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>
27 /* This module implements the buffer package used by the local transaction
28 * system (cm). It is initialized by calling cm_Init, which calls buf_Init;
29 * it must be initalized before any of its main routines are called.
31 * Each buffer is hashed into a hash table by file ID and offset, and if its
32 * reference count is zero, it is also in a free list.
34 * There are two locks involved in buffer processing. The global lock
35 * buf_globalLock protects all of the global variables defined in this module,
36 * the reference counts and hash pointers in the actual cm_buf_t structures,
37 * and the LRU queue pointers in the buffer structures.
39 * The mutexes in the buffer structures protect the remaining fields in the
40 * buffers, as well the data itself.
42 * The locking hierarchy here is this:
44 * - resv multiple simul. buffers reservation
45 * - lock buffer I/O flags
46 * - lock buffer's mutex
47 * - lock buf_globalLock
51 /* global debugging log */
52 osi_log_t *buf_logp = NULL;
54 /* Global lock protecting hash tables and free lists */
55 osi_rwlock_t buf_globalLock;
57 /* ptr to head of the free list (most recently used) and the
58 * tail (the guy to remove first). We use osi_Q* functions
59 * to put stuff in buf_freeListp, and maintain the end
62 cm_buf_t *buf_freeListp;
63 cm_buf_t *buf_freeListEndp;
65 /* a pointer to a list of all buffers, just so that we can find them
66 * easily for debugging, and for the incr syncer. Locked under
71 /* defaults setup; these variables may be manually assigned into
72 * before calling cm_Init, as a way of changing these defaults.
74 long buf_nbuffers = CM_BUF_BUFFERS;
75 long buf_nOrigBuffers;
76 long buf_bufferSize = CM_BUF_SIZE;
77 long buf_hashSize = CM_BUF_HASHSIZE;
87 /* buffer reservation variables */
88 long buf_reservedBufs;
89 long buf_maxReservedBufs;
90 int buf_reserveWaiting;
92 /* callouts for reading and writing data, etc */
93 cm_buf_ops_t *cm_buf_opsp;
95 /* pointer to hash table; size computed dynamically */
96 cm_buf_t **buf_hashTablepp;
98 /* another hash table */
99 cm_buf_t **buf_fileHashTablepp;
102 /* for experimental disk caching support in Win95 client */
103 cm_buf_t *buf_diskFreeListp;
104 cm_buf_t *buf_diskFreeListEndp;
105 cm_buf_t *buf_diskAllp;
106 extern int cm_diskCacheEnabled;
107 #endif /* DISKCACHE95 */
109 /* hold a reference to an already held buffer */
110 void buf_Hold(cm_buf_t *bp)
112 lock_ObtainWrite(&buf_globalLock);
114 lock_ReleaseWrite(&buf_globalLock);
117 /* incremental sync daemon. Writes 1/10th of all the buffers every 5000 ms */
118 void buf_IncrSyncer(long parm)
120 cm_buf_t *bp; /* buffer we're hacking on; held */
121 long i; /* counter */
122 long nAtOnce; /* how many to do at once */
125 lock_ObtainWrite(&buf_globalLock);
128 lock_ReleaseWrite(&buf_globalLock);
129 nAtOnce = buf_nbuffers / 10;
132 i = SleepEx(5000, 1);
133 if (i != 0) continue;
138 /* now go through our percentage of the buffers */
139 for(i=0; i<nAtOnce; i++) {
140 /* don't want its identity changing while we're
141 * messing with it, so must do all of this with
145 /* start cleaning the buffer; don't touch log pages since
146 * the log code counts on knowing exactly who is writing
147 * a log page at any given instant.
150 req.flags |= CM_REQ_NORETRY;
151 buf_CleanAsync(bp, &req);
153 /* now advance to the next buffer; the allp chain never changes,
154 * and so can be followed even when holding no locks.
156 lock_ObtainWrite(&buf_globalLock);
157 buf_LockedRelease(bp);
159 if (!bp) bp = buf_allp;
161 lock_ReleaseWrite(&buf_globalLock);
162 } /* for loop over a bunch of buffers */
163 } /* whole daemon's while loop */
167 /* Create a security attribute structure suitable for use when the cache file
168 * is created. What we mainly want is that only the administrator should be
169 * able to do anything with the file. We create an ACL with only one entry,
170 * an entry that grants all rights to the administrator.
172 PSECURITY_ATTRIBUTES CreateCacheFileSA()
174 PSECURITY_ATTRIBUTES psa;
175 PSECURITY_DESCRIPTOR psd;
176 SID_IDENTIFIER_AUTHORITY authority = SECURITY_NT_AUTHORITY;
178 DWORD AdminSIDlength;
182 /* Get Administrator SID */
183 AllocateAndInitializeSid(&authority, 2,
184 SECURITY_BUILTIN_DOMAIN_RID,
185 DOMAIN_ALIAS_RID_ADMINS,
189 /* Create Administrator-only ACL */
190 AdminSIDlength = GetLengthSid(AdminSID);
191 ACLlength = sizeof(ACL) + sizeof(ACCESS_ALLOWED_ACE)
192 + AdminSIDlength - sizeof(DWORD);
193 AdminOnlyACL = GlobalAlloc(GMEM_FIXED, ACLlength);
194 InitializeAcl(AdminOnlyACL, ACLlength, ACL_REVISION);
195 AddAccessAllowedAce(AdminOnlyACL, ACL_REVISION,
196 STANDARD_RIGHTS_ALL | SPECIFIC_RIGHTS_ALL,
199 /* Create security descriptor */
200 psd = GlobalAlloc(GMEM_FIXED, sizeof(SECURITY_DESCRIPTOR));
201 InitializeSecurityDescriptor(psd, SECURITY_DESCRIPTOR_REVISION);
202 SetSecurityDescriptorDacl(psd, TRUE, AdminOnlyACL, FALSE);
204 /* Create security attributes structure */
205 psa = GlobalAlloc(GMEM_FIXED, sizeof(SECURITY_ATTRIBUTES));
206 psa->nLength = sizeof(SECURITY_ATTRIBUTES);
207 psa->lpSecurityDescriptor = psd;
208 psa->bInheritHandle = TRUE;
215 /* Free a security attribute structure created by CreateCacheFileSA() */
216 VOID FreeCacheFileSA(PSECURITY_ATTRIBUTES psa)
221 GetSecurityDescriptorDacl(psa->lpSecurityDescriptor, &b1, &pAcl, &b2);
223 GlobalFree(psa->lpSecurityDescriptor);
228 /* initialize the buffer package; called with no locks
229 * held during the initialization phase.
231 long buf_Init(cm_buf_ops_t *opsp)
233 static osi_once_t once;
239 PSECURITY_ATTRIBUTES psa;
247 /* Get system info; all we really want is the allocation granularity */
248 GetSystemInfo(&sysInfo);
251 /* Have to be able to reserve a whole chunk */
252 if (((buf_nbuffers - 3) * buf_bufferSize) < cm_chunkSize)
253 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");
264 * Cache file mapping constrained by
265 * system allocation granularity;
266 * round up, assuming granularity is a power of two
268 cs = buf_nbuffers * buf_bufferSize;
269 cs = (cs + (sysInfo.dwAllocationGranularity - 1))
270 & ~(sysInfo.dwAllocationGranularity - 1);
271 if (cs != buf_nbuffers * buf_bufferSize) {
272 buf_nbuffers = cs / buf_bufferSize;
273 afsi_log("Cache size rounded up to %d buffers",
278 /* remember this for those who want to reset it */
279 buf_nOrigBuffers = buf_nbuffers;
281 /* lower hash size to a prime number */
282 buf_hashSize = osi_PrimeLessThan(buf_hashSize);
284 /* create hash table */
285 buf_hashTablepp = malloc(buf_hashSize * sizeof(cm_buf_t *));
286 memset((void *)buf_hashTablepp, 0,
287 buf_hashSize * sizeof(cm_buf_t *));
289 /* another hash table */
290 buf_fileHashTablepp = malloc(buf_hashSize * sizeof(cm_buf_t *));
291 memset((void *)buf_fileHashTablepp, 0,
292 buf_hashSize * sizeof(cm_buf_t *));
294 /* min value for which this works */
298 /* Reserve buffer space by mapping cache file */
299 psa = CreateCacheFileSA();
300 hf = CreateFile(cm_CachePath,
301 GENERIC_READ | GENERIC_WRITE,
302 FILE_SHARE_READ | FILE_SHARE_WRITE,
305 FILE_ATTRIBUTE_NORMAL,
307 if (hf == INVALID_HANDLE_VALUE) {
308 afsi_log("create file error %d", GetLastError());
309 return CM_ERROR_INVAL;
311 FreeCacheFileSA(psa);
313 hm = CreateFileMapping(hf,
316 0, buf_nbuffers * buf_bufferSize,
319 if (GetLastError() == ERROR_DISK_FULL) {
320 afsi_log("Error creating cache file mapping: disk full");
321 return CM_ERROR_TOOMANYBUFS;
323 return CM_ERROR_INVAL;
325 data = MapViewOfFile(hm,
328 buf_nbuffers * buf_bufferSize);
332 return CM_ERROR_INVAL;
336 /* djgpp doesn't support memory mapped files */
337 data = malloc(buf_nbuffers * buf_bufferSize);
340 /* create buffer headers and put in free list */
341 bp = malloc(buf_nbuffers * sizeof(cm_buf_t));
343 for(i=0; i<buf_nbuffers; i++) {
344 /* allocate and zero some storage */
345 memset(bp, 0, sizeof(cm_buf_t));
347 /* thread on list of all buffers */
351 osi_QAdd((osi_queue_t **)&buf_freeListp, &bp->q);
352 bp->flags |= CM_BUF_INLRU;
353 lock_InitializeMutex(&bp->mx, "Buffer mutex");
355 /* grab appropriate number of bytes from aligned zone */
358 /* setup last buffer pointer */
360 buf_freeListEndp = bp;
364 data += buf_bufferSize;
367 /* none reserved at first */
368 buf_reservedBufs = 0;
370 /* just for safety's sake */
371 buf_maxReservedBufs = buf_nbuffers - 3;
373 /* init the buffer trace log */
374 buf_logp = osi_LogCreate("buffer", 10);
378 /* and create the incr-syncer */
379 phandle = thrd_Create(0, 0,
380 (ThreadFunc) buf_IncrSyncer, 0, 0, &pid,
383 osi_assertx(phandle != NULL, "buf: can't create incremental sync proc");
385 CloseHandle(phandle);
392 /* add nbuffers to the buffer pool, if possible.
393 * Called with no locks held.
395 long buf_AddBuffers(long nbuffers)
405 * Cache file mapping constrained by
406 * system allocation granularity;
407 * round up, assuming granularity is a power of two;
408 * assume existing cache size is already rounded
410 cs = nbuffers * buf_bufferSize;
411 cs = (cs + (sysInfo.dwAllocationGranularity - 1))
412 & ~(sysInfo.dwAllocationGranularity - 1);
413 if (cs != nbuffers * buf_bufferSize) {
414 nbuffers = cs / buf_bufferSize;
417 /* Reserve additional buffer space by remapping cache file */
418 hm = CreateFileMapping(CacheHandle,
421 0, (buf_nbuffers + nbuffers) * buf_bufferSize,
424 if (GetLastError() == ERROR_DISK_FULL)
425 return CM_ERROR_TOOMANYBUFS;
427 return CM_ERROR_INVAL;
429 data = MapViewOfFile(hm,
431 0, buf_nbuffers * buf_bufferSize,
432 nbuffers * buf_bufferSize);
435 return CM_ERROR_INVAL;
439 data = malloc(buf_nbuffers * buf_bufferSize);
442 /* Create buffer headers and put in free list */
443 bp = malloc(nbuffers * sizeof(*bp));
445 for(i=0; i<nbuffers; i++) {
446 memset(bp, 0, sizeof(*bp));
448 lock_InitializeMutex(&bp->mx, "cm_buf_t");
450 /* grab appropriate number of bytes from aligned zone */
453 bp->flags |= CM_BUF_INLRU;
455 lock_ObtainWrite(&buf_globalLock);
456 /* note that buf_allp chain is covered by buf_globalLock now */
459 osi_QAdd((osi_queue_t **) &buf_freeListp, &bp->q);
460 if (!buf_freeListEndp) buf_freeListEndp = bp;
462 lock_ReleaseWrite(&buf_globalLock);
465 data += buf_bufferSize;
467 } /* for loop over all buffers */
472 /* interface to set the number of buffers to an exact figure.
473 * Called with no locks held.
475 long buf_SetNBuffers(long nbuffers)
477 if (nbuffers < 10) return CM_ERROR_INVAL;
478 if (nbuffers == buf_nbuffers) return 0;
479 else if (nbuffers > buf_nbuffers)
480 return buf_AddBuffers(nbuffers - buf_nbuffers);
481 else return CM_ERROR_INVAL;
484 /* release a buffer. Buffer must be referenced, but unlocked. */
485 void buf_Release(cm_buf_t *bp)
487 lock_ObtainWrite(&buf_globalLock);
488 buf_LockedRelease(bp);
489 lock_ReleaseWrite(&buf_globalLock);
492 /* wait for reading or writing to clear; called with write-locked
493 * buffer, and returns with locked buffer.
495 void buf_WaitIO(cm_buf_t *bp)
498 /* if no IO is happening, we're done */
499 if (!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)))
502 /* otherwise I/O is happening, but some other thread is waiting for
503 * the I/O already. Wait for that guy to figure out what happened,
504 * and then check again.
506 bp->flags |= CM_BUF_WAITING;
507 osi_SleepM((long) bp, &bp->mx);
508 lock_ObtainMutex(&bp->mx);
509 osi_Log1(buf_logp, "buf_WaitIO conflict wait done for 0x%x", bp);
512 /* if we get here, the IO is done, but we may have to wakeup people waiting for
513 * the I/O to complete. Do so.
515 if (bp->flags & CM_BUF_WAITING) {
516 bp->flags &= ~CM_BUF_WAITING;
517 osi_Wakeup((long) bp);
519 osi_Log1(buf_logp, "WaitIO finished wait for bp 0x%x", (long) bp);
522 /* code to drop reference count while holding buf_globalLock */
523 void buf_LockedRelease(cm_buf_t *bp)
525 /* ensure that we're in the LRU queue if our ref count is 0 */
526 osi_assert(bp->refCount > 0);
527 if (--bp->refCount == 0) {
528 if (!(bp->flags & CM_BUF_INLRU)) {
529 osi_QAdd((osi_queue_t **) &buf_freeListp, &bp->q);
531 /* watch for transition from empty to one element */
532 if (!buf_freeListEndp)
533 buf_freeListEndp = buf_freeListp;
534 bp->flags |= CM_BUF_INLRU;
539 /* find a buffer, if any, for a particular file ID and offset. Assumes
540 * that buf_globalLock is write locked when called.
542 cm_buf_t *buf_LockedFind(struct cm_scache *scp, osi_hyper_t *offsetp)
547 i = BUF_HASH(&scp->fid, offsetp);
548 for(bp = buf_hashTablepp[i]; bp; bp=bp->hashp) {
549 if (cm_FidCmp(&scp->fid, &bp->fid) == 0
550 && offsetp->LowPart == bp->offset.LowPart
551 && offsetp->HighPart == bp->offset.HighPart) {
557 /* return whatever we found, if anything */
561 /* find a buffer with offset *offsetp for vnode *scp. Called
562 * with no locks held.
564 cm_buf_t *buf_Find(struct cm_scache *scp, osi_hyper_t *offsetp)
568 lock_ObtainWrite(&buf_globalLock);
569 bp = buf_LockedFind(scp, offsetp);
570 lock_ReleaseWrite(&buf_globalLock);
575 /* start cleaning I/O on this buffer. Buffer must be write locked, and is returned
578 * Makes sure that there's only one person writing this block
579 * at any given time, and also ensures that the log is forced sufficiently far,
580 * if this buffer contains logged data.
582 void buf_LockedCleanAsync(cm_buf_t *bp, cm_req_t *reqp)
587 while ((bp->flags & (CM_BUF_WRITING | CM_BUF_DIRTY)) == CM_BUF_DIRTY) {
588 lock_ReleaseMutex(&bp->mx);
590 code = (*cm_buf_opsp->Writep)(&bp->fid, &bp->offset,
591 buf_bufferSize, 0, bp->userp,
594 lock_ObtainMutex(&bp->mx);
598 /* Disk cache support */
599 /* write buffer to disk cache (synchronous for now) */
600 diskcache_Update(bp->dcp, bp->datap, buf_bufferSize, bp->dataVersion);
601 #endif /* DISKCACHE95 */
604 /* do logging after call to GetLastError, or else */
605 osi_Log2(buf_logp, "buf_CleanAsync starts I/O on 0x%x, done=%d", bp, code);
607 /* if someone was waiting for the I/O that just completed or failed,
610 if (bp->flags & CM_BUF_WAITING) {
611 /* turn off flags and wakeup users */
612 bp->flags &= ~CM_BUF_WAITING;
613 osi_Wakeup((long) bp);
617 /* Called with a zero-ref count buffer and with the buf_globalLock write locked.
618 * recycles the buffer, and leaves it ready for reuse with a ref count of 1.
619 * The buffer must already be clean, and no I/O should be happening to it.
621 void buf_Recycle(cm_buf_t *bp)
626 cm_buf_t *prevBp, *nextBp;
628 /* if we get here, we know that the buffer still has a 0 ref count,
629 * and that it is clean and has no currently pending I/O. This is
630 * the dude to return.
631 * Remember that as long as the ref count is 0, we know that we won't
632 * have any lock conflicts, so we can grab the buffer lock out of
633 * order in the locking hierarchy.
636 "buf_Recycle recycles 0x%x, off 0x%x",
637 bp, bp->offset.LowPart);
639 osi_assert(bp->refCount == 0);
640 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING | CM_BUF_DIRTY)));
641 lock_AssertWrite(&buf_globalLock);
643 if (bp->flags & CM_BUF_INHASH) {
644 /* Remove from hash */
646 i = BUF_HASH(&bp->fid, &bp->offset);
647 lbpp = &(buf_hashTablepp[i]);
648 for(tbp = *lbpp; tbp; lbpp = &tbp->hashp, tbp = *lbpp) {
649 if (tbp == bp) break;
652 /* we better find it */
653 osi_assertx(tbp != NULL, "buf_GetNewLocked: hash table screwup");
655 *lbpp = bp->hashp; /* hash out */
657 /* Remove from file hash */
659 i = BUF_FILEHASH(&bp->fid);
660 prevBp = bp->fileHashBackp;
661 nextBp = bp->fileHashp;
663 prevBp->fileHashp = nextBp;
665 buf_fileHashTablepp[i] = nextBp;
667 nextBp->fileHashBackp = prevBp;
669 bp->flags &= ~CM_BUF_INHASH;
672 /* bump the soft reference counter now, to invalidate softRefs; no
673 * wakeup is required since people don't sleep waiting for this
678 /* make the fid unrecognizable */
679 memset(&bp->fid, 0, sizeof(bp->fid));
682 /* recycle a buffer, removing it from the free list, hashing in its new identity
683 * and returning it write-locked so that no one can use it. Called without
684 * any locks held, and can return an error if it loses the race condition and
685 * finds that someone else created the desired buffer.
687 * If success is returned, the buffer is returned write-locked.
689 * May be called with null scp and offsetp, if we're just trying to reclaim some
690 * space from the buffer pool. In that case, the buffer will be returned
691 * without being hashed into the hash table.
693 long buf_GetNewLocked(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
695 cm_buf_t *bp; /* buffer we're dealing with */
696 cm_buf_t *nextBp; /* next buffer in file hash chain */
700 cm_InitReq(&req); /* just in case */
704 lock_ObtainWrite(&buf_globalLock);
705 /* check to see if we lost the race */
707 if (bp = buf_LockedFind(scp, offsetp)) {
709 lock_ReleaseWrite(&buf_globalLock);
710 return CM_BUF_EXISTS;
714 /* for debugging, assert free list isn't empty, although we
715 * really should try waiting for a running tranasction to finish
716 * instead of this; or better, we should have a transaction
717 * throttler prevent us from entering this situation.
719 osi_assertx(buf_freeListEndp != NULL, "buf_GetNewLocked: no free buffers");
721 /* look at all buffers in free list, some of which may temp.
722 * have high refcounts and which then should be skipped,
723 * starting cleaning I/O for those which are dirty. If we find
724 * a clean buffer, we rehash it, lock it and return it.
726 for(bp = buf_freeListEndp; bp; bp=(cm_buf_t *) osi_QPrev(&bp->q)) {
727 /* check to see if it really has zero ref count. This
728 * code can bump refcounts, at least, so it may not be
731 if (bp->refCount > 0) continue;
733 /* we don't have to lock buffer itself, since the ref
734 * count is 0 and we know it will stay zero as long as
735 * we hold the global lock.
738 /* don't recycle someone in our own chunk */
739 if (!cm_FidCmp(&bp->fid, &scp->fid)
740 && (bp->offset.LowPart & (-cm_chunkSize))
741 == (offsetp->LowPart & (-cm_chunkSize)))
744 /* if this page is being filled (!) or cleaned, see if
745 * the I/O has completed. If not, skip it, otherwise
746 * do the final processing for the I/O.
748 if (bp->flags & (CM_BUF_READING | CM_BUF_WRITING)) {
749 /* probably shouldn't do this much work while
750 * holding the big lock? Watch for contention
756 if (bp->flags & CM_BUF_DIRTY) {
757 /* if the buffer is dirty, start cleaning it and
758 * move on to the next buffer. We do this with
759 * just the lock required to minimize contention
763 lock_ReleaseWrite(&buf_globalLock);
765 /* grab required lock and clean; this only
766 * starts the I/O. By the time we're back,
767 * it'll still be marked dirty, but it will also
768 * have the WRITING flag set, so we won't get
771 buf_CleanAsync(bp, &req);
773 /* now put it back and go around again */
778 /* if we get here, we know that the buffer still has a 0
779 * ref count, and that it is clean and has no currently
780 * pending I/O. This is the dude to return.
781 * Remember that as long as the ref count is 0, we know
782 * that we won't have any lock conflicts, so we can grab
783 * the buffer lock out of order in the locking hierarchy.
787 /* clean up junk flags */
788 bp->flags &= ~(CM_BUF_EOF | CM_BUF_ERROR);
789 bp->dataVersion = -1; /* unknown so far */
791 /* now hash in as our new buffer, and give it the
792 * appropriate label, if requested.
795 bp->flags |= CM_BUF_INHASH;
797 bp->offset = *offsetp;
798 i = BUF_HASH(&scp->fid, offsetp);
799 bp->hashp = buf_hashTablepp[i];
800 buf_hashTablepp[i] = bp;
801 i = BUF_FILEHASH(&scp->fid);
802 nextBp = buf_fileHashTablepp[i];
803 bp->fileHashp = nextBp;
804 bp->fileHashBackp = NULL;
806 nextBp->fileHashBackp = bp;
807 buf_fileHashTablepp[i] = bp;
810 /* prepare to return it. Start by giving it a good
814 /* and since it has a non-zero ref count, we should move
815 * it from the lru queue. It better be still there,
816 * since we've held the global (big) lock since we found
819 osi_assertx(bp->flags & CM_BUF_INLRU,
820 "buf_GetNewLocked: LRU screwup");
821 if (buf_freeListEndp == bp) {
822 /* we're the last guy in this queue, so maintain it */
823 buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
825 osi_QRemove((osi_queue_t **) &buf_freeListp, &bp->q);
826 bp->flags &= ~CM_BUF_INLRU;
828 /* finally, grab the mutex so that people don't use it
829 * before the caller fills it with data. Again, no one
830 * should have been able to get to this dude to lock it.
832 osi_assertx(lock_TryMutex(&bp->mx),
833 "buf_GetNewLocked: TryMutex failed");
835 lock_ReleaseWrite(&buf_globalLock);
838 } /* for all buffers in lru queue */
839 lock_ReleaseWrite(&buf_globalLock);
840 } /* while loop over everything */
844 /* get a page, returning it held but unlocked. Doesn't fill in the page
845 * with I/O, since we're going to write the whole thing new.
847 long buf_GetNew(struct cm_scache *scp, osi_hyper_t *offsetp, cm_buf_t **bufpp)
851 osi_hyper_t pageOffset;
855 pageOffset.HighPart = offsetp->HighPart;
856 pageOffset.LowPart = offsetp->LowPart & ~(buf_bufferSize-1);
858 lock_ObtainWrite(&buf_globalLock);
859 bp = buf_LockedFind(scp, &pageOffset);
860 lock_ReleaseWrite(&buf_globalLock);
862 /* lock it and break out */
863 lock_ObtainMutex(&bp->mx);
867 /* otherwise, we have to create a page */
868 code = buf_GetNewLocked(scp, &pageOffset, &bp);
870 /* check if the buffer was created in a race condition branch.
871 * If so, go around so we can hold a reference to it.
873 if (code == CM_BUF_EXISTS) continue;
875 /* something else went wrong */
876 if (code != 0) return code;
878 /* otherwise, we have a locked buffer that we just created */
881 } /* big while loop */
884 if (bp->flags & CM_BUF_READING)
887 /* once it has been read once, we can unlock it and return it, still
888 * with its refcount held.
890 lock_ReleaseMutex(&bp->mx);
892 osi_Log3(buf_logp, "buf_GetNew returning bp 0x%x for file 0x%x, offset 0x%x",
893 bp, (long) scp, offsetp->LowPart);
897 /* get a page, returning it held but unlocked. Make sure it is complete */
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;
907 #endif /* DISKCACHE95 */
910 pageOffset.HighPart = offsetp->HighPart;
911 pageOffset.LowPart = offsetp->LowPart & ~(buf_bufferSize-1);
913 lock_ObtainWrite(&buf_globalLock);
914 bp = buf_LockedFind(scp, &pageOffset);
915 lock_ReleaseWrite(&buf_globalLock);
917 /* lock it and break out */
918 lock_ObtainMutex(&bp->mx);
922 /* touch disk chunk to update LRU info */
923 diskcache_Touch(bp->dcp);
924 #endif /* DISKCACHE95 */
927 /* otherwise, we have to create a page */
928 code = buf_GetNewLocked(scp, &pageOffset, &bp);
930 /* check if the buffer was created in a race condition branch.
931 * If so, go around so we can hold a reference to it.
933 if (code == CM_BUF_EXISTS) continue;
935 /* something else went wrong */
936 if (code != 0) return code;
938 /* otherwise, we have a locked buffer that we just created */
941 } /* big while loop */
943 /* if we get here, we have a locked buffer that may have just been
944 * created, in which case it needs to be filled with data.
947 /* load the page; freshly created pages should be idle */
948 osi_assert(!(bp->flags & (CM_BUF_READING | CM_BUF_WRITING)));
950 /* setup offset, event */
951 #ifndef DJGPP /* doesn't seem to be used */
952 bp->over.Offset = bp->offset.LowPart;
953 bp->over.OffsetHigh = bp->offset.HighPart;
956 /* start the I/O; may drop lock */
957 bp->flags |= CM_BUF_READING;
958 code = (*cm_buf_opsp->Readp)(bp, buf_bufferSize, &tcount, NULL);
961 code = diskcache_Get(&bp->fid, &bp->offset, bp->datap, buf_bufferSize, &bp->dataVersion, &tcount, &dcp);
962 bp->dcp = dcp; /* pointer to disk cache struct. */
963 #endif /* DISKCACHE95 */
966 /* failure or queued */
967 #ifndef DJGPP /* cm_bufRead always returns 0 */
968 if (code != ERROR_IO_PENDING) {
971 bp->flags |= CM_BUF_ERROR;
972 bp->flags &= ~CM_BUF_READING;
973 if (bp->flags & CM_BUF_WAITING) {
974 bp->flags &= ~CM_BUF_WAITING;
975 osi_Wakeup((long) bp);
977 lock_ReleaseMutex(&bp->mx);
984 /* otherwise, I/O completed instantly and we're done, except
985 * for padding the xfr out with 0s and checking for EOF
987 if (tcount < (unsigned long) buf_bufferSize) {
988 memset(bp->datap+tcount, 0, buf_bufferSize - tcount);
990 bp->flags |= CM_BUF_EOF;
992 bp->flags &= ~CM_BUF_READING;
993 if (bp->flags & CM_BUF_WAITING) {
994 bp->flags &= ~CM_BUF_WAITING;
995 osi_Wakeup((long) bp);
1001 /* wait for reads, either that which we started above, or that someone
1002 * else started. We don't care if we return a buffer being cleaned.
1004 if (bp->flags & CM_BUF_READING)
1007 /* once it has been read once, we can unlock it and return it, still
1008 * with its refcount held.
1010 lock_ReleaseMutex(&bp->mx);
1013 /* now remove from queue; will be put in at the head (farthest from
1014 * being recycled) when we're done in buf_Release.
1016 lock_ObtainWrite(&buf_globalLock);
1017 if (bp->flags & CM_BUF_INLRU) {
1018 if (buf_freeListEndp == bp)
1019 buf_freeListEndp = (cm_buf_t *) osi_QPrev(&bp->q);
1020 osi_QRemove((osi_queue_t **) &buf_freeListp, &bp->q);
1021 bp->flags &= ~CM_BUF_INLRU;
1023 lock_ReleaseWrite(&buf_globalLock);
1025 osi_Log3(buf_logp, "buf_Get returning bp 0x%x for file 0x%x, offset 0x%x",
1026 bp, (long) scp, offsetp->LowPart);
1030 /* count # of elements in the free list;
1031 * we don't bother doing the proper locking for accessing dataVersion or flags
1032 * since it is a pain, and this is really just an advisory call. If you need
1033 * to do better at some point, rewrite this function.
1035 long buf_CountFreeList(void)
1041 lock_ObtainRead(&buf_globalLock);
1042 for(bufp = buf_freeListp; bufp; bufp = (cm_buf_t *) osi_QNext(&bufp->q)) {
1043 /* if the buffer doesn't have an identity, or if the buffer
1044 * has been invalidate (by having its DV stomped upon), then
1045 * count it as free, since it isn't really being utilized.
1047 if (!(bufp->flags & CM_BUF_INHASH) || bufp->dataVersion <= 0)
1050 lock_ReleaseRead(&buf_globalLock);
1054 /* clean a buffer synchronously */
1055 void buf_CleanAsync(cm_buf_t *bp, cm_req_t *reqp)
1057 lock_ObtainMutex(&bp->mx);
1058 buf_LockedCleanAsync(bp, reqp);
1059 lock_ReleaseMutex(&bp->mx);
1062 /* wait for a buffer's cleaning to finish */
1063 void buf_CleanWait(cm_buf_t *bp)
1065 lock_ObtainMutex(&bp->mx);
1066 if (bp->flags & CM_BUF_WRITING) {
1069 lock_ReleaseMutex(&bp->mx);
1072 /* set the dirty flag on a buffer, and set associated write-ahead log,
1073 * if there is one. Allow one to be added to a buffer, but not changed.
1075 * The buffer must be locked before calling this routine.
1077 void buf_SetDirty(cm_buf_t *bp)
1079 osi_assert(bp->refCount > 0);
1081 osi_Log1(buf_logp, "buf_SetDirty 0x%x", bp);
1084 bp->flags |= CM_BUF_DIRTY;
1086 /* and turn off EOF flag, since it has associated data now */
1087 bp->flags &= ~CM_BUF_EOF;
1090 /* clean all buffers, reset log pointers and invalidate all buffers.
1091 * Called with no locks held, and returns with same.
1093 * This function is guaranteed to clean and remove the log ptr of all the
1094 * buffers that were dirty or had non-zero log ptrs before the call was
1095 * made. That's sufficient to clean up any garbage left around by recovery,
1096 * which is all we're counting on this for; there may be newly created buffers
1097 * added while we're running, but that should be OK.
1099 * In an environment where there are no transactions (artificially imposed, for
1100 * example, when switching the database to raw mode), this function is used to
1101 * make sure that all updates have been written to the disk. In that case, we don't
1102 * really require that we forget the log association between pages and logs, but
1103 * it also doesn't hurt. Since raw mode I/O goes through this buffer package, we don't
1104 * have to worry about invalidating data in the buffers.
1106 * This function is used at the end of recovery as paranoia to get the recovered
1107 * database out to disk. It removes all references to the recovery log and cleans
1110 long buf_CleanAndReset(void)
1116 lock_ObtainWrite(&buf_globalLock);
1117 for(i=0; i<buf_hashSize; i++) {
1118 for(bp = buf_hashTablepp[i]; bp; bp = bp->hashp) {
1120 lock_ReleaseWrite(&buf_globalLock);
1122 /* now no locks are held; clean buffer and go on */
1124 buf_CleanAsync(bp, &req);
1127 /* relock and release buffer */
1128 lock_ObtainWrite(&buf_globalLock);
1129 buf_LockedRelease(bp);
1130 } /* over one bucket */
1131 } /* for loop over all hash buckets */
1134 lock_ReleaseWrite(&buf_globalLock);
1136 /* and we're done */
1140 /* called without global lock being held, reserves buffers for callers
1141 * that need more than one held (not locked) at once.
1143 void buf_ReserveBuffers(long nbuffers)
1145 lock_ObtainWrite(&buf_globalLock);
1147 if (buf_reservedBufs + nbuffers > buf_maxReservedBufs) {
1148 buf_reserveWaiting = 1;
1149 osi_Log1(buf_logp, "buf_ReserveBuffers waiting for %d bufs", nbuffers);
1150 osi_SleepW((long) &buf_reservedBufs, &buf_globalLock);
1151 lock_ObtainWrite(&buf_globalLock);
1154 buf_reservedBufs += nbuffers;
1158 lock_ReleaseWrite(&buf_globalLock);
1161 int buf_TryReserveBuffers(long nbuffers)
1165 lock_ObtainWrite(&buf_globalLock);
1166 if (buf_reservedBufs + nbuffers > buf_maxReservedBufs) {
1170 buf_reservedBufs += nbuffers;
1173 lock_ReleaseWrite(&buf_globalLock);
1177 /* called without global lock held, releases reservation held by
1178 * buf_ReserveBuffers.
1180 void buf_UnreserveBuffers(long nbuffers)
1182 lock_ObtainWrite(&buf_globalLock);
1183 buf_reservedBufs -= nbuffers;
1184 if (buf_reserveWaiting) {
1185 buf_reserveWaiting = 0;
1186 osi_Wakeup((long) &buf_reservedBufs);
1188 lock_ReleaseWrite(&buf_globalLock);
1191 /* truncate the buffers past sizep, zeroing out the page, if we don't
1192 * end on a page boundary.
1194 * Requires cm_bufCreateLock to be write locked.
1196 long buf_Truncate(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp,
1200 cm_buf_t *nbufp; /* next buffer, if didRelease */
1207 /* assert that cm_bufCreateLock is held in write mode */
1208 lock_AssertWrite(&scp->bufCreateLock);
1210 i = BUF_FILEHASH(&scp->fid);
1212 lock_ObtainWrite(&buf_globalLock);
1213 bufp = buf_fileHashTablepp[i];
1215 lock_ReleaseWrite(&buf_globalLock);
1220 lock_ReleaseWrite(&buf_globalLock);
1221 for(; bufp; bufp = nbufp) {
1223 lock_ObtainMutex(&bufp->mx);
1225 bufEnd.HighPart = 0;
1226 bufEnd.LowPart = buf_bufferSize;
1227 bufEnd = LargeIntegerAdd(bufEnd, bufp->offset);
1229 if (cm_FidCmp(&bufp->fid, &scp->fid) == 0 &&
1230 LargeIntegerLessThan(*sizep, bufEnd)) {
1233 lock_ObtainMutex(&scp->mx);
1235 /* make sure we have a callback (so we have the right value for
1236 * the length), and wait for it to be safe to do a truncate.
1238 code = cm_SyncOp(scp, bufp, userp, reqp, 0,
1239 CM_SCACHESYNC_NEEDCALLBACK
1240 | CM_SCACHESYNC_GETSTATUS
1241 | CM_SCACHESYNC_SETSIZE
1242 | CM_SCACHESYNC_BUFLOCKED);
1243 /* if we succeeded in our locking, and this applies to the right
1244 * file, and the truncate request overlaps the buffer either
1245 * totally or partially, then do something.
1247 if (code == 0 && cm_FidCmp(&bufp->fid, &scp->fid) == 0
1248 && LargeIntegerLessThan(*sizep, bufEnd)) {
1250 lock_ObtainWrite(&buf_globalLock);
1252 /* destroy the buffer, turning off its dirty bit, if
1253 * we're truncating the whole buffer. Otherwise, set
1254 * the dirty bit, and clear out the tail of the buffer
1255 * if we just overlap some.
1257 if (LargeIntegerLessThanOrEqualTo(*sizep, bufp->offset)) {
1258 /* truncating the entire page */
1259 bufp->flags &= ~CM_BUF_DIRTY;
1260 bufp->dataVersion = -1; /* known bad */
1261 bufp->dirtyCounter++;
1264 /* don't set dirty, since dirty implies
1265 * currently up-to-date. Don't need to do this,
1266 * since we'll update the length anyway.
1268 * Zero out remainder of the page, in case we
1269 * seek and write past EOF, and make this data
1272 bufferPos = sizep->LowPart & (buf_bufferSize - 1);
1273 osi_assert(bufferPos != 0);
1274 memset(bufp->datap + bufferPos, 0,
1275 buf_bufferSize - bufferPos);
1278 lock_ReleaseWrite(&buf_globalLock);
1282 lock_ReleaseMutex(&scp->mx);
1283 lock_ReleaseMutex(&bufp->mx);
1285 lock_ObtainWrite(&buf_globalLock);
1286 nbufp = bufp->fileHashp;
1287 if (nbufp) nbufp->refCount++;
1288 buf_LockedRelease(bufp);
1289 lock_ReleaseWrite(&buf_globalLock);
1292 /* bail out early if we fail */
1294 /* at this point, nbufp is held; bufp has already been
1297 if (nbufp) buf_Release(nbufp);
1306 long buf_FlushCleanPages(cm_scache_t *scp, cm_user_t *userp, cm_req_t *reqp)
1309 cm_buf_t *bp; /* buffer we're hacking on */
1314 i = BUF_FILEHASH(&scp->fid);
1317 lock_ObtainWrite(&buf_globalLock);
1318 bp = buf_fileHashTablepp[i];
1319 if (bp) bp->refCount++;
1320 lock_ReleaseWrite(&buf_globalLock);
1321 for(; bp; bp = nbp) {
1322 didRelease = 0; /* haven't released this buffer yet */
1324 /* clean buffer synchronously */
1325 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1326 lock_ObtainMutex(&bp->mx);
1328 /* start cleaning the buffer, and wait for it to finish */
1329 buf_LockedCleanAsync(bp, reqp);
1331 lock_ReleaseMutex(&bp->mx);
1333 code = (*cm_buf_opsp->Stabilizep)(scp, userp, reqp);
1334 if (code) goto skip;
1336 lock_ObtainWrite(&buf_globalLock);
1337 /* actually, we only know that buffer is clean if ref
1338 * count is 1, since we don't have buffer itself locked.
1340 if (!(bp->flags & CM_BUF_DIRTY)) {
1341 if (bp->refCount == 1) { /* bp is held above */
1342 buf_LockedRelease(bp);
1343 nbp = bp->fileHashp;
1344 if (nbp) nbp->refCount++;
1349 lock_ReleaseWrite(&buf_globalLock);
1351 (*cm_buf_opsp->Unstabilizep)(scp, userp);
1356 lock_ObtainWrite(&buf_globalLock);
1357 if (nbp = bp->fileHashp) nbp->refCount++;
1358 buf_LockedRelease(bp);
1359 lock_ReleaseWrite(&buf_globalLock);
1361 } /* for loop over a bunch of buffers */
1367 long buf_CleanVnode(struct cm_scache *scp, cm_user_t *userp, cm_req_t *reqp)
1370 cm_buf_t *bp; /* buffer we're hacking on */
1371 cm_buf_t *nbp; /* next one */
1374 i = BUF_FILEHASH(&scp->fid);
1377 lock_ObtainWrite(&buf_globalLock);
1378 bp = buf_fileHashTablepp[i];
1379 if (bp) bp->refCount++;
1380 lock_ReleaseWrite(&buf_globalLock);
1381 for(; bp; bp = nbp) {
1382 /* clean buffer synchronously */
1383 if (cm_FidCmp(&bp->fid, &scp->fid) == 0) {
1385 lock_ObtainMutex(&bp->mx);
1386 if (bp->userp) cm_ReleaseUser(bp->userp);
1388 lock_ReleaseMutex(&bp->mx);
1391 buf_CleanAsync(bp, reqp);
1393 lock_ObtainMutex(&bp->mx);
1394 if (bp->flags & CM_BUF_ERROR) {
1395 if (code == 0 || code == -1) code = bp->error;
1396 if (code == 0) code = -1;
1398 lock_ReleaseMutex(&bp->mx);
1401 lock_ObtainWrite(&buf_globalLock);
1402 buf_LockedRelease(bp);
1403 nbp = bp->fileHashp;
1404 if (nbp) nbp->refCount++;
1405 lock_ReleaseWrite(&buf_globalLock);
1406 } /* for loop over a bunch of buffers */