRU;
1036 lock_ReleaseWrite(&buf_globalLock);
1038 osi_Log3(buf_logp, "buf_Get returning bp 0x%p for scp 0x%p, offset 0x%x",
1039 bp, scp, offsetp->LowPart);
1041 buf_ValidateBufQueues();
1042 #endif /* TESTING */
1046 /* count # of elements in the free list;
1047 * we don't bother doing the proper locking for accessing dataVersion or flags
1048 * since it is a pain, and this is really just an advisory call. If you need
1049 * to do better at some point, rewrite this function.
1051 long buf_CountFreeList(void)
1057 lock_ObtainRead(&buf_globalLock);
1058 for(bufp = cm_data.buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1059 /* if the buffer doesn't have an identity, or if the buffer
1060 * has been invalidate (by having its DV stomped upon), then
1061 * count it as free, since it isn't really being utilized.
1063 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion <= 0)
1066 lock_ReleaseRead(&buf_globalLock);
1070 /* clean a buffer synchronously */
1071 void buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
1073 osi_assert(bp->magic == CM_BUF_MAGIC);
1075 lock_ObtainMutex(&bp->mx);
1076 buf_LockedCleanAsync(bp, reqp);
1077 lock_ReleaseMutex(&bp->mx);
1080 /* wait for a buffer's cleaning to finish */
1081 void buf_CleanWait(cm_scache_t * scp, cm_buf_t *bp)
1083 osi_assert(bp->magic == CM_BUF_MAGIC);
1085 lock_ObtainMutex(&bp->mx);
1086 if (bp->flags & CM_BUF_WRITING) {
1087 buf_WaitIO(scp, bp);
1089 lock_ReleaseMutex(&bp->mx);
1092 /* set the dirty flag on a buffer, and set associated write-ahead log,
1093 * if there is one. Allow one to be added to a buffer, but not changed.
1095 * The buffer must be locked before calling this routine.
1097 void buf_SetDirty(cm_buf_t *bp)
1099 osi_assert(bp->magic == CM_BUF_MAGIC);
1100 osi_assert(bp->refCount > 0);
1102 osi_Log1(buf_logp, "buf_SetDirty 0x%p", bp);
1105 bp->flags |= CM_BUF_DIRTY;
1107 /* and turn off EOF flag, since it has associated data now */
1108 bp->flags &= ~CM_BUF_EOF;
1111 /* clean all buffers, reset log pointers and invalidate all buffers.
1112 * Called with no locks held, and returns with same.
1114 * This function is guaranteed to clean and remove the log ptr of all the
1115 * buffers that were dirty or had non-zero log ptrs before the call was
1116 * made. That's sufficient to clean up any garbage left around by recovery,
1117 * which is all we're counting on this for; there may be newly created buffers
1118 * added while we're running, but that should be OK.
1120 * In an environment where there are no transactions (artificially imposed, for
1121 * example, when switching the database to raw mode), this function is used to
1122 * make sure that all updates have been written to the disk. In that case, we don't
1123 * really require that we forget the log association between pages and logs, but
1124 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1125 * have to worry about invalidating data in the buffers.
1127 * This function is used at the end of recovery as paranoia to get the recovered
1128 * database out to disk. It removes all references to the recovery log and cleans
1131 long buf_CleanAndReset(void)
1137 lock_ObtainWrite(&buf_globalLock);
1138 for(i=0; i<cm_data.buf_hashSize; i++) {
1139 for(bp = cm_data.buf_hashTablepp[i]; bp; bp = bp->hashp) {
1140 if ((bp->flags & CM_BUF_DIRTY) == CM_BUF_DIRTY) {
1142 lock_ReleaseWrite(&buf_globalLock);
1144 /* now no locks are held; clean buffer and go on */
1146 buf_CleanAsync(bp, &req);
1147 buf_CleanWait(NULL, bp);
1149 /* relock and release buffer */
1150 lock_ObtainWrite(&buf_globalLock);
1151 buf_LockedRelease(bp);
1153 } /* over one bucket */
1154 } /* for loop over all hash buckets */
1157 lock_ReleaseWrite(&buf_globalLock);
1160 buf_ValidateBufQueues();
1161 #endif /* TESTING */
1163 /* and we're done */
1167 /* called without global lock being held, reserves buffers for callers
1168 * that need more than one held (not locked) at once.
1170 void buf_ReserveBuffers(afs_uint64 nbuffers)
1172 lock_ObtainWrite(&buf_globalLock);
1174 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1175 cm_data.buf_reserveWaiting = 1;
1176 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1177 osi_SleepW((LONG_PTR) &cm_data.buf_reservedBufs, &buf_globalLock);
1178 lock_ObtainWrite(&buf_globalLock);
1181 cm_data.buf_reservedBufs += nbuffers;
1185 lock_ReleaseWrite(&buf_globalLock);
1188 int buf_TryReserveBuffers(afs_uint64 nbuffers)
1192 lock_ObtainWrite(&buf_globalLock);
1193 if (cm_data.buf_reservedBufs + nbuffers > cm_data.buf_maxReservedBufs) {
1197 cm_data.buf_reservedBufs += nbuffers;
1200 lock_ReleaseWrite(&buf_globalLock);
1204 /* called without global lock held, releases reservation held by
1205 * buf_ReserveBuffers.
1207 void buf_UnreserveBuffers(afs_uint64 nbuffers)
1209 lock_ObtainWrite(&buf_globalLock);
1210 cm_data.buf_reservedBufs -= nbuffers;
1211 if (cm_data.buf_reserveWaiting) {
1212 cm_data.buf_reserveWaiting = 0;
1213 osi_Wakeup((LONG_PTR) &cm_data.buf_reservedBufs);
1215 lock_ReleaseWrite(&buf_globalLock);
1218 /* truncate the buffers past sizep, zeroing out the page, if we don't
1219 * end on a page boundary.
1221 * Requires cm_bufCreateLock to be write locked.
1223 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1227 cm_buf_t *nbufp; /* next buffer, if didRelease */
1234 /* assert that cm_bufCreateLock is held in write mode */
1235 lock_AssertWrite(&scp->bufCreateLock);
1237 i = BUF_FILEHASH(&scp->fid);
1239 lock_ObtainWrite(&buf_globalLock);
1240 bufp = cm_data.buf_fileHashTablepp[i];
1242 lock_ReleaseWrite(&buf_globalLock);
1247 lock_ReleaseWrite(&buf_globalLock);
1248 for(; bufp; bufp = nbufp) {
1250 lock_ObtainMutex(&bufp->mx);
1252 bufEnd.HighPart = 0;
1253 bufEnd.LowPart = cm_data.buf_blockSize;
1254 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1256 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1257 LargeIntegerLessThan(*sizep, bufEnd)) {
1258 buf_WaitIO(scp, bufp);
1260 lock_ObtainMutex(&scp->mx);
1262 /* make sure we have a callback (so we have the right value for
1263 * the length), and wait for it to be safe to do a truncate.
1265 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1266 CM_SCACHESYNC_NEEDCALLBACK
1267 | CM_SCACHESYNC_GETSTATUS
1268 | CM_SCACHESYNC_SETSIZE
1269 | CM_SCACHESYNC_BUFLOCKED);
1270 /* if we succeeded in our locking, and this applies to the right
1271 * file, and the truncate request overlaps the buffer either
1272 * totally or partially, then do something.
1274 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1275 && LargeIntegerLessThan(*sizep, bufEnd)) {
1277 lock_ObtainWrite(&buf_globalLock);
1279 /* destroy the buffer, turning off its dirty bit, if
1280 * we're truncating the whole buffer. Otherwise, set
1281 * the dirty bit, and clear out the tail of the buffer
1282 * if we just overlap some.
1284 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1285 /* truncating the entire page */
1286 bufp->flags &= ~CM_BUF_DIRTY;
1287 bufp->dataVersion = -1; /* known bad */
1288 bufp->dirtyCounter++;
1291 /* don't set dirty, since dirty implies
1292 * currently up-to-date. Don't need to do this,
1293 * since we'll update the length anyway.
1295 * Zero out remainder of the page, in case we
1296 * seek and write past EOF, and make this data
1299 bufferPos = sizep->LowPart & (cm_data.buf_blockSize - 1);
1300 osi_assert(bufferPos != 0);
1301 memset(bufp->datap + bufferPos, 0,
1302 cm_data.buf_blockSize - bufferPos);
1305 lock_ReleaseWrite(&buf_globalLock);
1308 lock_ReleaseMutex(&scp->mx);
1309 lock_ReleaseMutex(&bufp->mx);
1311 lock_ObtainWrite(&buf_globalLock);
1312 nbufp = bufp->fileHashp;
1313 if (nbufp) nbufp->refCount++;
1314 buf_LockedRelease(bufp);
1315 lock_ReleaseWrite(&buf_globalLock);
1318 /* bail out early if we fail */
1320 /* at this point, nbufp is held; bufp has already been
1327 buf_ValidateBufQueues();
1328 #endif /* TESTING */
1335 buf_ValidateBufQueues();
1336 #endif /* TESTING */
1342 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1345 cm_buf_t *bp; /* buffer we're hacking on */
1350 i = BUF_FILEHASH(&scp->fid);
1353 lock_ObtainWrite(&buf_globalLock);
1354 bp = cm_data.buf_fileHashTablepp[i];
1357 lock_ReleaseWrite(&buf_globalLock);
1358 for (; bp; bp = nbp) {
1359 didRelease = 0; /* haven't released this buffer yet */
1361 /* clean buffer synchronously */
1362 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1363 lock_ObtainMutex(&bp->mx);
1365 /* start cleaning the buffer, and wait for it to finish */
1366 buf_LockedCleanAsync(bp, reqp);
1367 buf_WaitIO(scp, bp);
1368 lock_ReleaseMutex(&bp->mx);
1370 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1374 lock_ObtainWrite(&buf_globalLock);
1375 /* actually, we only know that buffer is clean if ref
1376 * count is 1, since we don't have buffer itself locked.
1378 if (!(bp->flags & CM_BUF_DIRTY)) {
1379 if (bp->refCount == 1) { /* bp is held above */
1380 buf_LockedRelease(bp);
1381 nbp = bp->fileHashp;
1388 lock_ReleaseWrite(&buf_globalLock);
1390 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1395 lock_ObtainWrite(&buf_globalLock);
1396 if (nbp = bp->fileHashp)
1398 buf_LockedRelease(bp);
1399 lock_ReleaseWrite(&buf_globalLock);
1401 } /* for loop over a bunch of buffers */
1404 buf_ValidateBufQueues();
1405 #endif /* TESTING */
1411 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1414 cm_buf_t *bp; /* buffer we're hacking on */
1415 cm_buf_t *nbp; /* next one */
1418 i = BUF_FILEHASH(&scp->fid);
1421 lock_ObtainWrite(&buf_globalLock);
1422 bp = cm_data.buf_fileHashTablepp[i];
1425 lock_ReleaseWrite(&buf_globalLock);
1426 for (; bp; bp = nbp) {
1427 /* clean buffer synchronously */
1428 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1431 lock_ObtainMutex(&bp->mx);
1433 cm_ReleaseUser(bp->userp);
1435 lock_ReleaseMutex(&bp->mx);
1437 buf_CleanAsync(bp, reqp);
1438 buf_CleanWait(scp, bp);
1439 lock_ObtainMutex(&bp->mx);
1440 if (bp->flags & CM_BUF_ERROR) {
1441 if (code == 0 || code == -1)
1446 lock_ReleaseMutex(&bp->mx);
1449 lock_ObtainWrite(&buf_globalLock);
1450 buf_LockedRelease(bp);
1451 nbp = bp->fileHashp;
1454 lock_ReleaseWrite(&buf_globalLock);
1455 } /* for loop over a bunch of buffers */
1458 buf_ValidateBufQueues();
1459 #endif /* TESTING */
1467 buf_ValidateBufQueues(void)
1469 cm_buf_t * bp, *bpb, *bpf, *bpa;
1470 afs_uint32 countf=0, countb=0, counta=0;
1472 lock_ObtainRead(&buf_globalLock);
1473 for (bp = cm_data.buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
1474 if (bp->magic != CM_BUF_MAGIC)
1480 for (bp = cm_data.buf_freeListp; bp; bp=(cm_buf_t *) osi_QNext(&bp->q)) {
1481 if (bp->magic != CM_BUF_MAGIC)
1487 for (bp = cm_data.buf_allp; bp; bp=bp->allp) {
1488 if (bp->magic != CM_BUF_MAGIC)
1493 lock_ReleaseRead(&buf_globalLock);
1495 if (countb != countf)
1498 if (counta != cm_data.buf_nbuffers)
1501 #endif /* TESTING */
1503 /* dump the contents of the buf_hashTablepp. */
1504 int cm_DumpBufHashTable(FILE *outputFile, char *cookie, int lock)
1511 if (cm_data.buf_hashTablepp == NULL)
1515 lock_ObtainRead(&buf_globalLock);
1517 StringCbPrintfA(output, sizeof(output), "%s - dumping buf_HashTable - buf_hashSize=%d\n",
1518 cookie, cm_data.buf_hashSize);
1519 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1521 for (i = 0; i < cm_data.buf_hashSize; i++)
1523 for (bp = cm_data.buf_hashTablepp[i]; bp; bp=bp->hashp)
1527 StringCbPrintfA(output, sizeof(output), "vnode=%d, unique=%d), size=%d refCount=%d\n",
1528 cookie, (void *)bp, i, bp->fid.cell, bp->fid.volume,
1529 bp->fid.vnode, bp->fid.unique, bp->size, bp->refCount);
1530 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1535 StringCbPrintfA(output, sizeof(output), "%s - Done dumping buf_HashTable.\n", cookie);
1536 WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
1539 lock_ReleaseRead(&buf_globalLock);
1543 void buf_ForceTrace(BOOL flush)
1552 len = GetTempPath(sizeof(buf)-10, buf);
1553 StringCbCopyA(&buf[len], sizeof(buf)-len, "/afs-buffer.log");
1554 handle = CreateFile(buf, GENERIC_WRITE, FILE_SHARE_READ,
1555 NULL, CREATE_ALWAYS, FILE_ATTRIBUTE_NORMAL, NULL);
1556 if (handle == INVALID_HANDLE_VALUE) {
1557 osi_panic("Cannot create log file", __FILE__, __LINE__);
1559 osi_LogPrint(buf_logp, handle);
1561 FlushFileBuffers(handle);
1562 CloseHandle(handle);