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
13 #include <afsconfig.h>
14 #include "../afs/param.h"
18 #include "../afs/sysincludes.h" /*Standard vendor system headers*/
19 #include "../afs/afsincludes.h" /*AFS-based standard headers*/
20 #include "../afs/afs_stats.h" /* statistics */
21 #include "../afs/afs_cbqueue.h"
22 #include "../afs/afs_osidnlc.h"
24 /* Forward declarations. */
25 static void afs_GetDownD(int anumber, int *aneedSpace);
26 static void afs_FreeDiscardedDCache(void);
27 static void afs_DiscardDCache(struct dcache *);
28 static void afs_FreeDCache(struct dcache *);
31 * --------------------- Exported definitions ---------------------
33 afs_lock_t afs_xdcache; /*Lock: alloc new disk cache entries*/
34 afs_int32 afs_freeDCList; /*Free list for disk cache entries*/
35 afs_int32 afs_freeDCCount; /*Count of elts in freeDCList*/
36 afs_int32 afs_discardDCList; /*Discarded disk cache entries*/
37 afs_int32 afs_discardDCCount; /*Count of elts in discardDCList*/
38 struct dcache *afs_freeDSList; /*Free list for disk slots */
39 struct dcache *afs_Initial_freeDSList; /*Initial list for above*/
40 ino_t cacheInode; /*Inode for CacheItems file*/
41 struct osi_file *afs_cacheInodep = 0; /* file for CacheItems inode */
42 struct afs_q afs_DLRU; /*dcache LRU*/
43 afs_int32 afs_dhashsize = 1024;
44 afs_int32 *afs_dvhashTbl; /*Data cache hash table*/
45 afs_int32 *afs_dchashTbl; /*Data cache hash table*/
46 afs_int32 *afs_dvnextTbl; /*Dcache hash table links */
47 afs_int32 *afs_dcnextTbl; /*Dcache hash table links */
48 struct dcache **afs_indexTable; /*Pointers to dcache entries*/
49 afs_hyper_t *afs_indexTimes; /*Dcache entry Access times*/
50 afs_int32 *afs_indexUnique; /*dcache entry Fid.Unique */
51 unsigned char *afs_indexFlags; /*(only one) Is there data there?*/
52 afs_hyper_t afs_indexCounter; /*Fake time for marking index
54 afs_int32 afs_cacheFiles =0; /*Size of afs_indexTable*/
55 afs_int32 afs_cacheBlocks; /*1K blocks in cache*/
56 afs_int32 afs_cacheStats; /*Stat entries in cache*/
57 afs_int32 afs_blocksUsed; /*Number of blocks in use*/
58 afs_int32 afs_blocksDiscarded; /*Blocks freed but not truncated */
59 afs_int32 afs_fsfragsize = 1023; /*Underlying Filesystem minimum unit
60 *of disk allocation usually 1K
61 *this value is (truefrag -1 ) to
62 *save a bunch of subtracts... */
63 #ifdef AFS_64BIT_CLIENT
64 #ifdef AFS_VM_RDWR_ENV
65 afs_size_t afs_vmMappingEnd; /* for large files (>= 2GB) the VM
66 * mapping an 32bit addressing machines
67 * can only be used below the 2 GB
68 * line. From this point upwards we
69 * must do direct I/O into the cache
70 * files. The value should be on a
72 #endif /* AFS_VM_RDWR_ENV */
73 #endif /* AFS_64BIT_CLIENT */
75 /* The following is used to ensure that new dcache's aren't obtained when
76 * the cache is nearly full.
78 int afs_WaitForCacheDrain = 0;
79 int afs_TruncateDaemonRunning = 0;
80 int afs_CacheTooFull = 0;
82 afs_int32 afs_dcentries; /* In-memory dcache entries */
85 int dcacheDisabled = 0;
87 extern struct dcache *afs_UFSGetDSlot();
88 extern struct volume *afs_UFSGetVolSlot();
89 extern int osi_UFSTruncate(), afs_osi_Read(), afs_osi_Write(), osi_UFSClose();
90 extern int afs_UFSRead(), afs_UFSWrite();
91 extern int afs_UFSHandleLink();
92 static int afs_UFSCacheFetchProc(), afs_UFSCacheStoreProc();
93 struct afs_cacheOps afs_UfsCacheOps = {
101 afs_UFSCacheFetchProc,
102 afs_UFSCacheStoreProc,
108 struct afs_cacheOps afs_MemCacheOps = {
110 afs_MemCacheTruncate,
116 afs_MemCacheFetchProc,
117 afs_MemCacheStoreProc,
123 int cacheDiskType; /*Type of backing disk for cache*/
124 struct afs_cacheOps *afs_cacheType;
133 * Warn about failing to store a file.
136 * acode : Associated error code.
137 * avolume : Volume involved.
138 * aflags : How to handle the output:
139 * aflags & 1: Print out on console
140 * aflags & 2: Print out on controlling tty
143 * Call this from close call when vnodeops is RCS unlocked.
146 void afs_StoreWarn(register afs_int32 acode, afs_int32 avolume, register afs_int32 aflags)
148 static char problem_fmt[] =
149 "afs: failed to store file in volume %d (%s)\n";
150 static char problem_fmt_w_error[] =
151 "afs: failed to store file in volume %d (error %d)\n";
152 static char netproblems[] = "network problems";
153 static char partfull[] = "partition full";
154 static char overquota[] = "over quota";
156 AFS_STATCNT(afs_StoreWarn);
162 afs_warn(problem_fmt, avolume, netproblems);
164 afs_warnuser(problem_fmt, avolume, netproblems);
167 if (acode == ENOSPC) {
172 afs_warn(problem_fmt, avolume, partfull);
174 afs_warnuser(problem_fmt, avolume, partfull);
178 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
179 * Instead ENOSPC will be sent...
181 if (acode == EDQUOT) {
186 afs_warn(problem_fmt, avolume, overquota);
188 afs_warnuser(problem_fmt, avolume, overquota);
196 afs_warn(problem_fmt_w_error, avolume, acode);
198 afs_warnuser(problem_fmt_w_error, avolume, acode);
202 void afs_MaybeWakeupTruncateDaemon(void)
204 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
205 afs_CacheTooFull = 1;
206 if (!afs_TruncateDaemonRunning)
207 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
208 } else if (!afs_TruncateDaemonRunning &&
209 afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
210 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
214 /* Keep statistics on run time for afs_CacheTruncateDaemon. This is a
215 * struct so we need only export one symbol for AIX.
217 static struct CTD_stats {
218 osi_timeval_t CTD_beforeSleep;
219 osi_timeval_t CTD_afterSleep;
220 osi_timeval_t CTD_sleepTime;
221 osi_timeval_t CTD_runTime;
225 u_int afs_min_cache = 0;
226 void afs_CacheTruncateDaemon(void)
228 osi_timeval_t CTD_tmpTime;
231 u_int dc_hiwat = (100-CM_DCACHECOUNTFREEPCT+CM_DCACHEEXTRAPCT)*afs_cacheFiles/100;
232 afs_min_cache = (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize)>>10;
234 osi_GetuTime(&CTD_stats.CTD_afterSleep);
235 afs_TruncateDaemonRunning = 1;
237 cb_lowat = ((CM_DCACHESPACEFREEPCT-CM_DCACHEEXTRAPCT)
238 * afs_cacheBlocks) / 100;
239 MObtainWriteLock(&afs_xdcache,266);
240 if (afs_CacheTooFull) {
241 int space_needed, slots_needed;
242 /* if we get woken up, we should try to clean something out */
243 for (counter = 0; counter < 10; counter++) {
244 space_needed = afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
245 slots_needed = dc_hiwat - afs_freeDCCount - afs_discardDCCount;
246 afs_GetDownD(slots_needed, &space_needed);
247 if ((space_needed <= 0) && (slots_needed <= 0)) {
250 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
253 if (!afs_CacheIsTooFull())
254 afs_CacheTooFull = 0;
256 MReleaseWriteLock(&afs_xdcache);
259 * This is a defensive check to try to avoid starving threads
260 * that may need the global lock so thay can help free some
261 * cache space. If this thread won't be sleeping or truncating
262 * any cache files then give up the global lock so other
263 * threads get a chance to run.
265 if ((afs_termState!=AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull &&
266 (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
267 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
271 * This is where we free the discarded cache elements.
273 while(afs_blocksDiscarded && !afs_WaitForCacheDrain &&
274 (afs_termState!=AFSOP_STOP_TRUNCDAEMON))
276 afs_FreeDiscardedDCache();
279 /* See if we need to continue to run. Someone may have
280 * signalled us while we were executing.
282 if (!afs_WaitForCacheDrain && !afs_CacheTooFull &&
283 (afs_termState!=AFSOP_STOP_TRUNCDAEMON))
285 /* Collect statistics on truncate daemon. */
286 CTD_stats.CTD_nSleeps++;
287 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
288 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
289 CTD_stats.CTD_beforeSleep);
290 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
292 afs_TruncateDaemonRunning = 0;
293 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
294 afs_TruncateDaemonRunning = 1;
296 osi_GetuTime(&CTD_stats.CTD_afterSleep);
297 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
298 CTD_stats.CTD_afterSleep);
299 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
301 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
303 afs_termState = AFSOP_STOP_AFSDB;
305 afs_termState = AFSOP_STOP_RXEVENT;
307 afs_osi_Wakeup(&afs_termState);
318 * Make adjustment for the new size in the disk cache entry
320 * Major Assumptions Here:
321 * Assumes that frag size is an integral power of two, less one,
322 * and that this is a two's complement machine. I don't
323 * know of any filesystems which violate this assumption...
326 * adc : Ptr to dcache entry.
327 * anewsize : New size desired.
330 void afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
332 register afs_int32 oldSize;
334 AFS_STATCNT(afs_AdjustSize);
336 adc->dflags |= DFEntryMod;
337 oldSize = ((adc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
338 adc->f.chunkBytes = newSize;
339 newSize = ((newSize + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
340 if (newSize > oldSize) {
341 /* We're growing the file, wakeup the daemon */
342 afs_MaybeWakeupTruncateDaemon();
344 afs_blocksUsed += (newSize - oldSize);
345 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
353 * This routine is responsible for moving at least one entry (but up
354 * to some number of them) from the LRU queue to the free queue.
357 * anumber : Number of entries that should ideally be moved.
358 * aneedSpace : How much space we need (1K blocks);
361 * The anumber parameter is just a hint; at least one entry MUST be
362 * moved, or we'll panic. We must be called with afs_xdcache
363 * write-locked. We should try to satisfy both anumber and aneedspace,
364 * whichever is more demanding - need to do several things:
365 * 1. only grab up to anumber victims if aneedSpace <= 0, not
366 * the whole set of MAXATONCE.
367 * 2. dynamically choose MAXATONCE to reflect severity of
368 * demand: something like (*aneedSpace >> (logChunk - 9))
369 * N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
370 * indicates that the cache is not properly configured/tuned or
371 * something. We should be able to automatically correct that problem.
374 #define MAXATONCE 16 /* max we can obtain at once */
375 static void afs_GetDownD(int anumber, int *aneedSpace)
379 struct VenusFid *afid;
383 register struct vcache *tvc;
384 afs_uint32 victims[MAXATONCE];
385 struct dcache *victimDCs[MAXATONCE];
386 afs_hyper_t victimTimes[MAXATONCE];/* youngest (largest LRU time) first */
387 afs_uint32 victimPtr; /* next free item in victim arrays */
388 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
389 afs_uint32 maxVictimPtr; /* where it is */
392 AFS_STATCNT(afs_GetDownD);
393 if (CheckLock(&afs_xdcache) != -1)
394 osi_Panic("getdownd nolock");
395 /* decrement anumber first for all dudes in free list */
396 /* SHOULD always decrement anumber first, even if aneedSpace >0,
397 * because we should try to free space even if anumber <=0 */
398 if (!aneedSpace || *aneedSpace <= 0) {
399 anumber -= afs_freeDCCount;
400 if (anumber <= 0) return; /* enough already free */
402 /* bounds check parameter */
403 if (anumber > MAXATONCE)
404 anumber = MAXATONCE; /* all we can do */
407 * The phase variable manages reclaims. Set to 0, the first pass,
408 * we don't reclaim active entries. Set to 1, we reclaim even active
412 for (i = 0; i < afs_cacheFiles; i++)
413 /* turn off all flags */
414 afs_indexFlags[i] &= ~IFFlag;
416 while (anumber > 0 || (aneedSpace && *aneedSpace >0)) {
417 /* find oldest entries for reclamation */
418 maxVictimPtr = victimPtr = 0;
419 hzero(maxVictimTime);
420 /* select victims from access time array */
421 for (i = 0; i < afs_cacheFiles; i++) {
422 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
423 /* skip if dirty or already free */
426 tdc = afs_indexTable[i];
427 if (tdc && (tdc->refCount != 0)) {
428 /* Referenced; can't use it! */
431 hset(vtime, afs_indexTimes[i]);
433 /* if we've already looked at this one, skip it */
434 if (afs_indexFlags[i] & IFFlag) continue;
436 if (victimPtr < MAXATONCE) {
437 /* if there's at least one free victim slot left */
438 victims[victimPtr] = i;
439 hset(victimTimes[victimPtr], vtime);
440 if (hcmp(vtime, maxVictimTime) > 0) {
441 hset(maxVictimTime, vtime);
442 maxVictimPtr = victimPtr;
446 else if (hcmp(vtime, maxVictimTime) < 0) {
448 * We're older than youngest victim, so we replace at
451 /* find youngest (largest LRU) victim */
453 if (j == victimPtr) osi_Panic("getdownd local");
455 hset(victimTimes[j], vtime);
456 /* recompute maxVictimTime */
457 hset(maxVictimTime, vtime);
458 for(j = 0; j < victimPtr; j++)
459 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
460 hset(maxVictimTime, victimTimes[j]);
466 /* now really reclaim the victims */
467 j = 0; /* flag to track if we actually got any of the victims */
468 /* first, hold all the victims, since we're going to release the lock
469 * during the truncate operation.
471 for(i=0; i < victimPtr; i++) {
472 tdc = afs_GetDSlot(victims[i], 0);
473 /* We got tdc->tlock(R) here */
474 if (tdc->refCount == 1)
478 ReleaseReadLock(&tdc->tlock);
479 if (!victimDCs[i]) afs_PutDCache(tdc);
481 for(i = 0; i < victimPtr; i++) {
482 /* q is first elt in dcache entry */
484 /* now, since we're dropping the afs_xdcache lock below, we
485 * have to verify, before proceeding, that there are no other
486 * references to this dcache entry, even now. Note that we
487 * compare with 1, since we bumped it above when we called
488 * afs_GetDSlot to preserve the entry's identity.
490 if (tdc && tdc->refCount == 1) {
491 unsigned char chunkFlags;
492 afs_size_t tchunkoffset;
494 /* xdcache is lower than the xvcache lock */
495 MReleaseWriteLock(&afs_xdcache);
496 MObtainReadLock(&afs_xvcache);
497 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
498 MReleaseReadLock(&afs_xvcache);
499 MObtainWriteLock(&afs_xdcache, 527);
501 if (tdc->refCount > 1) skip = 1;
503 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
504 chunkFlags = afs_indexFlags[tdc->index];
505 if (phase == 0 && osi_Active(tvc)) skip = 1;
506 if (phase > 0 && osi_Active(tvc) && (tvc->states & CDCLock)
507 && (chunkFlags & IFAnyPages)) skip = 1;
508 if (chunkFlags & IFDataMod) skip = 1;
509 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
510 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
511 ICL_TYPE_INT32, tdc->index,
513 ICL_HANDLE_OFFSET(tchunkoffset));
515 #if defined(AFS_SUN5_ENV)
517 * Now we try to invalidate pages. We do this only for
518 * Solaris. For other platforms, it's OK to recycle a
519 * dcache entry out from under a page, because the strategy
520 * function can call afs_GetDCache().
522 if (!skip && (chunkFlags & IFAnyPages)) {
525 MReleaseWriteLock(&afs_xdcache);
526 MObtainWriteLock(&tvc->vlock, 543);
527 if (tvc->multiPage) {
531 /* block locking pages */
532 tvc->vstates |= VPageCleaning;
533 /* block getting new pages */
535 MReleaseWriteLock(&tvc->vlock);
536 /* One last recheck */
537 MObtainWriteLock(&afs_xdcache, 333);
538 chunkFlags = afs_indexFlags[tdc->index];
539 if (tdc->refCount > 1
540 || (chunkFlags & IFDataMod)
541 || (osi_Active(tvc) && (tvc->states & CDCLock)
542 && (chunkFlags & IFAnyPages))) {
544 MReleaseWriteLock(&afs_xdcache);
547 MReleaseWriteLock(&afs_xdcache);
549 code = osi_VM_GetDownD(tvc, tdc);
551 MObtainWriteLock(&afs_xdcache,269);
552 /* we actually removed all pages, clean and dirty */
554 afs_indexFlags[tdc->index] &= ~(IFDirtyPages| IFAnyPages);
557 MReleaseWriteLock(&afs_xdcache);
559 MObtainWriteLock(&tvc->vlock, 544);
560 if (--tvc->activeV == 0 && (tvc->vstates & VRevokeWait)) {
561 tvc->vstates &= ~VRevokeWait;
562 afs_osi_Wakeup((char *)&tvc->vstates);
565 if (tvc->vstates & VPageCleaning) {
566 tvc->vstates &= ~VPageCleaning;
567 afs_osi_Wakeup((char *)&tvc->vstates);
570 MReleaseWriteLock(&tvc->vlock);
572 #endif /* AFS_SUN5_ENV */
574 MReleaseWriteLock(&afs_xdcache);
578 MObtainWriteLock(&afs_xdcache, 528);
579 if (afs_indexFlags[tdc->index] &
580 (IFDataMod | IFDirtyPages | IFAnyPages)) skip = 1;
581 if (tdc->refCount > 1) skip = 1;
583 #if defined(AFS_SUN5_ENV)
585 /* no vnode, so IFDirtyPages is spurious (we don't
586 * sweep dcaches on vnode recycling, so we can have
587 * DIRTYPAGES set even when all pages are gone). Just
589 * Hold vcache lock to prevent vnode from being
590 * created while we're clearing IFDirtyPages.
592 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
596 /* skip this guy and mark him as recently used */
597 afs_indexFlags[tdc->index] |= IFFlag;
598 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
599 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
600 ICL_TYPE_INT32, tdc->index,
602 ICL_HANDLE_OFFSET(tchunkoffset));
605 /* flush this dude from the data cache and reclaim;
606 * first, make sure no one will care that we damage
607 * it, by removing it from all hash tables. Then,
608 * melt it down for parts. Note that any concurrent
609 * (new possibility!) calls to GetDownD won't touch
610 * this guy because his reference count is > 0. */
611 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
612 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
613 ICL_TYPE_INT32, tdc->index,
615 ICL_HANDLE_OFFSET(tchunkoffset));
617 AFS_STATCNT(afs_gget);
619 afs_HashOutDCache(tdc);
620 if (tdc->f.chunkBytes != 0) {
623 *aneedSpace -= (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
628 afs_DiscardDCache(tdc);
633 j = 1; /* we reclaimed at least one victim */
640 /* Phase is 0 and no one was found, so try phase 1 (ignore
641 * osi_Active flag) */
644 for (i = 0; i < afs_cacheFiles; i++)
645 /* turn off all flags */
646 afs_indexFlags[i] &= ~IFFlag;
650 /* found no one in phase 1, we're hosed */
651 if (victimPtr == 0) break;
653 } /* big while loop */
660 * Description: remove adc from any hash tables that would allow it to be located
661 * again by afs_FindDCache or afs_GetDCache.
663 * Parameters: adc -- pointer to dcache entry to remove from hash tables.
665 * Locks: Must have the afs_xdcache lock write-locked to call this function.
667 int afs_HashOutDCache(struct dcache *adc)
672 AFS_STATCNT(afs_glink);
674 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
676 /* if this guy is in the hash table, pull him out */
677 if (adc->f.fid.Fid.Volume != 0) {
678 /* remove entry from first hash chains */
679 i = DCHash(&adc->f.fid, adc->f.chunk);
680 us = afs_dchashTbl[i];
681 if (us == adc->index) {
682 /* first dude in the list */
683 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
686 /* somewhere on the chain */
687 while (us != NULLIDX) {
688 if (afs_dcnextTbl[us] == adc->index) {
689 /* found item pointing at the one to delete */
690 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
693 us = afs_dcnextTbl[us];
695 if (us == NULLIDX) osi_Panic("dcache hc");
697 /* remove entry from *other* hash chain */
698 i = DVHash(&adc->f.fid);
699 us = afs_dvhashTbl[i];
700 if (us == adc->index) {
701 /* first dude in the list */
702 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
705 /* somewhere on the chain */
706 while (us != NULLIDX) {
707 if (afs_dvnextTbl[us] == adc->index) {
708 /* found item pointing at the one to delete */
709 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
712 us = afs_dvnextTbl[us];
714 if (us == NULLIDX) osi_Panic("dcache hv");
718 /* prevent entry from being found on a reboot (it is already out of
719 * the hash table, but after a crash, we just look at fid fields of
720 * stable (old) entries).
722 adc->f.fid.Fid.Volume = 0; /* invalid */
724 /* mark entry as modified */
725 adc->dflags |= DFEntryMod;
729 } /*afs_HashOutDCache */
736 * Flush the given dcache entry, pulling it from hash chains
737 * and truncating the associated cache file.
740 * adc: Ptr to dcache entry to flush.
743 * This routine must be called with the afs_xdcache lock held
747 void afs_FlushDCache(register struct dcache *adc)
749 AFS_STATCNT(afs_FlushDCache);
751 * Bump the number of cache files flushed.
753 afs_stats_cmperf.cacheFlushes++;
755 /* remove from all hash tables */
756 afs_HashOutDCache(adc);
758 /* Free its space; special case null operation, since truncate operation
759 * in UFS is slow even in this case, and this allows us to pre-truncate
760 * these files at more convenient times with fewer locks set
761 * (see afs_GetDownD).
763 if (adc->f.chunkBytes != 0) {
764 afs_DiscardDCache(adc);
765 afs_MaybeWakeupTruncateDaemon();
770 if (afs_WaitForCacheDrain) {
771 if (afs_blocksUsed <=
772 (CM_CACHESIZEDRAINEDPCT*afs_cacheBlocks)/100) {
773 afs_WaitForCacheDrain = 0;
774 afs_osi_Wakeup(&afs_WaitForCacheDrain);
777 } /*afs_FlushDCache*/
783 * Description: put a dcache entry on the free dcache entry list.
785 * Parameters: adc -- dcache entry to free
787 * Environment: called with afs_xdcache lock write-locked.
789 static void afs_FreeDCache(register struct dcache *adc)
791 /* Thread on free list, update free list count and mark entry as
792 * freed in its indexFlags element. Also, ensure DCache entry gets
793 * written out (set DFEntryMod).
796 afs_dvnextTbl[adc->index] = afs_freeDCList;
797 afs_freeDCList = adc->index;
799 afs_indexFlags[adc->index] |= IFFree;
800 adc->dflags |= DFEntryMod;
802 if (afs_WaitForCacheDrain) {
803 if ((afs_blocksUsed - afs_blocksDiscarded) <=
804 (CM_CACHESIZEDRAINEDPCT*afs_cacheBlocks)/100) {
805 afs_WaitForCacheDrain = 0;
806 afs_osi_Wakeup(&afs_WaitForCacheDrain);
815 * Discard the cache element by moving it to the discardDCList.
816 * This puts the cache element into a quasi-freed state, where
817 * the space may be reused, but the file has not been truncated.
819 * Major Assumptions Here:
820 * Assumes that frag size is an integral power of two, less one,
821 * and that this is a two's complement machine. I don't
822 * know of any filesystems which violate this assumption...
825 * adc : Ptr to dcache entry.
828 * Must be called with afs_xdcache write-locked.
831 static void afs_DiscardDCache(register struct dcache *adc)
833 register afs_int32 size;
835 AFS_STATCNT(afs_DiscardDCache);
837 osi_Assert(adc->refCount == 1);
839 size = ((adc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
840 afs_blocksDiscarded += size;
841 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
843 afs_dvnextTbl[adc->index] = afs_discardDCList;
844 afs_discardDCList = adc->index;
845 afs_discardDCCount++;
847 adc->f.fid.Fid.Volume = 0;
848 adc->dflags |= DFEntryMod;
849 afs_indexFlags[adc->index] |= IFDiscarded;
851 if (afs_WaitForCacheDrain) {
852 if ((afs_blocksUsed - afs_blocksDiscarded) <=
853 (CM_CACHESIZEDRAINEDPCT*afs_cacheBlocks)/100) {
854 afs_WaitForCacheDrain = 0;
855 afs_osi_Wakeup(&afs_WaitForCacheDrain);
859 } /*afs_DiscardDCache*/
862 * afs_FreeDiscardedDCache
865 * Free the next element on the list of discarded cache elements.
867 static void afs_FreeDiscardedDCache(void)
869 register struct dcache *tdc;
870 register struct osi_file *tfile;
871 register afs_int32 size;
873 AFS_STATCNT(afs_FreeDiscardedDCache);
875 MObtainWriteLock(&afs_xdcache,510);
876 if (!afs_blocksDiscarded) {
877 MReleaseWriteLock(&afs_xdcache);
882 * Get an entry from the list of discarded cache elements
884 tdc = afs_GetDSlot(afs_discardDCList, 0);
885 osi_Assert(tdc->refCount == 1);
886 ReleaseReadLock(&tdc->tlock);
888 afs_discardDCList = afs_dvnextTbl[tdc->index];
889 afs_dvnextTbl[tdc->index] = NULLIDX;
890 afs_discardDCCount--;
891 size = ((tdc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
892 afs_blocksDiscarded -= size;
893 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
894 /* We can lock because we just took it off the free list */
895 ObtainWriteLock(&tdc->lock, 626);
896 MReleaseWriteLock(&afs_xdcache);
899 * Truncate the element to reclaim its space
901 tfile = afs_CFileOpen(tdc->f.inode);
902 afs_CFileTruncate(tfile, 0);
903 afs_CFileClose(tfile);
904 afs_AdjustSize(tdc, 0);
907 * Free the element we just truncated
909 MObtainWriteLock(&afs_xdcache,511);
910 afs_indexFlags[tdc->index] &= ~IFDiscarded;
912 ReleaseWriteLock(&tdc->lock);
914 MReleaseWriteLock(&afs_xdcache);
918 * afs_MaybeFreeDiscardedDCache
921 * Free as many entries from the list of discarded cache elements
922 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
927 int afs_MaybeFreeDiscardedDCache(void)
930 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
932 while (afs_blocksDiscarded &&
933 (afs_blocksUsed > (CM_WAITFORDRAINPCT*afs_cacheBlocks)/100)) {
934 afs_FreeDiscardedDCache();
943 * Try to free up a certain number of disk slots.
946 * anumber : Targeted number of disk slots to free up.
949 * Must be called with afs_xdcache write-locked.
951 static void afs_GetDownDSlot(int anumber)
953 struct afs_q *tq, *nq;
959 AFS_STATCNT(afs_GetDownDSlot);
960 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
961 osi_Panic("diskless getdowndslot");
963 if (CheckLock(&afs_xdcache) != -1)
964 osi_Panic("getdowndslot nolock");
966 /* decrement anumber first for all dudes in free list */
967 for(tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
970 return; /* enough already free */
972 for(cnt=0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
974 tdc = (struct dcache *) tq; /* q is first elt in dcache entry */
975 nq = QPrev(tq); /* in case we remove it */
976 if (tdc->refCount == 0) {
977 if ((ix=tdc->index) == NULLIDX) osi_Panic("getdowndslot");
978 /* pull the entry out of the lruq and put it on the free list */
981 /* write-through if modified */
982 if (tdc->dflags & DFEntryMod) {
983 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
985 * ask proxy to do this for us - we don't have the stack space
987 while (tdc->dflags & DFEntryMod) {
990 s = SPLOCK(afs_sgibklock);
991 if (afs_sgibklist == NULL) {
992 /* if slot is free, grab it. */
994 SV_SIGNAL(&afs_sgibksync);
996 /* wait for daemon to (start, then) finish. */
997 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1001 tdc->dflags &= ~DFEntryMod;
1002 afs_WriteDCache(tdc, 1);
1009 struct osi_file * f = (struct osi_file *)tdc->ihint;
1017 /* finally put the entry in the free list */
1018 afs_indexTable[ix] = NULL;
1019 afs_indexFlags[ix] &= ~IFEverUsed;
1020 tdc->index = NULLIDX;
1021 tdc->lruq.next = (struct afs_q *) afs_freeDSList;
1022 afs_freeDSList = tdc;
1026 } /*afs_GetDownDSlot*/
1033 * Increment the reference count on a disk cache entry,
1034 * which already has a non-zero refcount. In order to
1035 * increment the refcount of a zero-reference entry, you
1036 * have to hold afs_xdcache.
1039 * adc : Pointer to the dcache entry to increment.
1042 * Nothing interesting.
1044 int afs_RefDCache(struct dcache *adc)
1046 ObtainWriteLock(&adc->tlock, 627);
1047 if (adc->refCount < 0)
1048 osi_Panic("RefDCache: negative refcount");
1050 ReleaseWriteLock(&adc->tlock);
1059 * Decrement the reference count on a disk cache entry.
1062 * ad : Ptr to the dcache entry to decrement.
1065 * Nothing interesting.
1067 int afs_PutDCache(register struct dcache *adc)
1069 AFS_STATCNT(afs_PutDCache);
1070 ObtainWriteLock(&adc->tlock, 276);
1071 if (adc->refCount <= 0)
1072 osi_Panic("putdcache");
1074 ReleaseWriteLock(&adc->tlock);
1083 * Try to discard all data associated with this file from the
1087 * avc : Pointer to the cache info for the file.
1090 * Both pvnLock and lock are write held.
1092 void afs_TryToSmush(register struct vcache *avc, struct AFS_UCRED *acred,
1095 register struct dcache *tdc;
1098 AFS_STATCNT(afs_TryToSmush);
1099 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1100 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length));
1101 sync = 1; /* XX Temp testing XX*/
1103 #if defined(AFS_SUN5_ENV)
1104 ObtainWriteLock(&avc->vlock, 573);
1105 avc->activeV++; /* block new getpages */
1106 ReleaseWriteLock(&avc->vlock);
1109 /* Flush VM pages */
1110 osi_VM_TryToSmush(avc, acred, sync);
1113 * Get the hash chain containing all dce's for this fid
1115 i = DVHash(&avc->fid);
1116 MObtainWriteLock(&afs_xdcache,277);
1117 for(index = afs_dvhashTbl[i]; index != NULLIDX; index=i) {
1118 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1119 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1120 int releaseTlock = 1;
1121 tdc = afs_GetDSlot(index, NULL);
1122 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1124 if ((afs_indexFlags[index] & IFDataMod) == 0 &&
1125 tdc->refCount == 1) {
1126 ReleaseReadLock(&tdc->tlock);
1128 afs_FlushDCache(tdc);
1131 afs_indexTable[index] = 0;
1133 if (releaseTlock) ReleaseReadLock(&tdc->tlock);
1137 #if defined(AFS_SUN5_ENV)
1138 ObtainWriteLock(&avc->vlock, 545);
1139 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1140 avc->vstates &= ~VRevokeWait;
1141 afs_osi_Wakeup((char *)&avc->vstates);
1143 ReleaseWriteLock(&avc->vlock);
1145 MReleaseWriteLock(&afs_xdcache);
1147 * It's treated like a callback so that when we do lookups we'll invalidate the unique bit if any
1148 * trytoSmush occured during the lookup call
1157 * Given the cached info for a file and a byte offset into the
1158 * file, make sure the dcache entry for that file and containing
1159 * the given byte is available, returning it to our caller.
1162 * avc : Pointer to the (held) vcache entry to look in.
1163 * abyte : Which byte we want to get to.
1166 * Pointer to the dcache entry covering the file & desired byte,
1167 * or NULL if not found.
1170 * The vcache entry is held upon entry.
1173 struct dcache *afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1176 register afs_int32 i, index;
1177 register struct dcache *tdc;
1179 AFS_STATCNT(afs_FindDCache);
1180 chunk = AFS_CHUNK(abyte);
1183 * Hash on the [fid, chunk] and get the corresponding dcache index
1184 * after write-locking the dcache.
1186 i = DCHash(&avc->fid, chunk);
1187 MObtainWriteLock(&afs_xdcache,278);
1188 for(index = afs_dchashTbl[i]; index != NULLIDX;) {
1189 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1190 tdc = afs_GetDSlot(index, NULL);
1191 ReleaseReadLock(&tdc->tlock);
1192 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1193 break; /* leaving refCount high for caller */
1197 index = afs_dcnextTbl[index];
1199 MReleaseWriteLock(&afs_xdcache);
1200 if (index != NULLIDX) {
1201 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1202 hadd32(afs_indexCounter, 1);
1208 } /*afs_FindDCache*/
1212 * afs_UFSCacheStoreProc
1215 * Called upon store.
1218 * acall : Ptr to the Rx call structure involved.
1219 * afile : Ptr to the related file descriptor.
1220 * alen : Size of the file in bytes.
1221 * avc : Ptr to the vcache entry.
1222 * shouldWake : is it "safe" to return early from close() ?
1223 * abytesToXferP : Set to the number of bytes to xfer.
1224 * NOTE: This parameter is only used if AFS_NOSTATS
1226 * abytesXferredP : Set to the number of bytes actually xferred.
1227 * NOTE: This parameter is only used if AFS_NOSTATS
1231 * Nothing interesting.
1233 static int afs_UFSCacheStoreProc(register struct rx_call *acall,
1234 struct osi_file *afile, register afs_int32 alen, struct vcache *avc,
1235 int *shouldWake, afs_size_t *abytesToXferP, afs_size_t *abytesXferredP)
1237 afs_int32 code, got;
1238 register char *tbuffer;
1241 AFS_STATCNT(UFS_CacheStoreProc);
1245 * In this case, alen is *always* the amount of data we'll be trying
1248 (*abytesToXferP) = alen;
1249 (*abytesXferredP) = 0;
1250 #endif /* AFS_NOSTATS */
1252 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1253 ICL_TYPE_FID, &(avc->fid),
1254 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length),
1255 ICL_TYPE_INT32, alen);
1256 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1258 tlen = (alen > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : alen);
1259 got = afs_osi_Read(afile, -1, tbuffer, tlen);
1261 #if !defined(AFS_SUN5_ENV) && !defined(AFS_OSF_ENV) && !defined(AFS_SGI64_ENV) && !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN_ENV) && !defined(AFS_FBSD_ENV)
1262 || (got != tlen && getuerror())
1265 osi_FreeLargeSpace(tbuffer);
1268 afs_Trace2(afs_iclSetp, CM_TRACE_STOREPROC2,
1269 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(*tbuffer),
1270 ICL_TYPE_INT32, got);
1272 code = rx_Write(acall, tbuffer, got); /* writing 0 bytes will
1273 * push a short packet. Is that really what we want, just because the
1274 * data didn't come back from the disk yet? Let's try it and see. */
1277 (*abytesXferredP) += code;
1278 #endif /* AFS_NOSTATS */
1280 osi_FreeLargeSpace(tbuffer);
1285 * If file has been locked on server, we can allow the store
1288 if (shouldWake && *shouldWake && (rx_GetRemoteStatus(acall) & 1)) {
1289 *shouldWake = 0; /* only do this once */
1293 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1294 ICL_TYPE_FID, &(avc->fid),
1295 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length),
1296 ICL_TYPE_INT32, alen);
1297 osi_FreeLargeSpace(tbuffer);
1300 } /* afs_UFSCacheStoreProc*/
1304 * afs_UFSCacheFetchProc
1307 * Routine called on fetch; also tells people waiting for data
1308 * that more has arrived.
1311 * acall : Ptr to the Rx call structure.
1312 * afile : File descriptor for the cache file.
1313 * abase : Base offset to fetch.
1314 * adc : Ptr to the dcache entry for the file, write-locked.
1315 * avc : Ptr to the vcache entry for the file.
1316 * abytesToXferP : Set to the number of bytes to xfer.
1317 * NOTE: This parameter is only used if AFS_NOSTATS
1319 * abytesXferredP : Set to the number of bytes actually xferred.
1320 * NOTE: This parameter is only used if AFS_NOSTATS
1324 * Nothing interesting.
1327 static int afs_UFSCacheFetchProc(register struct rx_call *acall,
1328 struct osi_file *afile, afs_size_t abase, struct dcache *adc,
1329 struct vcache *avc, afs_size_t *abytesToXferP,
1330 afs_size_t *abytesXferredP, afs_int32 lengthFound)
1333 register afs_int32 code;
1334 register char *tbuffer;
1338 AFS_STATCNT(UFS_CacheFetchProc);
1339 osi_Assert(WriteLocked(&adc->lock));
1340 afile->offset = 0; /* Each time start from the beginning */
1341 length = lengthFound;
1343 (*abytesToXferP) = 0;
1344 (*abytesXferredP) = 0;
1345 #endif /* AFS_NOSTATS */
1346 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1347 adc->validPos = abase;
1351 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
1353 length = ntohl(length);
1354 if (code != sizeof(afs_int32)) {
1355 osi_FreeLargeSpace(tbuffer);
1356 code = rx_Error(acall);
1357 return (code?code:-1); /* try to return code, not -1 */
1361 * The fetch protocol is extended for the AFS/DFS translator
1362 * to allow multiple blocks of data, each with its own length,
1363 * to be returned. As long as the top bit is set, there are more
1366 * We do not do this for AFS file servers because they sometimes
1367 * return large negative numbers as the transfer size.
1369 if (avc->states & CForeign) {
1370 moredata = length & 0x80000000;
1371 length &= ~0x80000000;
1376 (*abytesToXferP) += length;
1377 #endif /* AFS_NOSTATS */
1378 while (length > 0) {
1379 tlen = (length > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : length);
1380 #ifdef RX_KERNEL_TRACE
1381 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP,
1382 ICL_TYPE_STRING, "before rx_Read");
1385 code = rx_Read(acall, tbuffer, tlen);
1387 #ifdef RX_KERNEL_TRACE
1388 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP,
1389 ICL_TYPE_STRING, "after rx_Read");
1392 (*abytesXferredP) += code;
1393 #endif /* AFS_NOSTATS */
1395 osi_FreeLargeSpace(tbuffer);
1396 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64READ,
1397 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
1398 ICL_TYPE_INT32, length);
1401 code = afs_osi_Write(afile, -1, tbuffer, tlen);
1403 osi_FreeLargeSpace(tbuffer);
1408 adc->validPos = abase;
1410 if (afs_osi_Wakeup(&adc->validPos) == 0)
1411 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE,
1412 ICL_TYPE_STRING, __FILE__,
1413 ICL_TYPE_INT32, __LINE__,
1414 ICL_TYPE_POINTER, adc,
1415 ICL_TYPE_INT32, adc->dflags);
1417 afs_osi_Wakeup(&adc->validPos);
1421 osi_FreeLargeSpace(tbuffer);
1424 } /* afs_UFSCacheFetchProc*/
1430 * This function is called to obtain a reference to data stored in
1431 * the disk cache, locating a chunk of data containing the desired
1432 * byte and returning a reference to the disk cache entry, with its
1433 * reference count incremented.
1437 * avc : Ptr to a vcache entry (unlocked)
1438 * abyte : Byte position in the file desired
1439 * areq : Request structure identifying the requesting user.
1440 * aflags : Settings as follows:
1442 * 2 : Return after creating entry.
1443 * 4 : called from afs_vnop_write.c
1444 * *alen contains length of data to be written.
1446 * aoffset : Set to the offset within the chunk where the resident
1448 * alen : Set to the number of bytes of data after the desired
1449 * byte (including the byte itself) which can be read
1453 * The vcache entry pointed to by avc is unlocked upon entry.
1457 struct AFSVolSync tsync;
1458 struct AFSFetchStatus OutStatus;
1459 struct AFSCallBack CallBack;
1463 * Update the vnode-to-dcache hint if we can get the vnode lock
1464 * right away. Assumes dcache entry is at least read-locked.
1466 void updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1468 if (!lockVc || 0 == NBObtainWriteLock(&v->lock,src)) {
1469 if (hsame(v->m.DataVersion, d->f.versionNo) && v->callback) {
1471 v->quick.stamp = d->stamp = MakeStamp();
1472 v->quick.minLoc = AFS_CHUNKTOBASE(d->f.chunk);
1473 /* Don't think I need these next two lines forever */
1474 v->quick.len = d->f.chunkBytes;
1477 if (lockVc) ReleaseWriteLock(&v->lock);
1481 /* avc - Write-locked unless aflags & 1 */
1482 struct dcache *afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1483 register struct vrequest *areq, afs_size_t *aoffset, afs_size_t *alen,
1486 register afs_int32 i, code, code1, shortcut , adjustsize=0;
1491 afs_size_t maxGoodLength; /* amount of good data at server */
1492 struct rx_call *tcall;
1493 afs_size_t Position = 0;
1494 #ifdef AFS_64BIT_CLIENT
1496 #endif /* AFS_64BIT_CLIENT */
1497 afs_int32 size, tlen; /* size of segment to transfer */
1498 afs_size_t lengthFound; /* as returned from server */
1499 struct tlocal1 *tsmall = 0;
1500 register struct dcache *tdc;
1501 register struct osi_file *file;
1502 register struct conn *tc;
1504 struct server *newCallback;
1505 char setNewCallback;
1506 char setVcacheStatus;
1507 char doVcacheUpdate;
1509 int doAdjustSize = 0;
1510 int doReallyAdjustSize = 0;
1511 int overWriteWholeChunk = 0;
1515 struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
1516 osi_timeval_t xferStartTime, /*FS xfer start time*/
1517 xferStopTime; /*FS xfer stop time*/
1518 afs_size_t bytesToXfer; /* # bytes to xfer*/
1519 afs_size_t bytesXferred; /* # bytes actually xferred*/
1520 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats*/
1521 int fromReplica; /*Are we reading from a replica?*/
1522 int numFetchLoops; /*# times around the fetch/analyze loop*/
1523 #endif /* AFS_NOSTATS */
1525 AFS_STATCNT(afs_GetDCache);
1530 setLocks = aflags & 1;
1533 * Determine the chunk number and offset within the chunk corresponding
1534 * to the desired byte.
1536 if (avc->fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1540 chunk = AFS_CHUNK(abyte);
1543 /* come back to here if we waited for the cache to drain. */
1546 setNewCallback = setVcacheStatus = 0;
1550 ObtainWriteLock(&avc->lock, 616);
1552 ObtainReadLock(&avc->lock);
1557 * avc->lock(R) if setLocks && !slowPass
1558 * avc->lock(W) if !setLocks || slowPass
1563 /* check hints first! (might could use bcmp or some such...) */
1564 if ((tdc = avc->h1.dchint)) {
1568 * The locking order between afs_xdcache and dcache lock matters.
1569 * The hint dcache entry could be anywhere, even on the free list.
1570 * Locking afs_xdcache ensures that noone is trying to pull dcache
1571 * entries from the free list, and thereby assuming them to be not
1572 * referenced and not locked.
1574 MObtainReadLock(&afs_xdcache);
1575 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1578 (tdc->index != NULLIDX) && !FidCmp(&tdc->f.fid, &avc->fid) &&
1579 chunk == tdc->f.chunk &&
1580 !(afs_indexFlags[tdc->index] & (IFFree|IFDiscarded))) {
1581 /* got the right one. It might not be the right version, and it
1582 * might be fetching, but it's the right dcache entry.
1584 /* All this code should be integrated better with what follows:
1585 * I can save a good bit more time under a write lock if I do..
1587 ObtainWriteLock(&tdc->tlock, 603);
1589 ReleaseWriteLock(&tdc->tlock);
1591 MReleaseReadLock(&afs_xdcache);
1594 if (hsame(tdc->f.versionNo, avc->m.DataVersion) &&
1595 !(tdc->dflags & DFFetching)) {
1597 afs_stats_cmperf.dcacheHits++;
1598 MObtainWriteLock(&afs_xdcache, 559);
1599 QRemove(&tdc->lruq);
1600 QAdd(&afs_DLRU, &tdc->lruq);
1601 MReleaseWriteLock(&afs_xdcache);
1604 * avc->lock(R) if setLocks && !slowPass
1605 * avc->lock(W) if !setLocks || slowPass
1611 if (dcLocked) ReleaseSharedLock(&tdc->lock);
1612 MReleaseReadLock(&afs_xdcache);
1620 * avc->lock(R) if setLocks && !slowPass
1621 * avc->lock(W) if !setLocks || slowPass
1622 * tdc->lock(S) if tdc
1625 if (!tdc) { /* If the hint wasn't the right dcache entry */
1627 * Hash on the [fid, chunk] and get the corresponding dcache index
1628 * after write-locking the dcache.
1633 * avc->lock(R) if setLocks && !slowPass
1634 * avc->lock(W) if !setLocks || slowPass
1637 i = DCHash(&avc->fid, chunk);
1638 /* check to make sure our space is fine */
1639 afs_MaybeWakeupTruncateDaemon();
1641 MObtainWriteLock(&afs_xdcache,280);
1643 for (index = afs_dchashTbl[i]; index != NULLIDX; ) {
1644 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1645 tdc = afs_GetDSlot(index, NULL);
1646 ReleaseReadLock(&tdc->tlock);
1649 * avc->lock(R) if setLocks && !slowPass
1650 * avc->lock(W) if !setLocks || slowPass
1653 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1654 /* Move it up in the beginning of the list */
1655 if (afs_dchashTbl[i] != index) {
1656 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1657 afs_dcnextTbl[index] = afs_dchashTbl[i];
1658 afs_dchashTbl[i] = index;
1660 MReleaseWriteLock(&afs_xdcache);
1661 ObtainSharedLock(&tdc->lock, 606);
1662 break; /* leaving refCount high for caller */
1668 index = afs_dcnextTbl[index];
1672 * If we didn't find the entry, we'll create one.
1674 if (index == NULLIDX) {
1677 * avc->lock(R) if setLocks
1678 * avc->lock(W) if !setLocks
1681 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1682 avc, ICL_TYPE_INT32, chunk);
1684 /* Make sure there is a free dcache entry for us to use */
1685 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1687 if (!setLocks) avc->states |= CDCLock;
1688 afs_GetDownD(5, (int*)0); /* just need slots */
1689 if (!setLocks) avc->states &= ~CDCLock;
1690 if (afs_discardDCList != NULLIDX || afs_freeDCList != NULLIDX)
1692 /* If we can't get space for 5 mins we give up and panic */
1693 if (++downDCount > 300)
1694 osi_Panic("getdcache");
1695 MReleaseWriteLock(&afs_xdcache);
1698 * avc->lock(R) if setLocks
1699 * avc->lock(W) if !setLocks
1701 afs_osi_Wait(1000, 0, 0);
1706 if (afs_discardDCList == NULLIDX ||
1707 ((aflags & 2) && afs_freeDCList != NULLIDX)) {
1709 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1710 tdc = afs_GetDSlot(afs_freeDCList, 0);
1711 osi_Assert(tdc->refCount == 1);
1712 ReleaseReadLock(&tdc->tlock);
1713 ObtainWriteLock(&tdc->lock, 604);
1714 afs_freeDCList = afs_dvnextTbl[tdc->index];
1717 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1718 tdc = afs_GetDSlot(afs_discardDCList, 0);
1719 osi_Assert(tdc->refCount == 1);
1720 ReleaseReadLock(&tdc->tlock);
1721 ObtainWriteLock(&tdc->lock, 605);
1722 afs_discardDCList = afs_dvnextTbl[tdc->index];
1723 afs_discardDCCount--;
1724 size = ((tdc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;
1725 afs_blocksDiscarded -= size;
1726 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1728 /* Truncate the chunk so zeroes get filled properly */
1729 file = afs_CFileOpen(tdc->f.inode);
1730 afs_CFileTruncate(file, 0);
1731 afs_CFileClose(file);
1732 afs_AdjustSize(tdc, 0);
1738 * avc->lock(R) if setLocks
1739 * avc->lock(W) if !setLocks
1745 * Fill in the newly-allocated dcache record.
1747 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1748 tdc->f.fid = avc->fid;
1749 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1750 hones(tdc->f.versionNo); /* invalid value */
1751 tdc->f.chunk = chunk;
1752 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1754 if (tdc->lruq.prev == &tdc->lruq) osi_Panic("lruq 1");
1757 * Now add to the two hash chains - note that i is still set
1758 * from the above DCHash call.
1760 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1761 afs_dchashTbl[i] = tdc->index;
1762 i = DVHash(&avc->fid);
1763 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1764 afs_dvhashTbl[i] = tdc->index;
1765 tdc->dflags = DFEntryMod;
1768 afs_MaybeWakeupTruncateDaemon();
1769 MReleaseWriteLock(&afs_xdcache);
1770 ConvertWToSLock(&tdc->lock);
1772 } /* vcache->dcache hint failed */
1776 * avc->lock(R) if setLocks && !slowPass
1777 * avc->lock(W) if !setLocks || slowPass
1781 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1782 ICL_TYPE_POINTER, tdc,
1783 ICL_TYPE_INT32, hgetlo(tdc->f.versionNo),
1784 ICL_TYPE_INT32, hgetlo(avc->m.DataVersion));
1786 * Here we have the entry in tdc, with its refCount incremented.
1787 * Note: we don't use the S-lock on avc; it costs concurrency when
1788 * storing a file back to the server.
1792 * Not a newly created file so we need to check the file's length and
1793 * compare data versions since someone could have changed the data or we're
1794 * reading a file written elsewhere. We only want to bypass doing no-op
1795 * read rpcs on newly created files (dv of 0) since only then we guarantee
1796 * that this chunk's data hasn't been filled by another client.
1798 size = AFS_CHUNKSIZE(abyte);
1799 if (aflags & 4) /* called from write */
1801 else /* called from read */
1802 tlen = tdc->validPos - abyte;
1803 Position = AFS_CHUNKTOBASE(chunk);
1804 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3,
1805 ICL_TYPE_INT32, tlen,
1806 ICL_TYPE_INT32, aflags,
1807 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(abyte),
1808 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(Position));
1809 if ((aflags & 4) && (hiszero(avc->m.DataVersion)))
1811 if ((aflags & 4) && (abyte == Position) && (tlen >= size))
1812 overWriteWholeChunk = 1;
1813 if (doAdjustSize || overWriteWholeChunk) {
1814 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1816 #ifdef AFS_SGI64_ENV
1817 if (doAdjustSize) adjustsize = NBPP;
1818 #else /* AFS_SGI64_ENV */
1819 if (doAdjustSize) adjustsize = 8192;
1820 #endif /* AFS_SGI64_ENV */
1821 #else /* AFS_SGI_ENV */
1822 if (doAdjustSize) adjustsize = 4096;
1823 #endif /* AFS_SGI_ENV */
1824 if (AFS_CHUNKTOBASE(chunk)+adjustsize >= avc->m.Length &&
1825 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1826 #if defined(AFS_SUN_ENV) || defined(AFS_OSF_ENV)
1827 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
1829 if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
1831 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1832 !hsame(avc->m.DataVersion, tdc->f.versionNo))
1833 doReallyAdjustSize = 1;
1835 if (doReallyAdjustSize || overWriteWholeChunk) {
1836 /* no data in file to read at this position */
1837 UpgradeSToWLock(&tdc->lock, 607);
1839 file = afs_CFileOpen(tdc->f.inode);
1840 afs_CFileTruncate(file, 0);
1841 afs_CFileClose(file);
1842 afs_AdjustSize(tdc, 0);
1843 hset(tdc->f.versionNo, avc->m.DataVersion);
1844 tdc->dflags |= DFEntryMod;
1846 ConvertWToSLock(&tdc->lock);
1851 * We must read in the whole chunk if the version number doesn't
1855 /* don't need data, just a unique dcache entry */
1856 ObtainWriteLock(&afs_xdcache, 608);
1857 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1858 hadd32(afs_indexCounter, 1);
1859 ReleaseWriteLock(&afs_xdcache);
1861 updateV2DC(setLocks, avc, tdc, 553);
1862 if (vType(avc) == VDIR)
1865 *aoffset = AFS_CHUNKOFFSET(abyte);
1866 if (tdc->validPos < abyte)
1867 *alen = (afs_size_t) 0;
1869 *alen = tdc->validPos - abyte;
1870 ReleaseSharedLock(&tdc->lock);
1873 ReleaseWriteLock(&avc->lock);
1875 ReleaseReadLock(&avc->lock);
1877 return tdc; /* check if we're done */
1882 * avc->lock(R) if setLocks && !slowPass
1883 * avc->lock(W) if !setLocks || slowPass
1886 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
1888 setNewCallback = setVcacheStatus = 0;
1892 * avc->lock(R) if setLocks && !slowPass
1893 * avc->lock(W) if !setLocks || slowPass
1896 if (!hsame(avc->m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
1898 * Version number mismatch.
1900 UpgradeSToWLock(&tdc->lock, 609);
1903 * If data ever existed for this vnode, and this is a text object,
1904 * do some clearing. Now, you'd think you need only do the flush
1905 * when VTEXT is on, but VTEXT is turned off when the text object
1906 * is freed, while pages are left lying around in memory marked
1907 * with this vnode. If we would reactivate (create a new text
1908 * object from) this vnode, we could easily stumble upon some of
1909 * these old pages in pagein. So, we always flush these guys.
1910 * Sun has a wonderful lack of useful invariants in this system.
1912 * avc->flushDV is the data version # of the file at the last text
1913 * flush. Clearly, at least, we don't have to flush the file more
1914 * often than it changes
1916 if (hcmp(avc->flushDV, avc->m.DataVersion) < 0) {
1918 * By here, the cache entry is always write-locked. We can
1919 * deadlock if we call osi_Flush with the cache entry locked...
1920 * Unlock the dcache too.
1922 ReleaseWriteLock(&tdc->lock);
1923 if (setLocks && !slowPass)
1924 ReleaseReadLock(&avc->lock);
1926 ReleaseWriteLock(&avc->lock);
1930 * Call osi_FlushPages in open, read/write, and map, since it
1931 * is too hard here to figure out if we should lock the
1934 if (setLocks && !slowPass)
1935 ObtainReadLock(&avc->lock);
1937 ObtainWriteLock(&avc->lock, 66);
1938 ObtainWriteLock(&tdc->lock, 610);
1943 * avc->lock(R) if setLocks && !slowPass
1944 * avc->lock(W) if !setLocks || slowPass
1948 /* Watch for standard race condition around osi_FlushText */
1949 if (hsame(avc->m.DataVersion, tdc->f.versionNo)) {
1950 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
1951 afs_stats_cmperf.dcacheHits++;
1952 ConvertWToSLock(&tdc->lock);
1956 /* Sleep here when cache needs to be drained. */
1957 if (setLocks && !slowPass &&
1958 (afs_blocksUsed > (CM_WAITFORDRAINPCT*afs_cacheBlocks)/100)) {
1959 /* Make sure truncate daemon is running */
1960 afs_MaybeWakeupTruncateDaemon();
1961 ObtainWriteLock(&tdc->tlock, 614);
1962 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
1963 ReleaseWriteLock(&tdc->tlock);
1964 ReleaseWriteLock(&tdc->lock);
1965 ReleaseReadLock(&avc->lock);
1966 while ((afs_blocksUsed-afs_blocksDiscarded) >
1967 (CM_WAITFORDRAINPCT*afs_cacheBlocks)/100) {
1968 afs_WaitForCacheDrain = 1;
1969 afs_osi_Sleep(&afs_WaitForCacheDrain);
1971 afs_MaybeFreeDiscardedDCache();
1972 /* need to check if someone else got the chunk first. */
1973 goto RetryGetDCache;
1976 /* Do not fetch data beyond truncPos. */
1977 maxGoodLength = avc->m.Length;
1978 if (avc->truncPos < maxGoodLength) maxGoodLength = avc->truncPos;
1979 Position = AFS_CHUNKBASE(abyte);
1980 if (vType(avc) == VDIR) {
1981 size = avc->m.Length;
1982 if (size > tdc->f.chunkBytes) {
1983 /* pre-reserve space for file */
1984 afs_AdjustSize(tdc, size);
1986 size = 999999999; /* max size for transfer */
1989 size = AFS_CHUNKSIZE(abyte); /* expected max size */
1990 /* don't read past end of good data on server */
1991 if (Position + size > maxGoodLength)
1992 size = maxGoodLength - Position;
1993 if (size < 0) size = 0; /* Handle random races */
1994 if (size > tdc->f.chunkBytes) {
1995 /* pre-reserve space for file */
1996 afs_AdjustSize(tdc, size); /* changes chunkBytes */
1997 /* max size for transfer still in size */
2000 if (afs_mariner && !tdc->f.chunk)
2001 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter );*/
2003 * Right now, we only have one tool, and it's a hammer. So, we
2004 * fetch the whole file.
2006 DZap(&tdc->f.inode); /* pages in cache may be old */
2008 if (file = tdc->ihint) {
2009 if (tdc->f.inode == file->inum )
2016 file = osi_UFSOpen(tdc->f.inode);
2021 file = afs_CFileOpen(tdc->f.inode);
2022 afs_RemoveVCB(&avc->fid);
2023 tdc->f.states |= DWriting;
2024 tdc->dflags |= DFFetching;
2025 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2026 if (tdc->mflags & DFFetchReq) {
2027 tdc->mflags &= ~DFFetchReq;
2029 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2030 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE,
2031 ICL_TYPE_STRING, __FILE__,
2032 ICL_TYPE_INT32, __LINE__,
2033 ICL_TYPE_POINTER, tdc,
2034 ICL_TYPE_INT32, tdc->dflags);
2036 afs_osi_Wakeup(&tdc->validPos);
2039 tsmall = (struct tlocal1 *) osi_AllocLargeSpace(sizeof(struct tlocal1));
2040 setVcacheStatus = 0;
2043 * Remember if we are doing the reading from a replicated volume,
2044 * and how many times we've zipped around the fetch/analyze loop.
2046 fromReplica = (avc->states & CRO) ? 1 : 0;
2048 accP = &(afs_stats_cmfullperf.accessinf);
2050 (accP->replicatedRefs)++;
2052 (accP->unreplicatedRefs)++;
2053 #endif /* AFS_NOSTATS */
2054 /* this is a cache miss */
2055 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2056 ICL_TYPE_FID, &(avc->fid),
2057 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(Position),
2058 ICL_TYPE_INT32, size);
2060 if (size) afs_stats_cmperf.dcacheMisses++;
2063 * Dynamic root support: fetch data from local memory.
2065 if (afs_IsDynroot(avc)) {
2069 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2071 dynrootDir += Position;
2072 dynrootLen -= Position;
2073 if (size > dynrootLen)
2075 if (size < 0) size = 0;
2076 code = afs_CFileWrite(file, 0, dynrootDir, size);
2084 tdc->validPos = Position + size;
2085 afs_CFileTruncate(file, size); /* prune it */
2088 * Not a dynamic vnode: do the real fetch.
2093 * avc->lock(R) if setLocks && !slowPass
2094 * avc->lock(W) if !setLocks || slowPass
2098 tc = afs_Conn(&avc->fid, areq, SHARED_LOCK);
2100 afs_int32 length_hi, length, bytes;
2104 (accP->numReplicasAccessed)++;
2106 #endif /* AFS_NOSTATS */
2107 if (!setLocks || slowPass) {
2108 avc->callback = tc->srvr->server;
2110 newCallback = tc->srvr->server;
2115 tcall = rx_NewCall(tc->id);
2118 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
2119 #ifdef AFS_64BIT_CLIENT
2120 length_hi = code = 0;
2121 if (!afs_serverHasNo64Bit(tc)) {
2124 code = StartRXAFS_FetchData64(tcall,
2125 (struct AFSFid *) &avc->fid.Fid,
2129 afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
2130 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code);
2132 bytes = rx_Read(tcall, (char *)&length_hi, sizeof(afs_int32));
2134 if (bytes == sizeof(afs_int32)) {
2135 length_hi = ntohl(length_hi);
2138 code = rx_Error(tcall);
2140 code1 = rx_EndCall(tcall, code);
2142 tcall = (struct rx_call *) 0;
2146 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
2147 if (Position > 0x7FFFFFFF) {
2154 tcall = rx_NewCall(tc->id);
2155 code = StartRXAFS_FetchData(tcall,
2156 (struct AFSFid *) &avc->fid.Fid, pos, size);
2159 afs_serverSetNo64Bit(tc);
2163 bytes = rx_Read(tcall, (char *)&length, sizeof(afs_int32));
2165 if (bytes == sizeof(afs_int32)) {
2166 length = ntohl(length);
2168 code = rx_Error(tcall);
2171 FillInt64(lengthFound, length_hi, length);
2172 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64LENG,
2173 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
2174 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(lengthFound));
2175 #else /* AFS_64BIT_CLIENT */
2177 code = StartRXAFS_FetchData(tcall,
2178 (struct AFSFid *) &avc->fid.Fid,
2183 bytes = rx_Read(tcall, (char *)&length, sizeof(afs_int32));
2185 if (bytes == sizeof(afs_int32)) {
2186 length = ntohl(length);
2188 code = rx_Error(tcall);
2191 #endif /* AFS_64BIT_CLIENT */
2195 xferP = &(afs_stats_cmfullperf.rpc.fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
2196 osi_GetuTime(&xferStartTime);
2198 code = afs_CacheFetchProc(tcall, file,
2199 (afs_size_t) Position, tdc, avc,
2200 &bytesToXfer, &bytesXferred, length);
2202 osi_GetuTime(&xferStopTime);
2203 (xferP->numXfers)++;
2205 (xferP->numSuccesses)++;
2206 afs_stats_XferSumBytes[AFS_STATS_FS_XFERIDX_FETCHDATA] += bytesXferred;
2207 (xferP->sumBytes) += (afs_stats_XferSumBytes[AFS_STATS_FS_XFERIDX_FETCHDATA] >> 10);
2208 afs_stats_XferSumBytes[AFS_STATS_FS_XFERIDX_FETCHDATA] &= 0x3FF;
2209 if (bytesXferred < xferP->minBytes)
2210 xferP->minBytes = bytesXferred;
2211 if (bytesXferred > xferP->maxBytes)
2212 xferP->maxBytes = bytesXferred;
2215 * Tally the size of the object. Note: we tally the actual size,
2216 * NOT the number of bytes that made it out over the wire.
2218 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
2219 (xferP->count[0])++;
2221 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET1)
2222 (xferP->count[1])++;
2224 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET2)
2225 (xferP->count[2])++;
2227 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET3)
2228 (xferP->count[3])++;
2230 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET4)
2231 (xferP->count[4])++;
2233 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET5)
2234 (xferP->count[5])++;
2236 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET6)
2237 (xferP->count[6])++;
2239 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET7)
2240 (xferP->count[7])++;
2242 (xferP->count[8])++;
2244 afs_stats_GetDiff(elapsedTime, xferStartTime, xferStopTime);
2245 afs_stats_AddTo((xferP->sumTime), elapsedTime);
2246 afs_stats_SquareAddTo((xferP->sqrTime), elapsedTime);
2247 if (afs_stats_TimeLessThan(elapsedTime, (xferP->minTime))) {
2248 afs_stats_TimeAssign((xferP->minTime), elapsedTime);
2250 if (afs_stats_TimeGreaterThan(elapsedTime, (xferP->maxTime))) {
2251 afs_stats_TimeAssign((xferP->maxTime), elapsedTime);
2255 code = afs_CacheFetchProc(tcall, file, Position, tdc, avc, 0, 0, length);
2256 #endif /* AFS_NOSTATS */
2260 code = EndRXAFS_FetchData(tcall,
2269 code1 = rx_EndCall(tcall, code);
2275 if ( !code && code1 )
2279 /* callback could have been broken (or expired) in a race here,
2280 * but we return the data anyway. It's as good as we knew about
2281 * when we started. */
2283 * validPos is updated by CacheFetchProc, and can only be
2284 * modifed under a dcache write lock, which we've blocked out
2286 size = tdc->validPos - Position; /* actual segment size */
2287 if (size < 0) size = 0;
2288 afs_CFileTruncate(file, size); /* prune it */
2291 if (!setLocks || slowPass) {
2292 ObtainWriteLock(&afs_xcbhash, 453);
2293 afs_DequeueCallback(avc);
2294 avc->states &= ~(CStatd | CUnique);
2295 avc->callback = NULL;
2296 ReleaseWriteLock(&afs_xcbhash);
2297 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2298 osi_dnlc_purgedp(avc);
2300 /* Something lost. Forget about performance, and go
2301 * back with a vcache write lock.
2303 afs_CFileTruncate(file, 0);
2304 afs_AdjustSize(tdc, 0);
2305 afs_CFileClose(file);
2306 osi_FreeLargeSpace(tsmall);
2308 ReleaseWriteLock(&tdc->lock);
2311 ReleaseReadLock(&avc->lock);
2313 goto RetryGetDCache;
2318 (afs_Analyze(tc, code, &avc->fid, areq,
2319 AFS_STATS_FS_RPCIDX_FETCHDATA,
2320 SHARED_LOCK, NULL));
2324 * avc->lock(R) if setLocks && !slowPass
2325 * avc->lock(W) if !setLocks || slowPass
2331 * In the case of replicated access, jot down info on the number of
2332 * attempts it took before we got through or gave up.
2335 if (numFetchLoops <= 1)
2336 (accP->refFirstReplicaOK)++;
2337 if (numFetchLoops > accP->maxReplicasPerRef)
2338 accP->maxReplicasPerRef = numFetchLoops;
2340 #endif /* AFS_NOSTATS */
2342 tdc->dflags &= ~DFFetching;
2344 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2345 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE,
2346 ICL_TYPE_STRING, __FILE__,
2347 ICL_TYPE_INT32, __LINE__,
2348 ICL_TYPE_POINTER, tdc,
2349 ICL_TYPE_INT32, tdc->dflags);
2351 afs_osi_Wakeup(&tdc->validPos);
2353 if (avc->execsOrWriters == 0) tdc->f.states &= ~DWriting;
2355 /* now, if code != 0, we have an error and should punt.
2356 * note that we have the vcache write lock, either because
2357 * !setLocks or slowPass.
2360 afs_CFileTruncate(file, 0);
2361 afs_AdjustSize(tdc, 0);
2362 afs_CFileClose(file);
2363 ZapDCE(tdc); /* sets DFEntryMod */
2364 if (vType(avc) == VDIR) {
2365 DZap(&tdc->f.inode);
2367 ReleaseWriteLock(&tdc->lock);
2369 ObtainWriteLock(&afs_xcbhash, 454);
2370 afs_DequeueCallback(avc);
2371 avc->states &= ~( CStatd | CUnique );
2372 ReleaseWriteLock(&afs_xcbhash);
2373 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2374 osi_dnlc_purgedp(avc);
2377 * avc->lock(W); assert(!setLocks || slowPass)
2379 osi_Assert(!setLocks || slowPass);
2384 /* otherwise we copy in the just-fetched info */
2385 afs_CFileClose(file);
2386 afs_AdjustSize(tdc, size); /* new size */
2388 * Copy appropriate fields into vcache. Status is
2389 * copied later where we selectively acquire the
2390 * vcache write lock.
2393 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2395 setVcacheStatus = 1;
2396 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh, tsmall->OutStatus.DataVersion);
2397 tdc->dflags |= DFEntryMod;
2398 afs_indexFlags[tdc->index] |= IFEverUsed;
2399 ConvertWToSLock(&tdc->lock);
2400 } /*Data version numbers don't match*/
2403 * Data version numbers match.
2405 afs_stats_cmperf.dcacheHits++;
2406 } /*Data version numbers match*/
2408 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2412 * avc->lock(R) if setLocks && !slowPass
2413 * avc->lock(W) if !setLocks || slowPass
2414 * tdc->lock(S) if tdc
2418 * See if this was a reference to a file in the local cell.
2420 if (afs_IsPrimaryCellNum(avc->fid.Cell))
2421 afs_stats_cmperf.dlocalAccesses++;
2423 afs_stats_cmperf.dremoteAccesses++;
2425 /* Fix up LRU info */
2428 MObtainWriteLock(&afs_xdcache, 602);
2429 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2430 hadd32(afs_indexCounter, 1);
2431 MReleaseWriteLock(&afs_xdcache);
2433 /* return the data */
2434 if (vType(avc) == VDIR)
2437 *aoffset = AFS_CHUNKOFFSET(abyte);
2438 *alen = (tdc->f.chunkBytes - *aoffset);
2439 ReleaseSharedLock(&tdc->lock);
2444 * avc->lock(R) if setLocks && !slowPass
2445 * avc->lock(W) if !setLocks || slowPass
2448 /* Fix up the callback and status values in the vcache */
2450 if (setLocks && !slowPass) {
2453 * This is our dirty little secret to parallel fetches.
2454 * We don't write-lock the vcache while doing the fetch,
2455 * but potentially we'll need to update the vcache after
2456 * the fetch is done.
2458 * Drop the read lock and try to re-obtain the write
2459 * lock. If the vcache still has the same DV, it's
2460 * ok to go ahead and install the new data.
2462 afs_hyper_t currentDV, statusDV;
2464 hset(currentDV, avc->m.DataVersion);
2466 if (setNewCallback && avc->callback != newCallback)
2470 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2471 tsmall->OutStatus.DataVersion);
2473 if (setVcacheStatus && avc->m.Length != tsmall->OutStatus.Length)
2475 if (setVcacheStatus && !hsame(currentDV, statusDV))
2479 ReleaseReadLock(&avc->lock);
2481 if (doVcacheUpdate) {
2482 ObtainWriteLock(&avc->lock, 615);
2483 if (!hsame(avc->m.DataVersion, currentDV)) {
2484 /* We lose. Someone will beat us to it. */
2486 ReleaseWriteLock(&avc->lock);
2491 /* With slow pass, we've already done all the updates */
2493 ReleaseWriteLock(&avc->lock);
2496 /* Check if we need to perform any last-minute fixes with a write-lock */
2497 if (!setLocks || doVcacheUpdate) {
2498 if (setNewCallback) avc->callback = newCallback;
2499 if (tsmall && setVcacheStatus) afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2500 if (setLocks) ReleaseWriteLock(&avc->lock);
2503 if (tsmall) osi_FreeLargeSpace(tsmall);
2510 * afs_WriteThroughDSlots
2513 * Sweep through the dcache slots and write out any modified
2514 * in-memory data back on to our caching store.
2520 * The afs_xdcache is write-locked through this whole affair.
2522 void afs_WriteThroughDSlots(void)
2524 register struct dcache *tdc;
2525 register afs_int32 i, touchedit=0;
2526 struct dcache **ents;
2527 int entmax, entcount;
2529 AFS_STATCNT(afs_WriteThroughDSlots);
2532 * Because of lock ordering, we can't grab dcache locks while
2533 * holding afs_xdcache. So we enter xdcache, get a reference
2534 * for every dcache entry, and exit xdcache.
2536 MObtainWriteLock(&afs_xdcache,283);
2537 entmax = afs_cacheFiles;
2538 ents = afs_osi_Alloc(entmax * sizeof(struct dcache *));
2540 for(i = 0; i < afs_cacheFiles; i++) {
2541 tdc = afs_indexTable[i];
2543 /* Grab tlock in case the existing refcount isn't zero */
2544 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2545 ObtainWriteLock(&tdc->tlock, 623);
2547 ReleaseWriteLock(&tdc->tlock);
2549 ents[entcount++] = tdc;
2552 MReleaseWriteLock(&afs_xdcache);
2555 * Now, for each dcache entry we found, check if it's dirty.
2556 * If so, get write-lock, get afs_xdcache, which protects
2557 * afs_cacheInodep, and flush it. Don't forget to put back
2560 for (i = 0; i < entcount; i++) {
2563 if (tdc->dflags & DFEntryMod) {
2566 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2568 /* Now that we have the write lock, double-check */
2569 if (wrLock && (tdc->dflags & DFEntryMod)) {
2570 tdc->dflags &= ~DFEntryMod;
2571 MObtainWriteLock(&afs_xdcache, 620);
2572 afs_WriteDCache(tdc, 1);
2573 MReleaseWriteLock(&afs_xdcache);
2576 if (wrLock) ReleaseWriteLock(&tdc->lock);
2581 afs_osi_Free(ents, entmax * sizeof(struct dcache *));
2583 MObtainWriteLock(&afs_xdcache, 617);
2584 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2585 /* Touch the file to make sure that the mtime on the file is kept
2586 * up-to-date to avoid losing cached files on cold starts because
2587 * their mtime seems old...
2589 struct afs_fheader theader;
2591 theader.magic = AFS_FHMAGIC;
2592 theader.firstCSize = AFS_FIRSTCSIZE;
2593 theader.otherCSize = AFS_OTHERCSIZE;
2594 theader.version = AFS_CI_VERSION;
2595 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2597 MReleaseWriteLock(&afs_xdcache);
2604 * Return a pointer to an freshly initialized dcache entry using
2605 * a memory-based cache. The tlock will be read-locked.
2608 * aslot : Dcache slot to look at.
2609 * tmpdc : Ptr to dcache entry.
2612 * Must be called with afs_xdcache write-locked.
2615 struct dcache *afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2617 register afs_int32 code;
2618 register struct dcache *tdc;
2619 register char *tfile;
2622 AFS_STATCNT(afs_MemGetDSlot);
2623 if (CheckLock(&afs_xdcache) != -1) osi_Panic("getdslot nolock");
2624 if (aslot < 0 || aslot >= afs_cacheFiles) osi_Panic("getdslot slot");
2625 tdc = afs_indexTable[aslot];
2627 QRemove(&tdc->lruq); /* move to queue head */
2628 QAdd(&afs_DLRU, &tdc->lruq);
2629 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2630 ObtainWriteLock(&tdc->tlock, 624);
2632 ConvertWToRLock(&tdc->tlock);
2635 if (tmpdc == NULL) {
2636 if (!afs_freeDSList) afs_GetDownDSlot(4);
2637 if (!afs_freeDSList) {
2638 /* none free, making one is better than a panic */
2639 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2640 tdc = (struct dcache *) afs_osi_Alloc(sizeof (struct dcache));
2641 #ifdef KERNEL_HAVE_PIN
2642 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2645 tdc = afs_freeDSList;
2646 afs_freeDSList = (struct dcache *) tdc->lruq.next;
2649 tdc->dflags = 0; /* up-to-date, not in free q */
2651 QAdd(&afs_DLRU, &tdc->lruq);
2652 if (tdc->lruq.prev == &tdc->lruq) osi_Panic("lruq 3");
2659 /* initialize entry */
2660 tdc->f.fid.Cell = 0;
2661 tdc->f.fid.Fid.Volume = 0;
2663 hones(tdc->f.versionNo);
2664 tdc->f.inode = aslot;
2665 tdc->dflags |= DFEntryMod;
2668 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2671 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2672 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2673 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2676 RWLOCK_INIT(&tdc->lock, "dcache lock");
2677 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2678 RWLOCK_INIT(&tdc->mflock, "dcache flock");
2679 ObtainReadLock(&tdc->tlock);
2682 afs_indexTable[aslot] = tdc;
2685 } /*afs_MemGetDSlot*/
2687 unsigned int last_error = 0, lasterrtime = 0;
2693 * Return a pointer to an freshly initialized dcache entry using
2694 * a UFS-based disk cache. The dcache tlock will be read-locked.
2697 * aslot : Dcache slot to look at.
2698 * tmpdc : Ptr to dcache entry.
2701 * afs_xdcache lock write-locked.
2703 struct dcache *afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2705 register afs_int32 code;
2706 register struct dcache *tdc;
2710 AFS_STATCNT(afs_UFSGetDSlot);
2711 if (CheckLock(&afs_xdcache) != -1) osi_Panic("getdslot nolock");
2712 if (aslot < 0 || aslot >= afs_cacheFiles) osi_Panic("getdslot slot");
2713 tdc = afs_indexTable[aslot];
2715 QRemove(&tdc->lruq); /* move to queue head */
2716 QAdd(&afs_DLRU, &tdc->lruq);
2717 /* Grab tlock in case refCount != 0 */
2718 ObtainWriteLock(&tdc->tlock, 625);
2720 ConvertWToRLock(&tdc->tlock);
2723 /* otherwise we should read it in from the cache file */
2725 * If we weren't passed an in-memory region to place the file info,
2726 * we have to allocate one.
2728 if (tmpdc == NULL) {
2729 if (!afs_freeDSList) afs_GetDownDSlot(4);
2730 if (!afs_freeDSList) {
2731 /* none free, making one is better than a panic */
2732 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2733 tdc = (struct dcache *) afs_osi_Alloc(sizeof (struct dcache));
2734 #ifdef KERNEL_HAVE_PIN
2735 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2738 tdc = afs_freeDSList;
2739 afs_freeDSList = (struct dcache *) tdc->lruq.next;
2742 tdc->dflags = 0; /* up-to-date, not in free q */
2744 QAdd(&afs_DLRU, &tdc->lruq);
2745 if (tdc->lruq.prev == &tdc->lruq) osi_Panic("lruq 3");
2754 * Seek to the aslot'th entry and read it in.
2756 code = afs_osi_Read(afs_cacheInodep, sizeof(struct fcache) * aslot +
2757 sizeof(struct afs_fheader),
2758 (char *)(&tdc->f), sizeof(struct fcache));
2760 if (code != sizeof(struct fcache))
2762 if (!afs_CellNumValid(tdc->f.fid.Cell))
2766 tdc->f.fid.Cell = 0;
2767 tdc->f.fid.Fid.Volume = 0;
2769 hones(tdc->f.versionNo);
2770 tdc->dflags |= DFEntryMod;
2771 #if !defined(AFS_SUN5_ENV) && !defined(AFS_OSF_ENV) && !defined(AFS_SGI64_ENV) && !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN_ENV) && !defined(AFS_FBSD_ENV)
2772 last_error = getuerror();
2774 lasterrtime = osi_Time();
2775 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2781 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2782 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2783 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2786 RWLOCK_INIT(&tdc->lock, "dcache lock");
2787 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2788 RWLOCK_INIT(&tdc->mflock, "dcache flock");
2789 ObtainReadLock(&tdc->tlock);
2792 * If we didn't read into a temporary dcache region, update the
2793 * slot pointer table.
2796 afs_indexTable[aslot] = tdc;
2799 } /*afs_UFSGetDSlot*/
2807 * write a particular dcache entry back to its home in the
2811 * adc : Pointer to the dcache entry to write.
2812 * atime : If true, set the modtime on the file to the current time.
2815 * Must be called with the afs_xdcache lock at least read-locked,
2816 * and dcache entry at least read-locked.
2817 * The reference count is not changed.
2820 int afs_WriteDCache(register struct dcache *adc, int atime)
2822 register struct osi_file *tfile;
2823 register afs_int32 code;
2825 if (cacheDiskType == AFS_FCACHE_TYPE_MEM) return 0;
2826 AFS_STATCNT(afs_WriteDCache);
2828 adc->f.modTime = osi_Time();
2830 * Seek to the right dcache slot and write the in-memory image out to disk.
2832 afs_cellname_write();
2833 code = afs_osi_Write(afs_cacheInodep, sizeof(struct fcache) * adc->index +
2834 sizeof(struct afs_fheader),
2835 (char *)(&adc->f), sizeof(struct fcache));
2836 if (code != sizeof(struct fcache)) return EIO;
2846 * Wake up users of a particular file waiting for stores to take
2850 * avc : Ptr to related vcache entry.
2853 * Nothing interesting.
2856 int afs_wakeup(register struct vcache *avc)
2859 register struct brequest *tb;
2861 AFS_STATCNT(afs_wakeup);
2862 for (i = 0; i < NBRS; i++, tb++) {
2863 /* if request is valid and for this file, we've found it */
2864 if (tb->refCount > 0 && avc == tb->vnode) {
2867 * If CSafeStore is on, then we don't awaken the guy
2868 * waiting for the store until the whole store has finished.
2869 * Otherwise, we do it now. Note that if CSafeStore is on,
2870 * the BStore routine actually wakes up the user, instead
2872 * I think this is redundant now because this sort of thing
2873 * is already being handled by the higher-level code.
2875 if ((avc->states & CSafeStore) == 0) {
2877 tb->flags |= BUVALID;
2878 if (tb->flags & BUWAIT) {
2879 tb->flags &= ~BUWAIT;
2894 * Given a file name and inode, set up that file to be an
2895 * active member in the AFS cache. This also involves checking
2896 * the usability of its data.
2899 * afile : Name of the cache file to initialize.
2900 * ainode : Inode of the file.
2903 * This function is called only during initialization.
2906 int afs_InitCacheFile(char *afile, ino_t ainode)
2908 register afs_int32 code;
2909 #if defined(AFS_LINUX22_ENV)
2910 struct dentry *filevp;
2912 struct vnode *filevp;
2916 struct osi_file *tfile;
2917 struct osi_stat tstat;
2918 register struct dcache *tdc;
2920 AFS_STATCNT(afs_InitCacheFile);
2921 index = afs_stats_cmperf.cacheNumEntries;
2922 if (index >= afs_cacheFiles) return EINVAL;
2924 MObtainWriteLock(&afs_xdcache,282);
2925 tdc = afs_GetDSlot(index, NULL);
2926 ReleaseReadLock(&tdc->tlock);
2927 MReleaseWriteLock(&afs_xdcache);
2929 ObtainWriteLock(&tdc->lock, 621);
2930 MObtainWriteLock(&afs_xdcache, 622);
2932 code = gop_lookupname(afile,
2938 ReleaseWriteLock(&afs_xdcache);
2939 ReleaseWriteLock(&tdc->lock);
2944 * We have a VN_HOLD on filevp. Get the useful info out and
2945 * return. We make use of the fact that the cache is in the
2946 * UFS file system, and just record the inode number.
2948 #ifdef AFS_LINUX22_ENV
2949 tdc->f.inode = VTOI(filevp->d_inode)->i_number;
2952 tdc->f.inode = afs_vnodeToInumber(filevp);
2956 AFS_RELE((struct vnode *)filevp);
2958 #endif /* AFS_LINUX22_ENV */
2961 tdc->f.inode = ainode;
2964 if ((tdc->f.states & DWriting) ||
2965 tdc->f.fid.Fid.Volume == 0) fileIsBad = 1;
2966 tfile = osi_UFSOpen(tdc->f.inode);
2967 code = afs_osi_Stat(tfile, &tstat);
2968 if (code) osi_Panic("initcachefile stat");
2971 * If file size doesn't match the cache info file, it's probably bad.
2973 if (tdc->f.chunkBytes != tstat.size) fileIsBad = 1;
2974 tdc->f.chunkBytes = 0;
2977 * If file changed within T (120?) seconds of cache info file, it's
2978 * probably bad. In addition, if slot changed within last T seconds,
2979 * the cache info file may be incorrectly identified, and so slot
2982 if (cacheInfoModTime < tstat.mtime + 120) fileIsBad = 1;
2983 if (cacheInfoModTime < tdc->f.modTime + 120) fileIsBad = 1;
2984 /* In case write through is behind, make sure cache items entry is
2985 * at least as new as the chunk.
2987 if (tdc->f.modTime < tstat.mtime) fileIsBad = 1;
2989 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
2990 if (tstat.size != 0)
2991 osi_UFSTruncate(tfile, 0);
2992 /* put entry in free cache slot list */
2993 afs_dvnextTbl[tdc->index] = afs_freeDCList;
2994 afs_freeDCList = index;
2996 afs_indexFlags[index] |= IFFree;
2997 afs_indexUnique[index] = 0;
3001 * We must put this entry in the appropriate hash tables.
3002 * Note that i is still set from the above DCHash call
3004 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3005 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3006 afs_dchashTbl[code] = tdc->index;
3007 code = DVHash(&tdc->f.fid);
3008 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3009 afs_dvhashTbl[code] = tdc->index;
3010 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3012 /* has nontrivial amt of data */
3013 afs_indexFlags[index] |= IFEverUsed;
3014 afs_stats_cmperf.cacheFilesReused++;
3016 * Initialize index times to file's mod times; init indexCounter
3019 hset32(afs_indexTimes[index], tstat.atime);
3020 if (hgetlo(afs_indexCounter) < tstat.atime) {
3021 hset32(afs_indexCounter, tstat.atime);
3023 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3024 } /*File is not bad*/
3026 osi_UFSClose(tfile);
3027 tdc->f.states &= ~DWriting;
3028 tdc->dflags &= ~DFEntryMod;
3029 /* don't set f.modTime; we're just cleaning up */
3030 afs_WriteDCache(tdc, 0);
3031 ReleaseWriteLock(&afs_xdcache);
3032 ReleaseWriteLock(&tdc->lock);
3034 afs_stats_cmperf.cacheNumEntries++;
3039 /*Max # of struct dcache's resident at any time*/
3041 * If 'dchint' is enabled then in-memory dcache min is increased because of
3050 * Initialize dcache related variables.
3052 void afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk,
3055 register struct dcache *tdp;
3059 afs_freeDCList = NULLIDX;
3060 afs_discardDCList = NULLIDX;
3061 afs_freeDCCount = 0;
3062 afs_freeDSList = NULL;
3063 hzero(afs_indexCounter);
3065 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3071 if (achunk < 0 || achunk > 30)
3072 achunk = 13; /* Use default */
3073 AFS_SETCHUNKSIZE(achunk);
3079 if(aflags & AFSCALL_INIT_MEMCACHE) {
3081 * Use a memory cache instead of a disk cache
3083 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3084 afs_cacheType = &afs_MemCacheOps;
3085 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3086 ablocks = afiles * (AFS_FIRSTCSIZE/1024);
3087 /* ablocks is reported in 1K blocks */
3088 code = afs_InitMemCache(afiles * AFS_FIRSTCSIZE, AFS_FIRSTCSIZE, aflags);
3090 printf("afsd: memory cache too large for available memory.\n");
3091 printf("afsd: AFS files cannot be accessed.\n\n");
3093 afiles = ablocks = 0;
3096 printf("Memory cache: Allocating %d dcache entries...", aDentries);
3098 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3099 afs_cacheType = &afs_UfsCacheOps;
3102 if (aDentries > 512)
3103 afs_dhashsize = 2048;
3104 /* initialize hash tables */
3105 afs_dvhashTbl = (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3106 afs_dchashTbl = (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3107 for(i=0;i< afs_dhashsize;i++) {
3108 afs_dvhashTbl[i] = NULLIDX;
3109 afs_dchashTbl[i] = NULLIDX;
3111 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3112 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3113 for(i=0;i< afiles;i++) {
3114 afs_dvnextTbl[i] = NULLIDX;
3115 afs_dcnextTbl[i] = NULLIDX;
3118 /* Allocate and zero the pointer array to the dcache entries */
3119 afs_indexTable = (struct dcache **)
3120 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3121 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3122 afs_indexTimes = (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3123 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3124 afs_indexUnique = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3125 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3126 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3127 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3129 /* Allocate and thread the struct dcache entries themselves */
3130 tdp = afs_Initial_freeDSList =
3131 (struct dcache *) afs_osi_Alloc(aDentries * sizeof(struct dcache));
3132 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3133 #ifdef KERNEL_HAVE_PIN
3134 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles);/* XXX */
3135 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3136 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3137 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3138 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3139 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3140 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3141 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3142 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3145 afs_freeDSList = &tdp[0];
3146 for(i=0; i < aDentries-1; i++) {
3147 tdp[i].lruq.next = (struct afs_q *) (&tdp[i+1]);
3149 tdp[aDentries-1].lruq.next = (struct afs_q *) 0;
3151 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal = afs_cacheBlocks = ablocks;
3152 afs_ComputeCacheParms(); /* compute parms based on cache size */
3154 afs_dcentries = aDentries;
3163 void shutdown_dcache(void)
3167 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3168 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3169 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3170 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3171 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3172 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3173 afs_osi_Free(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3174 #ifdef KERNEL_HAVE_PIN
3175 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3176 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3177 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3178 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3179 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3180 unpin((u_char *)afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3181 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3185 for(i=0;i< afs_dhashsize;i++) {
3186 afs_dvhashTbl[i] = NULLIDX;
3187 afs_dchashTbl[i] = NULLIDX;
3191 afs_blocksUsed = afs_dcentries = 0;
3192 hzero(afs_indexCounter);
3194 afs_freeDCCount = 0;
3195 afs_freeDCList = NULLIDX;
3196 afs_discardDCList = NULLIDX;
3197 afs_freeDSList = afs_Initial_freeDSList = 0;
3199 LOCK_INIT(&afs_xdcache, "afs_xdcache");