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"
17 #include "afs/sysincludes.h" /*Standard vendor system headers */
18 #include "afsincludes.h" /*AFS-based standard headers */
19 #include "afs/afs_stats.h" /* statistics */
20 #include "afs/afs_cbqueue.h"
21 #include "afs/afs_osidnlc.h"
23 /* Forward declarations. */
24 static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
25 static void afs_FreeDiscardedDCache(void);
26 static void afs_DiscardDCache(struct dcache *);
27 static void afs_FreeDCache(struct dcache *);
29 static afs_int32 afs_DCGetBucket(struct vcache *);
30 static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
31 static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
32 static void afs_DCSizeInit(void);
33 static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);
36 * --------------------- Exported definitions ---------------------
39 afs_int32 afs_blocksUsed_0; /*!< 1K blocks in cache - in theory is zero */
40 afs_int32 afs_blocksUsed_1; /*!< 1K blocks in cache */
41 afs_int32 afs_blocksUsed_2; /*!< 1K blocks in cache */
42 afs_int32 afs_pct1 = -1;
43 afs_int32 afs_pct2 = -1;
44 afs_uint32 afs_tpct1 = 0;
45 afs_uint32 afs_tpct2 = 0;
46 afs_uint32 splitdcache = 0;
48 afs_lock_t afs_xdcache; /*!< Lock: alloc new disk cache entries */
49 afs_int32 afs_freeDCList; /*!< Free list for disk cache entries */
50 afs_int32 afs_freeDCCount; /*!< Count of elts in freeDCList */
51 afs_int32 afs_discardDCList; /*!< Discarded disk cache entries */
52 afs_int32 afs_discardDCCount; /*!< Count of elts in discardDCList */
53 struct dcache *afs_freeDSList; /*!< Free list for disk slots */
54 struct dcache *afs_Initial_freeDSList; /*!< Initial list for above */
55 afs_dcache_id_t cacheInode; /*!< Inode for CacheItems file */
56 struct osi_file *afs_cacheInodep = 0; /*!< file for CacheItems inode */
57 struct afs_q afs_DLRU; /*!< dcache LRU */
58 afs_int32 afs_dhashsize = 1024;
59 afs_int32 *afs_dvhashTbl; /*!< Data cache hash table: hashed by FID + chunk number. */
60 afs_int32 *afs_dchashTbl; /*!< Data cache hash table: hashed by FID. */
61 afs_int32 *afs_dvnextTbl; /*!< Dcache hash table links */
62 afs_int32 *afs_dcnextTbl; /*!< Dcache hash table links */
63 struct dcache **afs_indexTable; /*!< Pointers to dcache entries */
64 afs_hyper_t *afs_indexTimes; /*!< Dcache entry Access times */
65 afs_int32 *afs_indexUnique; /*!< dcache entry Fid.Unique */
66 unsigned char *afs_indexFlags; /*!< (only one) Is there data there? */
67 afs_hyper_t afs_indexCounter; /*!< Fake time for marking index
69 afs_int32 afs_cacheFiles = 0; /*!< Size of afs_indexTable */
70 afs_int32 afs_cacheBlocks; /*!< 1K blocks in cache */
71 afs_int32 afs_cacheStats; /*!< Stat entries in cache */
72 afs_int32 afs_blocksUsed; /*!< Number of blocks in use */
73 afs_int32 afs_blocksDiscarded; /*!<Blocks freed but not truncated */
74 afs_int32 afs_fsfragsize = AFS_MIN_FRAGSIZE; /*!< Underlying Filesystem minimum unit
75 *of disk allocation usually 1K
76 *this value is (truefrag -1 ) to
77 *save a bunch of subtracts... */
78 #ifdef AFS_64BIT_CLIENT
79 #ifdef AFS_VM_RDWR_ENV
80 afs_size_t afs_vmMappingEnd; /* !< For large files (>= 2GB) the VM
81 * mapping an 32bit addressing machines
82 * can only be used below the 2 GB
83 * line. From this point upwards we
84 * must do direct I/O into the cache
85 * files. The value should be on a
87 #endif /* AFS_VM_RDWR_ENV */
88 #endif /* AFS_64BIT_CLIENT */
90 /* The following is used to ensure that new dcache's aren't obtained when
91 * the cache is nearly full.
93 int afs_WaitForCacheDrain = 0;
94 int afs_TruncateDaemonRunning = 0;
95 int afs_CacheTooFull = 0;
97 afs_int32 afs_dcentries; /*!< In-memory dcache entries */
100 int dcacheDisabled = 0;
102 struct afs_cacheOps afs_UfsCacheOps = {
103 #ifndef HAVE_STRUCT_LABEL_SUPPORT
116 .truncate = osi_UFSTruncate,
117 .fread = afs_osi_Read,
118 .fwrite = afs_osi_Write,
119 .close = osi_UFSClose,
120 .vreadUIO = afs_UFSReadUIO,
121 .vwriteUIO = afs_UFSWriteUIO,
122 .GetDSlot = afs_UFSGetDSlot,
123 .GetVolSlot = afs_UFSGetVolSlot,
124 .HandleLink = afs_UFSHandleLink,
128 struct afs_cacheOps afs_MemCacheOps = {
129 #ifndef HAVE_STRUCT_LABEL_SUPPORT
131 afs_MemCacheTruncate,
141 .open = afs_MemCacheOpen,
142 .truncate = afs_MemCacheTruncate,
143 .fread = afs_MemReadBlk,
144 .fwrite = afs_MemWriteBlk,
145 .close = afs_MemCacheClose,
146 .vreadUIO = afs_MemReadUIO,
147 .vwriteUIO = afs_MemWriteUIO,
148 .GetDSlot = afs_MemGetDSlot,
149 .GetVolSlot = afs_MemGetVolSlot,
150 .HandleLink = afs_MemHandleLink,
154 int cacheDiskType; /*Type of backing disk for cache */
155 struct afs_cacheOps *afs_cacheType;
158 * Where is this vcache's entry associated dcache located/
159 * \param avc The vcache entry.
160 * \return Bucket index:
165 afs_DCGetBucket(struct vcache *avc)
170 /* This should be replaced with some sort of user configurable function */
171 if (avc->f.states & CRO) {
173 } else if (avc->f.states & CBackup) {
183 * Readjust a dcache's size.
185 * \param adc The dcache to be adjusted.
186 * \param oldSize Old size for the dcache.
187 * \param newSize The new size to be adjusted to.
191 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
193 afs_int32 adjustSize = newSize - oldSize;
201 afs_blocksUsed_0 += adjustSize;
202 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
205 afs_blocksUsed_1 += adjustSize;
206 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
209 afs_blocksUsed_2 += adjustSize;
210 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
218 * Move a dcache from one bucket to another.
220 * \param adc Operate on this dcache.
221 * \param size Size in bucket (?).
222 * \param newBucket Destination bucket.
226 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
231 /* Substract size from old bucket. */
235 afs_blocksUsed_0 -= size;
238 afs_blocksUsed_1 -= size;
241 afs_blocksUsed_2 -= size;
245 /* Set new bucket and increase destination bucket size. */
246 adc->bucket = newBucket;
251 afs_blocksUsed_0 += size;
254 afs_blocksUsed_1 += size;
257 afs_blocksUsed_2 += size;
265 * Init split caches size.
270 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
279 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
284 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
285 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
287 /* Short cut: if we don't know about it, try to kill it */
288 if (phase < 2 && afs_blocksUsed_0)
291 if (afs_pct1 > afs_tpct1)
293 if (afs_pct2 > afs_tpct2)
295 return 0; /* unlikely */
300 * Warn about failing to store a file.
302 * \param acode Associated error code.
303 * \param avolume Volume involved.
304 * \param aflags How to handle the output:
305 * aflags & 1: Print out on console
306 * aflags & 2: Print out on controlling tty
308 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
312 afs_StoreWarn(afs_int32 acode, afs_int32 avolume,
315 static char problem_fmt[] =
316 "afs: failed to store file in volume %d (%s)\n";
317 static char problem_fmt_w_error[] =
318 "afs: failed to store file in volume %d (error %d)\n";
319 static char netproblems[] = "network problems";
320 static char partfull[] = "partition full";
321 static char overquota[] = "over quota";
323 AFS_STATCNT(afs_StoreWarn);
329 afs_warn(problem_fmt, avolume, netproblems);
331 afs_warnuser(problem_fmt, avolume, netproblems);
332 } else if (acode == ENOSPC) {
337 afs_warn(problem_fmt, avolume, partfull);
339 afs_warnuser(problem_fmt, avolume, partfull);
342 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
343 * Instead ENOSPC will be sent...
345 if (acode == EDQUOT) {
350 afs_warn(problem_fmt, avolume, overquota);
352 afs_warnuser(problem_fmt, avolume, overquota);
360 afs_warn(problem_fmt_w_error, avolume, acode);
362 afs_warnuser(problem_fmt_w_error, avolume, acode);
367 * Try waking up truncation daemon, if it's worth it.
370 afs_MaybeWakeupTruncateDaemon(void)
372 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
373 afs_CacheTooFull = 1;
374 if (!afs_TruncateDaemonRunning)
375 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
376 } else if (!afs_TruncateDaemonRunning
377 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
378 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
385 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
386 * struct so we need only export one symbol for AIX.
388 static struct CTD_stats {
389 osi_timeval_t CTD_beforeSleep;
390 osi_timeval_t CTD_afterSleep;
391 osi_timeval_t CTD_sleepTime;
392 osi_timeval_t CTD_runTime;
396 u_int afs_min_cache = 0;
399 * Keeps the cache clean and free by truncating uneeded files, when used.
404 afs_CacheTruncateDaemon(void)
406 osi_timeval_t CTD_tmpTime;
410 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
412 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
414 osi_GetuTime(&CTD_stats.CTD_afterSleep);
415 afs_TruncateDaemonRunning = 1;
417 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
418 ObtainWriteLock(&afs_xdcache, 266);
419 if (afs_CacheTooFull) {
420 int space_needed, slots_needed;
421 /* if we get woken up, we should try to clean something out */
422 for (counter = 0; counter < 10; counter++) {
424 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
425 if (space_needed < 0)
428 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
429 if (slots_needed < 0)
431 if (slots_needed || space_needed)
432 afs_GetDownD(slots_needed, &space_needed, 0);
433 if ((space_needed <= 0) && (slots_needed <= 0)) {
434 afs_CacheTooFull = 0;
437 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
440 if (!afs_CacheIsTooFull())
441 afs_CacheTooFull = 0;
442 } /* end of cache cleanup */
443 ReleaseWriteLock(&afs_xdcache);
446 * This is a defensive check to try to avoid starving threads
447 * that may need the global lock so thay can help free some
448 * cache space. If this thread won't be sleeping or truncating
449 * any cache files then give up the global lock so other
450 * threads get a chance to run.
452 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
453 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
454 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
458 * This is where we free the discarded cache elements.
460 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
461 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
462 afs_FreeDiscardedDCache();
465 /* See if we need to continue to run. Someone may have
466 * signalled us while we were executing.
468 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
469 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
470 /* Collect statistics on truncate daemon. */
471 CTD_stats.CTD_nSleeps++;
472 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
473 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
474 CTD_stats.CTD_beforeSleep);
475 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
477 afs_TruncateDaemonRunning = 0;
478 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
479 afs_TruncateDaemonRunning = 1;
481 osi_GetuTime(&CTD_stats.CTD_afterSleep);
482 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
483 CTD_stats.CTD_afterSleep);
484 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
486 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
487 afs_termState = AFSOP_STOP_AFSDB;
488 afs_osi_Wakeup(&afs_termState);
496 * Make adjustment for the new size in the disk cache entry
498 * \note Major Assumptions Here:
499 * Assumes that frag size is an integral power of two, less one,
500 * and that this is a two's complement machine. I don't
501 * know of any filesystems which violate this assumption...
503 * \param adc Ptr to dcache entry.
504 * \param anewsize New size desired.
509 afs_AdjustSize(struct dcache *adc, afs_int32 newSize)
513 AFS_STATCNT(afs_AdjustSize);
515 adc->dflags |= DFEntryMod;
516 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
517 adc->f.chunkBytes = newSize;
520 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
521 afs_DCAdjustSize(adc, oldSize, newSize);
522 if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {
524 /* We're growing the file, wakeup the daemon */
525 afs_MaybeWakeupTruncateDaemon();
527 afs_blocksUsed += (newSize - oldSize);
528 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
533 * This routine is responsible for moving at least one entry (but up
534 * to some number of them) from the LRU queue to the free queue.
536 * \param anumber Number of entries that should ideally be moved.
537 * \param aneedSpace How much space we need (1K blocks);
540 * The anumber parameter is just a hint; at least one entry MUST be
541 * moved, or we'll panic. We must be called with afs_xdcache
542 * write-locked. We should try to satisfy both anumber and aneedspace,
543 * whichever is more demanding - need to do several things:
544 * 1. only grab up to anumber victims if aneedSpace <= 0, not
545 * the whole set of MAXATONCE.
546 * 2. dynamically choose MAXATONCE to reflect severity of
547 * demand: something like (*aneedSpace >> (logChunk - 9))
549 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
550 * indicates that the cache is not properly configured/tuned or
551 * something. We should be able to automatically correct that problem.
554 #define MAXATONCE 16 /* max we can obtain at once */
556 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
560 struct VenusFid *afid;
565 afs_uint32 victims[MAXATONCE];
566 struct dcache *victimDCs[MAXATONCE];
567 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
568 afs_uint32 victimPtr; /* next free item in victim arrays */
569 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
570 afs_uint32 maxVictimPtr; /* where it is */
574 AFS_STATCNT(afs_GetDownD);
576 if (CheckLock(&afs_xdcache) != -1)
577 osi_Panic("getdownd nolock");
578 /* decrement anumber first for all dudes in free list */
579 /* SHOULD always decrement anumber first, even if aneedSpace >0,
580 * because we should try to free space even if anumber <=0 */
581 if (!aneedSpace || *aneedSpace <= 0) {
582 anumber -= afs_freeDCCount;
584 return; /* enough already free */
588 /* bounds check parameter */
589 if (anumber > MAXATONCE)
590 anumber = MAXATONCE; /* all we can do */
592 /* rewrite so phases include a better eligiblity for gc test*/
594 * The phase variable manages reclaims. Set to 0, the first pass,
595 * we don't reclaim active entries, or other than target bucket.
596 * Set to 1, we reclaim even active ones in target bucket.
597 * Set to 2, we reclaim any inactive one.
598 * Set to 3, we reclaim even active ones. On Solaris, we also reclaim
599 * entries whose corresponding vcache has a nonempty multiPage list, when
608 for (i = 0; i < afs_cacheFiles; i++)
609 /* turn off all flags */
610 afs_indexFlags[i] &= ~IFFlag;
612 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
613 /* find oldest entries for reclamation */
614 maxVictimPtr = victimPtr = 0;
615 hzero(maxVictimTime);
616 curbucket = afs_DCWhichBucket(phase, buckethint);
617 /* select victims from access time array */
618 for (i = 0; i < afs_cacheFiles; i++) {
619 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
620 /* skip if dirty or already free */
623 tdc = afs_indexTable[i];
624 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
626 /* Wrong bucket; can't use it! */
629 if (tdc && (tdc->refCount != 0)) {
630 /* Referenced; can't use it! */
633 hset(vtime, afs_indexTimes[i]);
635 /* if we've already looked at this one, skip it */
636 if (afs_indexFlags[i] & IFFlag)
639 if (victimPtr < MAXATONCE) {
640 /* if there's at least one free victim slot left */
641 victims[victimPtr] = i;
642 hset(victimTimes[victimPtr], vtime);
643 if (hcmp(vtime, maxVictimTime) > 0) {
644 hset(maxVictimTime, vtime);
645 maxVictimPtr = victimPtr;
648 } else if (hcmp(vtime, maxVictimTime) < 0) {
650 * We're older than youngest victim, so we replace at
653 /* find youngest (largest LRU) victim */
656 osi_Panic("getdownd local");
658 hset(victimTimes[j], vtime);
659 /* recompute maxVictimTime */
660 hset(maxVictimTime, vtime);
661 for (j = 0; j < victimPtr; j++)
662 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
663 hset(maxVictimTime, victimTimes[j]);
669 /* now really reclaim the victims */
670 j = 0; /* flag to track if we actually got any of the victims */
671 /* first, hold all the victims, since we're going to release the lock
672 * during the truncate operation.
674 for (i = 0; i < victimPtr; i++) {
675 tdc = afs_GetValidDSlot(victims[i]);
676 /* We got tdc->tlock(R) here */
677 if (tdc && tdc->refCount == 1)
682 ReleaseReadLock(&tdc->tlock);
687 for (i = 0; i < victimPtr; i++) {
688 /* q is first elt in dcache entry */
690 /* now, since we're dropping the afs_xdcache lock below, we
691 * have to verify, before proceeding, that there are no other
692 * references to this dcache entry, even now. Note that we
693 * compare with 1, since we bumped it above when we called
694 * afs_GetValidDSlot to preserve the entry's identity.
696 if (tdc && tdc->refCount == 1) {
697 unsigned char chunkFlags;
698 afs_size_t tchunkoffset = 0;
700 /* xdcache is lower than the xvcache lock */
701 ReleaseWriteLock(&afs_xdcache);
702 ObtainReadLock(&afs_xvcache);
703 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
704 ReleaseReadLock(&afs_xvcache);
705 ObtainWriteLock(&afs_xdcache, 527);
707 if (tdc->refCount > 1)
710 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
711 chunkFlags = afs_indexFlags[tdc->index];
712 if (((phase & 1) == 0) && osi_Active(tvc))
714 if (((phase & 1) == 1) && osi_Active(tvc)
715 && (tvc->f.states & CDCLock)
716 && (chunkFlags & IFAnyPages))
718 if (chunkFlags & IFDataMod)
720 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
721 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
722 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
723 ICL_HANDLE_OFFSET(tchunkoffset));
725 #if defined(AFS_SUN5_ENV)
727 * Now we try to invalidate pages. We do this only for
728 * Solaris. For other platforms, it's OK to recycle a
729 * dcache entry out from under a page, because the strategy
730 * function can call afs_GetDCache().
732 if (!skip && (chunkFlags & IFAnyPages)) {
735 ReleaseWriteLock(&afs_xdcache);
736 ObtainWriteLock(&tvc->vlock, 543);
737 if (!QEmpty(&tvc->multiPage)) {
738 if (phase < 3 || osi_VM_MultiPageConflict(tvc, tdc)) {
743 /* block locking pages */
744 tvc->vstates |= VPageCleaning;
745 /* block getting new pages */
747 ReleaseWriteLock(&tvc->vlock);
748 /* One last recheck */
749 ObtainWriteLock(&afs_xdcache, 333);
750 chunkFlags = afs_indexFlags[tdc->index];
751 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
752 || (osi_Active(tvc) && (tvc->f.states & CDCLock)
753 && (chunkFlags & IFAnyPages))) {
755 ReleaseWriteLock(&afs_xdcache);
758 ReleaseWriteLock(&afs_xdcache);
760 code = osi_VM_GetDownD(tvc, tdc);
762 ObtainWriteLock(&afs_xdcache, 269);
763 /* we actually removed all pages, clean and dirty */
765 afs_indexFlags[tdc->index] &=
766 ~(IFDirtyPages | IFAnyPages);
769 ReleaseWriteLock(&afs_xdcache);
771 ObtainWriteLock(&tvc->vlock, 544);
772 if (--tvc->activeV == 0
773 && (tvc->vstates & VRevokeWait)) {
774 tvc->vstates &= ~VRevokeWait;
775 afs_osi_Wakeup((char *)&tvc->vstates);
778 if (tvc->vstates & VPageCleaning) {
779 tvc->vstates &= ~VPageCleaning;
780 afs_osi_Wakeup((char *)&tvc->vstates);
783 ReleaseWriteLock(&tvc->vlock);
785 #endif /* AFS_SUN5_ENV */
787 ReleaseWriteLock(&afs_xdcache);
790 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
791 ObtainWriteLock(&afs_xdcache, 528);
792 if (afs_indexFlags[tdc->index] &
793 (IFDataMod | IFDirtyPages | IFAnyPages))
795 if (tdc->refCount > 1)
798 #if defined(AFS_SUN5_ENV)
800 /* no vnode, so IFDirtyPages is spurious (we don't
801 * sweep dcaches on vnode recycling, so we can have
802 * DIRTYPAGES set even when all pages are gone). Just
804 * Hold vcache lock to prevent vnode from being
805 * created while we're clearing IFDirtyPages.
807 afs_indexFlags[tdc->index] &=
808 ~(IFDirtyPages | IFAnyPages);
812 /* skip this guy and mark him as recently used */
813 afs_indexFlags[tdc->index] |= IFFlag;
814 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
815 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
816 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
817 ICL_HANDLE_OFFSET(tchunkoffset));
819 /* flush this dude from the data cache and reclaim;
820 * first, make sure no one will care that we damage
821 * it, by removing it from all hash tables. Then,
822 * melt it down for parts. Note that any concurrent
823 * (new possibility!) calls to GetDownD won't touch
824 * this guy because his reference count is > 0. */
825 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
826 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
827 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
828 ICL_HANDLE_OFFSET(tchunkoffset));
829 AFS_STATCNT(afs_gget);
830 afs_HashOutDCache(tdc, 1);
831 if (tdc->f.chunkBytes != 0) {
835 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
840 afs_DiscardDCache(tdc);
845 j = 1; /* we reclaimed at least one victim */
850 } /* end of for victims loop */
853 /* Phase is 0 and no one was found, so try phase 1 (ignore
854 * osi_Active flag) */
857 for (i = 0; i < afs_cacheFiles; i++)
858 /* turn off all flags */
859 afs_indexFlags[i] &= ~IFFlag;
862 /* found no one in phases 0-5, we're hosed */
866 } /* big while loop */
874 * Remove adc from any hash tables that would allow it to be located
875 * again by afs_FindDCache or afs_GetDCache.
877 * \param adc Pointer to dcache entry to remove from hash tables.
879 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
883 afs_HashOutDCache(struct dcache *adc, int zap)
887 AFS_STATCNT(afs_glink);
889 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
891 /* if this guy is in the hash table, pull him out */
892 if (adc->f.fid.Fid.Volume != 0) {
893 /* remove entry from first hash chains */
894 i = DCHash(&adc->f.fid, adc->f.chunk);
895 us = afs_dchashTbl[i];
896 if (us == adc->index) {
897 /* first dude in the list */
898 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
900 /* somewhere on the chain */
901 while (us != NULLIDX) {
902 if (afs_dcnextTbl[us] == adc->index) {
903 /* found item pointing at the one to delete */
904 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
907 us = afs_dcnextTbl[us];
910 osi_Panic("dcache hc");
912 /* remove entry from *other* hash chain */
913 i = DVHash(&adc->f.fid);
914 us = afs_dvhashTbl[i];
915 if (us == adc->index) {
916 /* first dude in the list */
917 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
919 /* somewhere on the chain */
920 while (us != NULLIDX) {
921 if (afs_dvnextTbl[us] == adc->index) {
922 /* found item pointing at the one to delete */
923 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
926 us = afs_dvnextTbl[us];
929 osi_Panic("dcache hv");
934 /* prevent entry from being found on a reboot (it is already out of
935 * the hash table, but after a crash, we just look at fid fields of
936 * stable (old) entries).
938 adc->f.fid.Fid.Volume = 0; /* invalid */
940 /* mark entry as modified */
941 adc->dflags |= DFEntryMod;
946 } /*afs_HashOutDCache */
949 * Flush the given dcache entry, pulling it from hash chains
950 * and truncating the associated cache file.
952 * \param adc Ptr to dcache entry to flush.
955 * This routine must be called with the afs_xdcache lock held
959 afs_FlushDCache(struct dcache *adc)
961 AFS_STATCNT(afs_FlushDCache);
963 * Bump the number of cache files flushed.
965 afs_stats_cmperf.cacheFlushes++;
967 /* remove from all hash tables */
968 afs_HashOutDCache(adc, 1);
970 /* Free its space; special case null operation, since truncate operation
971 * in UFS is slow even in this case, and this allows us to pre-truncate
972 * these files at more convenient times with fewer locks set
973 * (see afs_GetDownD).
975 if (adc->f.chunkBytes != 0) {
976 afs_DiscardDCache(adc);
977 afs_MaybeWakeupTruncateDaemon();
982 if (afs_WaitForCacheDrain) {
983 if (afs_blocksUsed <=
984 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
985 afs_WaitForCacheDrain = 0;
986 afs_osi_Wakeup(&afs_WaitForCacheDrain);
989 } /*afs_FlushDCache */
993 * Put a dcache entry on the free dcache entry list.
995 * \param adc dcache entry to free.
997 * \note Environment: called with afs_xdcache lock write-locked.
1000 afs_FreeDCache(struct dcache *adc)
1002 /* Thread on free list, update free list count and mark entry as
1003 * freed in its indexFlags element. Also, ensure DCache entry gets
1004 * written out (set DFEntryMod).
1007 afs_dvnextTbl[adc->index] = afs_freeDCList;
1008 afs_freeDCList = adc->index;
1010 afs_indexFlags[adc->index] |= IFFree;
1011 adc->dflags |= DFEntryMod;
1013 if (afs_WaitForCacheDrain) {
1014 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1015 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1016 afs_WaitForCacheDrain = 0;
1017 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1020 } /* afs_FreeDCache */
1023 * Discard the cache element by moving it to the discardDCList.
1024 * This puts the cache element into a quasi-freed state, where
1025 * the space may be reused, but the file has not been truncated.
1027 * \note Major Assumptions Here:
1028 * Assumes that frag size is an integral power of two, less one,
1029 * and that this is a two's complement machine. I don't
1030 * know of any filesystems which violate this assumption...
1032 * \param adr Ptr to dcache entry.
1034 * \note Environment:
1035 * Must be called with afs_xdcache write-locked.
1039 afs_DiscardDCache(struct dcache *adc)
1043 AFS_STATCNT(afs_DiscardDCache);
1045 osi_Assert(adc->refCount == 1);
1047 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1048 afs_blocksDiscarded += size;
1049 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1051 afs_dvnextTbl[adc->index] = afs_discardDCList;
1052 afs_discardDCList = adc->index;
1053 afs_discardDCCount++;
1055 adc->f.fid.Fid.Volume = 0;
1056 adc->dflags |= DFEntryMod;
1057 afs_indexFlags[adc->index] |= IFDiscarded;
1059 if (afs_WaitForCacheDrain) {
1060 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1061 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1062 afs_WaitForCacheDrain = 0;
1063 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1067 } /*afs_DiscardDCache */
1070 * Free the next element on the list of discarded cache elements.
1073 afs_FreeDiscardedDCache(void)
1076 struct osi_file *tfile;
1079 AFS_STATCNT(afs_FreeDiscardedDCache);
1081 ObtainWriteLock(&afs_xdcache, 510);
1082 if (!afs_blocksDiscarded) {
1083 ReleaseWriteLock(&afs_xdcache);
1088 * Get an entry from the list of discarded cache elements
1090 tdc = afs_GetUnusedDSlot(afs_discardDCList);
1092 osi_Assert(tdc->refCount == 1);
1093 ReleaseReadLock(&tdc->tlock);
1095 afs_discardDCList = afs_dvnextTbl[tdc->index];
1096 afs_dvnextTbl[tdc->index] = NULLIDX;
1097 afs_discardDCCount--;
1098 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1099 afs_blocksDiscarded -= size;
1100 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1101 /* We can lock because we just took it off the free list */
1102 ObtainWriteLock(&tdc->lock, 626);
1103 ReleaseWriteLock(&afs_xdcache);
1106 * Truncate the element to reclaim its space
1108 tfile = afs_CFileOpen(&tdc->f.inode);
1109 afs_CFileTruncate(tfile, 0);
1110 afs_CFileClose(tfile);
1111 afs_AdjustSize(tdc, 0);
1112 afs_DCMoveBucket(tdc, 0, 0);
1115 * Free the element we just truncated
1117 ObtainWriteLock(&afs_xdcache, 511);
1118 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1119 afs_FreeDCache(tdc);
1120 tdc->f.states &= ~(DRO|DBackup|DRW);
1121 ReleaseWriteLock(&tdc->lock);
1123 ReleaseWriteLock(&afs_xdcache);
1127 * Free as many entries from the list of discarded cache elements
1128 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1133 afs_MaybeFreeDiscardedDCache(void)
1136 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1138 while (afs_blocksDiscarded
1139 && (afs_blocksUsed >
1140 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1141 afs_FreeDiscardedDCache();
1147 * Try to free up a certain number of disk slots.
1149 * \param anumber Targeted number of disk slots to free up.
1151 * \note Environment:
1152 * Must be called with afs_xdcache write-locked.
1156 afs_GetDownDSlot(int anumber)
1158 struct afs_q *tq, *nq;
1163 AFS_STATCNT(afs_GetDownDSlot);
1164 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1165 osi_Panic("diskless getdowndslot");
1167 if (CheckLock(&afs_xdcache) != -1)
1168 osi_Panic("getdowndslot nolock");
1170 /* decrement anumber first for all dudes in free list */
1171 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1174 return; /* enough already free */
1176 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1178 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1179 nq = QPrev(tq); /* in case we remove it */
1180 if (tdc->refCount == 0) {
1181 if ((ix = tdc->index) == NULLIDX)
1182 osi_Panic("getdowndslot");
1183 /* pull the entry out of the lruq and put it on the free list */
1184 QRemove(&tdc->lruq);
1186 /* write-through if modified */
1187 if (tdc->dflags & DFEntryMod) {
1188 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1190 * ask proxy to do this for us - we don't have the stack space
1192 while (tdc->dflags & DFEntryMod) {
1195 s = SPLOCK(afs_sgibklock);
1196 if (afs_sgibklist == NULL) {
1197 /* if slot is free, grab it. */
1198 afs_sgibklist = tdc;
1199 SV_SIGNAL(&afs_sgibksync);
1201 /* wait for daemon to (start, then) finish. */
1202 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1206 tdc->dflags &= ~DFEntryMod;
1207 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
1211 /* finally put the entry in the free list */
1212 afs_indexTable[ix] = NULL;
1213 afs_indexFlags[ix] &= ~IFEverUsed;
1214 tdc->index = NULLIDX;
1215 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1216 afs_freeDSList = tdc;
1220 } /*afs_GetDownDSlot */
1227 * Increment the reference count on a disk cache entry,
1228 * which already has a non-zero refcount. In order to
1229 * increment the refcount of a zero-reference entry, you
1230 * have to hold afs_xdcache.
1233 * adc : Pointer to the dcache entry to increment.
1236 * Nothing interesting.
1239 afs_RefDCache(struct dcache *adc)
1241 ObtainWriteLock(&adc->tlock, 627);
1242 if (adc->refCount < 0)
1243 osi_Panic("RefDCache: negative refcount");
1245 ReleaseWriteLock(&adc->tlock);
1254 * Decrement the reference count on a disk cache entry.
1257 * ad : Ptr to the dcache entry to decrement.
1260 * Nothing interesting.
1263 afs_PutDCache(struct dcache *adc)
1265 AFS_STATCNT(afs_PutDCache);
1266 ObtainWriteLock(&adc->tlock, 276);
1267 if (adc->refCount <= 0)
1268 osi_Panic("putdcache");
1270 ReleaseWriteLock(&adc->tlock);
1279 * Try to discard all data associated with this file from the
1283 * avc : Pointer to the cache info for the file.
1286 * Both pvnLock and lock are write held.
1289 afs_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
1294 AFS_STATCNT(afs_TryToSmush);
1295 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1296 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
1297 sync = 1; /* XX Temp testing XX */
1299 #if defined(AFS_SUN5_ENV)
1300 ObtainWriteLock(&avc->vlock, 573);
1301 avc->activeV++; /* block new getpages */
1302 ReleaseWriteLock(&avc->vlock);
1305 /* Flush VM pages */
1306 osi_VM_TryToSmush(avc, acred, sync);
1309 * Get the hash chain containing all dce's for this fid
1311 i = DVHash(&avc->f.fid);
1312 ObtainWriteLock(&afs_xdcache, 277);
1313 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1314 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1315 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1316 int releaseTlock = 1;
1317 tdc = afs_GetValidDSlot(index);
1319 /* afs_TryToSmush is best-effort; we may not actually discard
1320 * everything, so failure to discard a dcache due to an i/o
1324 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1326 if ((afs_indexFlags[index] & IFDataMod) == 0
1327 && tdc->refCount == 1) {
1328 ReleaseReadLock(&tdc->tlock);
1330 afs_FlushDCache(tdc);
1333 afs_indexTable[index] = 0;
1336 ReleaseReadLock(&tdc->tlock);
1340 #if defined(AFS_SUN5_ENV)
1341 ObtainWriteLock(&avc->vlock, 545);
1342 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1343 avc->vstates &= ~VRevokeWait;
1344 afs_osi_Wakeup((char *)&avc->vstates);
1346 ReleaseWriteLock(&avc->vlock);
1348 ReleaseWriteLock(&afs_xdcache);
1350 * It's treated like a callback so that when we do lookups we'll
1351 * invalidate the unique bit if any
1352 * trytoSmush occured during the lookup call
1358 * afs_DCacheMissingChunks
1361 * Given the cached info for a file, return the number of chunks that
1362 * are not available from the dcache.
1365 * avc: Pointer to the (held) vcache entry to look in.
1368 * The number of chunks which are not currently cached.
1371 * The vcache entry is held upon entry.
1375 afs_DCacheMissingChunks(struct vcache *avc)
1378 afs_size_t totalLength = 0;
1379 afs_uint32 totalChunks = 0;
1382 totalLength = avc->f.m.Length;
1383 if (avc->f.truncPos < totalLength)
1384 totalLength = avc->f.truncPos;
1386 /* Length is 0, no chunk missing. */
1387 if (totalLength == 0)
1390 /* If totalLength is a multiple of chunksize, the last byte appears
1391 * as being part of the next chunk, which does not exist.
1392 * Decrementing totalLength by one fixes that.
1395 totalChunks = (AFS_CHUNK(totalLength) + 1);
1397 /* If we're a directory, we only ever have one chunk, regardless of
1398 * the size of the dir.
1400 if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
1404 printf("Should have %d chunks for %u bytes\n",
1405 totalChunks, (totalLength + 1));
1407 i = DVHash(&avc->f.fid);
1408 ObtainWriteLock(&afs_xdcache, 1001);
1409 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1410 i = afs_dvnextTbl[index];
1411 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1412 tdc = afs_GetValidDSlot(index);
1414 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1417 ReleaseReadLock(&tdc->tlock);
1422 ReleaseWriteLock(&afs_xdcache);
1424 /*printf("Missing %d chunks\n", totalChunks);*/
1426 return (totalChunks);
1433 * Given the cached info for a file and a byte offset into the
1434 * file, make sure the dcache entry for that file and containing
1435 * the given byte is available, returning it to our caller.
1438 * avc : Pointer to the (held) vcache entry to look in.
1439 * abyte : Which byte we want to get to.
1442 * Pointer to the dcache entry covering the file & desired byte,
1443 * or NULL if not found.
1446 * The vcache entry is held upon entry.
1450 afs_FindDCache(struct vcache *avc, afs_size_t abyte)
1454 struct dcache *tdc = NULL;
1456 AFS_STATCNT(afs_FindDCache);
1457 chunk = AFS_CHUNK(abyte);
1460 * Hash on the [fid, chunk] and get the corresponding dcache index
1461 * after write-locking the dcache.
1463 i = DCHash(&avc->f.fid, chunk);
1464 ObtainWriteLock(&afs_xdcache, 278);
1465 for (index = afs_dchashTbl[i]; index != NULLIDX; index = afs_dcnextTbl[index]) {
1466 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1467 tdc = afs_GetValidDSlot(index);
1469 /* afs_FindDCache is best-effort; we may not find the given
1470 * file/offset, so if we cannot find the given dcache due to
1471 * i/o errors, that is okay. */
1474 ReleaseReadLock(&tdc->tlock);
1475 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1476 break; /* leaving refCount high for caller */
1481 if (index != NULLIDX) {
1482 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1483 hadd32(afs_indexCounter, 1);
1484 ReleaseWriteLock(&afs_xdcache);
1487 ReleaseWriteLock(&afs_xdcache);
1489 } /*afs_FindDCache */
1493 * Get a fresh dcache from the free or discarded list.
1495 * \param avc Who's dcache is this going to be?
1496 * \param chunk The position where it will be placed in.
1497 * \param lock How are locks held.
1498 * \param ashFid If this dcache going to be used for a shadow dir,
1501 * \note Required locks:
1503 * - avc (R if (lock & 1) set and W otherwise)
1504 * \note It write locks the new dcache. The caller must unlock it.
1506 * \return The new dcache.
1509 afs_AllocDCache(struct vcache *avc, afs_int32 chunk, afs_int32 lock,
1510 struct VenusFid *ashFid)
1512 struct dcache *tdc = NULL;
1513 afs_uint32 size = 0;
1514 struct osi_file *file;
1516 if (afs_discardDCList == NULLIDX
1517 || ((lock & 2) && afs_freeDCList != NULLIDX)) {
1519 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1520 tdc = afs_GetUnusedDSlot(afs_freeDCList);
1522 osi_Assert(tdc->refCount == 1);
1523 ReleaseReadLock(&tdc->tlock);
1524 ObtainWriteLock(&tdc->lock, 604);
1525 afs_freeDCList = afs_dvnextTbl[tdc->index];
1528 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1529 tdc = afs_GetUnusedDSlot(afs_discardDCList);
1531 osi_Assert(tdc->refCount == 1);
1532 ReleaseReadLock(&tdc->tlock);
1533 ObtainWriteLock(&tdc->lock, 605);
1534 afs_discardDCList = afs_dvnextTbl[tdc->index];
1535 afs_discardDCCount--;
1537 ((tdc->f.chunkBytes +
1538 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1539 tdc->f.states &= ~(DRO|DBackup|DRW);
1540 afs_DCMoveBucket(tdc, size, 0);
1541 afs_blocksDiscarded -= size;
1542 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1544 /* Truncate the chunk so zeroes get filled properly */
1545 file = afs_CFileOpen(&tdc->f.inode);
1546 afs_CFileTruncate(file, 0);
1547 afs_CFileClose(file);
1548 afs_AdjustSize(tdc, 0);
1554 * avc->lock(R) if setLocks
1555 * avc->lock(W) if !setLocks
1561 * Fill in the newly-allocated dcache record.
1563 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1565 /* Use shadow fid if provided. */
1566 tdc->f.fid = *ashFid;
1568 /* Use normal vcache's fid otherwise. */
1569 tdc->f.fid = avc->f.fid;
1570 if (avc->f.states & CRO)
1571 tdc->f.states = DRO;
1572 else if (avc->f.states & CBackup)
1573 tdc->f.states = DBackup;
1575 tdc->f.states = DRW;
1576 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1577 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1579 hones(tdc->f.versionNo); /* invalid value */
1580 tdc->f.chunk = chunk;
1581 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1583 if (tdc->lruq.prev == &tdc->lruq)
1584 osi_Panic("lruq 1");
1593 * This function is called to obtain a reference to data stored in
1594 * the disk cache, locating a chunk of data containing the desired
1595 * byte and returning a reference to the disk cache entry, with its
1596 * reference count incremented.
1600 * avc : Ptr to a vcache entry (unlocked)
1601 * abyte : Byte position in the file desired
1602 * areq : Request structure identifying the requesting user.
1603 * aflags : Settings as follows:
1605 * 2 : Return after creating entry.
1606 * 4 : called from afs_vnop_write.c
1607 * *alen contains length of data to be written.
1609 * aoffset : Set to the offset within the chunk where the resident
1611 * alen : Set to the number of bytes of data after the desired
1612 * byte (including the byte itself) which can be read
1616 * The vcache entry pointed to by avc is unlocked upon entry.
1620 * Update the vnode-to-dcache hint if we can get the vnode lock
1621 * right away. Assumes dcache entry is at least read-locked.
1624 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1626 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1627 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1630 ReleaseWriteLock(&v->lock);
1634 /* avc - Write-locked unless aflags & 1 */
1636 afs_GetDCache(struct vcache *avc, afs_size_t abyte,
1637 struct vrequest *areq, afs_size_t * aoffset,
1638 afs_size_t * alen, int aflags)
1640 afs_int32 i, code, shortcut;
1641 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1642 afs_int32 adjustsize = 0;
1648 afs_size_t Position = 0;
1649 afs_int32 size, tlen; /* size of segment to transfer */
1650 struct afs_FetchOutput *tsmall = 0;
1652 struct osi_file *file;
1653 struct afs_conn *tc;
1655 struct server *newCallback = NULL;
1656 char setNewCallback;
1657 char setVcacheStatus;
1658 char doVcacheUpdate;
1660 int doAdjustSize = 0;
1661 int doReallyAdjustSize = 0;
1662 int overWriteWholeChunk = 0;
1663 struct rx_connection *rxconn;
1666 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1667 int fromReplica; /*Are we reading from a replica? */
1668 int numFetchLoops; /*# times around the fetch/analyze loop */
1669 #endif /* AFS_NOSTATS */
1671 AFS_STATCNT(afs_GetDCache);
1675 setLocks = aflags & 1;
1678 * Determine the chunk number and offset within the chunk corresponding
1679 * to the desired byte.
1681 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1684 chunk = AFS_CHUNK(abyte);
1687 /* come back to here if we waited for the cache to drain. */
1690 setNewCallback = setVcacheStatus = 0;
1694 ObtainWriteLock(&avc->lock, 616);
1696 ObtainReadLock(&avc->lock);
1701 * avc->lock(R) if setLocks && !slowPass
1702 * avc->lock(W) if !setLocks || slowPass
1707 /* check hints first! (might could use bcmp or some such...) */
1708 if ((tdc = avc->dchint)) {
1712 * The locking order between afs_xdcache and dcache lock matters.
1713 * The hint dcache entry could be anywhere, even on the free list.
1714 * Locking afs_xdcache ensures that noone is trying to pull dcache
1715 * entries from the free list, and thereby assuming them to be not
1716 * referenced and not locked.
1718 ObtainReadLock(&afs_xdcache);
1719 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1721 if (dcLocked && (tdc->index != NULLIDX)
1722 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1723 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1724 /* got the right one. It might not be the right version, and it
1725 * might be fetching, but it's the right dcache entry.
1727 /* All this code should be integrated better with what follows:
1728 * I can save a good bit more time under a write lock if I do..
1730 ObtainWriteLock(&tdc->tlock, 603);
1732 ReleaseWriteLock(&tdc->tlock);
1734 ReleaseReadLock(&afs_xdcache);
1737 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1738 && !(tdc->dflags & DFFetching)) {
1740 afs_stats_cmperf.dcacheHits++;
1741 ObtainWriteLock(&afs_xdcache, 559);
1742 QRemove(&tdc->lruq);
1743 QAdd(&afs_DLRU, &tdc->lruq);
1744 ReleaseWriteLock(&afs_xdcache);
1747 * avc->lock(R) if setLocks && !slowPass
1748 * avc->lock(W) if !setLocks || slowPass
1755 ReleaseSharedLock(&tdc->lock);
1756 ReleaseReadLock(&afs_xdcache);
1764 * avc->lock(R) if setLocks && !slowPass
1765 * avc->lock(W) if !setLocks || slowPass
1766 * tdc->lock(S) if tdc
1769 if (!tdc) { /* If the hint wasn't the right dcache entry */
1770 int dslot_error = 0;
1772 * Hash on the [fid, chunk] and get the corresponding dcache index
1773 * after write-locking the dcache.
1778 * avc->lock(R) if setLocks && !slowPass
1779 * avc->lock(W) if !setLocks || slowPass
1782 i = DCHash(&avc->f.fid, chunk);
1783 /* check to make sure our space is fine */
1784 afs_MaybeWakeupTruncateDaemon();
1786 ObtainWriteLock(&afs_xdcache, 280);
1788 for (index = afs_dchashTbl[i]; index != NULLIDX; us = index, index = afs_dcnextTbl[index]) {
1789 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1790 tdc = afs_GetValidDSlot(index);
1792 /* we got an i/o error when trying to get the given dslot,
1793 * but do not bail out just yet; it is possible the dcache
1794 * we're looking for is elsewhere, so it doesn't matter if
1795 * we can't load this one. */
1799 ReleaseReadLock(&tdc->tlock);
1802 * avc->lock(R) if setLocks && !slowPass
1803 * avc->lock(W) if !setLocks || slowPass
1806 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1807 /* Move it up in the beginning of the list */
1808 if (afs_dchashTbl[i] != index) {
1809 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1810 afs_dcnextTbl[index] = afs_dchashTbl[i];
1811 afs_dchashTbl[i] = index;
1813 ReleaseWriteLock(&afs_xdcache);
1814 ObtainSharedLock(&tdc->lock, 606);
1815 break; /* leaving refCount high for caller */
1823 * If we didn't find the entry, we'll create one.
1825 if (index == NULLIDX) {
1828 * avc->lock(R) if setLocks
1829 * avc->lock(W) if !setLocks
1832 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1833 avc, ICL_TYPE_INT32, chunk);
1836 /* We couldn't find the dcache we want, but we hit some i/o
1837 * errors when trying to find it, so we're not sure if the
1838 * dcache we want is in the cache or not. Error out, so we
1839 * don't try to possibly create 2 separate dcaches for the
1840 * same exact data. */
1841 ReleaseWriteLock(&afs_xdcache);
1845 /* Make sure there is a free dcache entry for us to use */
1846 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1849 avc->f.states |= CDCLock;
1850 /* just need slots */
1851 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1853 avc->f.states &= ~CDCLock;
1854 if (afs_discardDCList != NULLIDX
1855 || afs_freeDCList != NULLIDX)
1857 /* If we can't get space for 5 mins we give up and panic */
1858 if (++downDCount > 300) {
1859 osi_Panic("getdcache");
1861 ReleaseWriteLock(&afs_xdcache);
1864 * avc->lock(R) if setLocks
1865 * avc->lock(W) if !setLocks
1867 afs_osi_Wait(1000, 0, 0);
1872 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1875 * Now add to the two hash chains - note that i is still set
1876 * from the above DCHash call.
1878 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1879 afs_dchashTbl[i] = tdc->index;
1880 i = DVHash(&avc->f.fid);
1881 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1882 afs_dvhashTbl[i] = tdc->index;
1883 tdc->dflags = DFEntryMod;
1885 afs_MaybeWakeupTruncateDaemon();
1886 ReleaseWriteLock(&afs_xdcache);
1887 ConvertWToSLock(&tdc->lock);
1892 /* vcache->dcache hint failed */
1895 * avc->lock(R) if setLocks && !slowPass
1896 * avc->lock(W) if !setLocks || slowPass
1899 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1900 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1901 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1902 hgetlo(avc->f.m.DataVersion));
1904 * Here we have the entry in tdc, with its refCount incremented.
1905 * Note: we don't use the S-lock on avc; it costs concurrency when
1906 * storing a file back to the server.
1910 * Not a newly created file so we need to check the file's length and
1911 * compare data versions since someone could have changed the data or we're
1912 * reading a file written elsewhere. We only want to bypass doing no-op
1913 * read rpcs on newly created files (dv of 0) since only then we guarantee
1914 * that this chunk's data hasn't been filled by another client.
1916 size = AFS_CHUNKSIZE(abyte);
1917 if (aflags & 4) /* called from write */
1919 else /* called from read */
1920 tlen = tdc->validPos - abyte;
1921 Position = AFS_CHUNKTOBASE(chunk);
1922 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1923 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1924 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1925 ICL_HANDLE_OFFSET(Position));
1926 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
1928 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
1929 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
1930 overWriteWholeChunk = 1;
1931 if (doAdjustSize || overWriteWholeChunk) {
1932 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1934 #ifdef AFS_SGI64_ENV
1937 #else /* AFS_SGI64_ENV */
1940 #endif /* AFS_SGI64_ENV */
1941 #else /* AFS_SGI_ENV */
1944 #endif /* AFS_SGI_ENV */
1945 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
1946 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1947 #if defined(AFS_SUN5_ENV)
1948 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
1950 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
1952 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1953 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
1954 doReallyAdjustSize = 1;
1956 if (doReallyAdjustSize || overWriteWholeChunk) {
1957 /* no data in file to read at this position */
1958 UpgradeSToWLock(&tdc->lock, 607);
1959 file = afs_CFileOpen(&tdc->f.inode);
1960 afs_CFileTruncate(file, 0);
1961 afs_CFileClose(file);
1962 afs_AdjustSize(tdc, 0);
1963 hset(tdc->f.versionNo, avc->f.m.DataVersion);
1964 tdc->dflags |= DFEntryMod;
1966 ConvertWToSLock(&tdc->lock);
1971 * We must read in the whole chunk if the version number doesn't
1975 /* don't need data, just a unique dcache entry */
1976 ObtainWriteLock(&afs_xdcache, 608);
1977 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1978 hadd32(afs_indexCounter, 1);
1979 ReleaseWriteLock(&afs_xdcache);
1981 updateV2DC(setLocks, avc, tdc, 553);
1982 if (vType(avc) == VDIR)
1985 *aoffset = AFS_CHUNKOFFSET(abyte);
1986 if (tdc->validPos < abyte)
1987 *alen = (afs_size_t) 0;
1989 *alen = tdc->validPos - abyte;
1990 ReleaseSharedLock(&tdc->lock);
1993 ReleaseWriteLock(&avc->lock);
1995 ReleaseReadLock(&avc->lock);
1997 return tdc; /* check if we're done */
2002 * avc->lock(R) if setLocks && !slowPass
2003 * avc->lock(W) if !setLocks || slowPass
2006 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2008 setNewCallback = setVcacheStatus = 0;
2012 * avc->lock(R) if setLocks && !slowPass
2013 * avc->lock(W) if !setLocks || slowPass
2016 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2018 * Version number mismatch.
2021 * If we are disconnected, then we can't do much of anything
2022 * because the data doesn't match the file.
2024 if (AFS_IS_DISCONNECTED) {
2025 ReleaseSharedLock(&tdc->lock);
2028 ReleaseWriteLock(&avc->lock);
2030 ReleaseReadLock(&avc->lock);
2032 /* Flush the Dcache */
2037 UpgradeSToWLock(&tdc->lock, 609);
2040 * If data ever existed for this vnode, and this is a text object,
2041 * do some clearing. Now, you'd think you need only do the flush
2042 * when VTEXT is on, but VTEXT is turned off when the text object
2043 * is freed, while pages are left lying around in memory marked
2044 * with this vnode. If we would reactivate (create a new text
2045 * object from) this vnode, we could easily stumble upon some of
2046 * these old pages in pagein. So, we always flush these guys.
2047 * Sun has a wonderful lack of useful invariants in this system.
2049 * avc->flushDV is the data version # of the file at the last text
2050 * flush. Clearly, at least, we don't have to flush the file more
2051 * often than it changes
2053 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2055 * By here, the cache entry is always write-locked. We can
2056 * deadlock if we call osi_Flush with the cache entry locked...
2057 * Unlock the dcache too.
2059 ReleaseWriteLock(&tdc->lock);
2060 if (setLocks && !slowPass)
2061 ReleaseReadLock(&avc->lock);
2063 ReleaseWriteLock(&avc->lock);
2067 * Call osi_FlushPages in open, read/write, and map, since it
2068 * is too hard here to figure out if we should lock the
2071 if (setLocks && !slowPass)
2072 ObtainReadLock(&avc->lock);
2074 ObtainWriteLock(&avc->lock, 66);
2075 ObtainWriteLock(&tdc->lock, 610);
2080 * avc->lock(R) if setLocks && !slowPass
2081 * avc->lock(W) if !setLocks || slowPass
2085 /* Watch for standard race condition around osi_FlushText */
2086 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2087 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2088 afs_stats_cmperf.dcacheHits++;
2089 ConvertWToSLock(&tdc->lock);
2093 /* Sleep here when cache needs to be drained. */
2094 if (setLocks && !slowPass
2095 && (afs_blocksUsed >
2096 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2097 /* Make sure truncate daemon is running */
2098 afs_MaybeWakeupTruncateDaemon();
2099 ObtainWriteLock(&tdc->tlock, 614);
2100 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2101 ReleaseWriteLock(&tdc->tlock);
2102 ReleaseWriteLock(&tdc->lock);
2103 ReleaseReadLock(&avc->lock);
2104 while ((afs_blocksUsed - afs_blocksDiscarded) >
2105 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2106 afs_WaitForCacheDrain = 1;
2107 afs_osi_Sleep(&afs_WaitForCacheDrain);
2109 afs_MaybeFreeDiscardedDCache();
2110 /* need to check if someone else got the chunk first. */
2111 goto RetryGetDCache;
2114 Position = AFS_CHUNKBASE(abyte);
2115 if (vType(avc) == VDIR) {
2116 size = avc->f.m.Length;
2117 if (size > tdc->f.chunkBytes) {
2118 /* pre-reserve space for file */
2119 afs_AdjustSize(tdc, size);
2121 size = 999999999; /* max size for transfer */
2123 afs_size_t maxGoodLength;
2125 /* estimate how much data we're expecting back from the server,
2126 * and reserve space in the dcache entry for it */
2128 maxGoodLength = avc->f.m.Length;
2129 if (avc->f.truncPos < maxGoodLength)
2130 maxGoodLength = avc->f.truncPos;
2132 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2133 if (Position + size > maxGoodLength)
2134 size = maxGoodLength - Position;
2136 size = 0; /* Handle random races */
2137 if (size > tdc->f.chunkBytes) {
2138 /* pre-reserve estimated space for file */
2139 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2143 /* For the actual fetch, do not limit the request to the
2144 * length of the file. If this results in a read past EOF on
2145 * the server, the server will just reply with less data than
2146 * requested. If we limit ourselves to only requesting data up
2147 * to the avc file length, we open ourselves up to races if the
2148 * file is extended on the server at about the same time.
2150 * However, we must restrict ourselves to the avc->f.truncPos
2151 * length, since this represents an outstanding local
2152 * truncation of the file that will be committed to the
2153 * fileserver when we actually write the fileserver contents.
2154 * If we do not restrict the fetch length based on
2155 * avc->f.truncPos, a different truncate operation extending
2156 * the file length could cause the old data after
2157 * avc->f.truncPos to reappear, instead of extending the file
2158 * with NUL bytes. */
2159 size = AFS_CHUNKSIZE(abyte);
2160 if (Position + size > avc->f.truncPos) {
2161 size = avc->f.truncPos - Position;
2168 if (afs_mariner && !tdc->f.chunk)
2169 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2171 * Right now, we only have one tool, and it's a hammer. So, we
2172 * fetch the whole file.
2174 DZap(tdc); /* pages in cache may be old */
2175 file = afs_CFileOpen(&tdc->f.inode);
2176 afs_RemoveVCB(&avc->f.fid);
2177 tdc->f.states |= DWriting;
2178 tdc->dflags |= DFFetching;
2179 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2180 if (tdc->mflags & DFFetchReq) {
2181 tdc->mflags &= ~DFFetchReq;
2182 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2183 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2184 __FILE__, ICL_TYPE_INT32, __LINE__,
2185 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2188 tsmall = osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2189 setVcacheStatus = 0;
2192 * Remember if we are doing the reading from a replicated volume,
2193 * and how many times we've zipped around the fetch/analyze loop.
2195 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2197 accP = &(afs_stats_cmfullperf.accessinf);
2199 (accP->replicatedRefs)++;
2201 (accP->unreplicatedRefs)++;
2202 #endif /* AFS_NOSTATS */
2203 /* this is a cache miss */
2204 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2205 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2206 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2209 afs_stats_cmperf.dcacheMisses++;
2212 * Dynamic root support: fetch data from local memory.
2214 if (afs_IsDynroot(avc)) {
2218 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2220 dynrootDir += Position;
2221 dynrootLen -= Position;
2222 if (size > dynrootLen)
2226 code = afs_CFileWrite(file, 0, dynrootDir, size);
2234 tdc->validPos = Position + size;
2235 afs_CFileTruncate(file, size); /* prune it */
2236 } else if (afs_IsDynrootMount(avc)) {
2240 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2242 dynrootDir += Position;
2243 dynrootLen -= Position;
2244 if (size > dynrootLen)
2248 code = afs_CFileWrite(file, 0, dynrootDir, size);
2256 tdc->validPos = Position + size;
2257 afs_CFileTruncate(file, size); /* prune it */
2260 * Not a dynamic vnode: do the real fetch.
2265 * avc->lock(R) if setLocks && !slowPass
2266 * avc->lock(W) if !setLocks || slowPass
2270 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
2275 (accP->numReplicasAccessed)++;
2277 #endif /* AFS_NOSTATS */
2278 if (!setLocks || slowPass) {
2279 avc->callback = tc->parent->srvr->server;
2281 newCallback = tc->parent->srvr->server;
2285 code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
2291 /* callback could have been broken (or expired) in a race here,
2292 * but we return the data anyway. It's as good as we knew about
2293 * when we started. */
2295 * validPos is updated by CacheFetchProc, and can only be
2296 * modifed under a dcache write lock, which we've blocked out
2298 size = tdc->validPos - Position; /* actual segment size */
2301 afs_CFileTruncate(file, size); /* prune it */
2303 if (!setLocks || slowPass) {
2304 ObtainWriteLock(&afs_xcbhash, 453);
2305 afs_DequeueCallback(avc);
2306 avc->f.states &= ~(CStatd | CUnique);
2307 avc->callback = NULL;
2308 ReleaseWriteLock(&afs_xcbhash);
2309 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2310 osi_dnlc_purgedp(avc);
2312 /* Something lost. Forget about performance, and go
2313 * back with a vcache write lock.
2315 afs_CFileTruncate(file, 0);
2316 afs_AdjustSize(tdc, 0);
2317 afs_CFileClose(file);
2318 osi_FreeLargeSpace(tsmall);
2320 ReleaseWriteLock(&tdc->lock);
2323 ReleaseReadLock(&avc->lock);
2325 goto RetryGetDCache;
2329 } while (afs_Analyze
2330 (tc, rxconn, code, &avc->f.fid, areq,
2331 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2335 * avc->lock(R) if setLocks && !slowPass
2336 * avc->lock(W) if !setLocks || slowPass
2342 * In the case of replicated access, jot down info on the number of
2343 * attempts it took before we got through or gave up.
2346 if (numFetchLoops <= 1)
2347 (accP->refFirstReplicaOK)++;
2348 if (numFetchLoops > accP->maxReplicasPerRef)
2349 accP->maxReplicasPerRef = numFetchLoops;
2351 #endif /* AFS_NOSTATS */
2353 tdc->dflags &= ~DFFetching;
2354 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2355 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2356 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2357 tdc, ICL_TYPE_INT32, tdc->dflags);
2358 if (avc->execsOrWriters == 0)
2359 tdc->f.states &= ~DWriting;
2361 /* now, if code != 0, we have an error and should punt.
2362 * note that we have the vcache write lock, either because
2363 * !setLocks or slowPass.
2366 afs_CFileTruncate(file, 0);
2367 afs_AdjustSize(tdc, 0);
2368 afs_CFileClose(file);
2369 ZapDCE(tdc); /* sets DFEntryMod */
2370 if (vType(avc) == VDIR) {
2373 tdc->f.states &= ~(DRO|DBackup|DRW);
2374 afs_DCMoveBucket(tdc, 0, 0);
2375 ReleaseWriteLock(&tdc->lock);
2377 if (!afs_IsDynroot(avc)) {
2378 ObtainWriteLock(&afs_xcbhash, 454);
2379 afs_DequeueCallback(avc);
2380 avc->f.states &= ~(CStatd | CUnique);
2381 ReleaseWriteLock(&afs_xcbhash);
2382 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2383 osi_dnlc_purgedp(avc);
2386 * avc->lock(W); assert(!setLocks || slowPass)
2388 osi_Assert(!setLocks || slowPass);
2394 /* otherwise we copy in the just-fetched info */
2395 afs_CFileClose(file);
2396 afs_AdjustSize(tdc, size); /* new size */
2398 * Copy appropriate fields into vcache. Status is
2399 * copied later where we selectively acquire the
2400 * vcache write lock.
2403 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2405 setVcacheStatus = 1;
2406 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2407 tsmall->OutStatus.DataVersion);
2408 tdc->dflags |= DFEntryMod;
2409 afs_indexFlags[tdc->index] |= IFEverUsed;
2410 ConvertWToSLock(&tdc->lock);
2411 } /*Data version numbers don't match */
2414 * Data version numbers match.
2416 afs_stats_cmperf.dcacheHits++;
2417 } /*Data version numbers match */
2419 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2423 * avc->lock(R) if setLocks && !slowPass
2424 * avc->lock(W) if !setLocks || slowPass
2425 * tdc->lock(S) if tdc
2429 * See if this was a reference to a file in the local cell.
2431 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2432 afs_stats_cmperf.dlocalAccesses++;
2434 afs_stats_cmperf.dremoteAccesses++;
2436 /* Fix up LRU info */
2439 ObtainWriteLock(&afs_xdcache, 602);
2440 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2441 hadd32(afs_indexCounter, 1);
2442 ReleaseWriteLock(&afs_xdcache);
2444 /* return the data */
2445 if (vType(avc) == VDIR)
2448 *aoffset = AFS_CHUNKOFFSET(abyte);
2449 *alen = (tdc->f.chunkBytes - *aoffset);
2450 ReleaseSharedLock(&tdc->lock);
2455 * avc->lock(R) if setLocks && !slowPass
2456 * avc->lock(W) if !setLocks || slowPass
2459 /* Fix up the callback and status values in the vcache */
2461 if (setLocks && !slowPass) {
2464 * This is our dirty little secret to parallel fetches.
2465 * We don't write-lock the vcache while doing the fetch,
2466 * but potentially we'll need to update the vcache after
2467 * the fetch is done.
2469 * Drop the read lock and try to re-obtain the write
2470 * lock. If the vcache still has the same DV, it's
2471 * ok to go ahead and install the new data.
2473 afs_hyper_t currentDV, statusDV;
2475 hset(currentDV, avc->f.m.DataVersion);
2477 if (setNewCallback && avc->callback != newCallback)
2481 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2482 tsmall->OutStatus.DataVersion);
2484 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2486 if (setVcacheStatus && !hsame(currentDV, statusDV))
2490 ReleaseReadLock(&avc->lock);
2492 if (doVcacheUpdate) {
2493 ObtainWriteLock(&avc->lock, 615);
2494 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2495 /* We lose. Someone will beat us to it. */
2497 ReleaseWriteLock(&avc->lock);
2502 /* With slow pass, we've already done all the updates */
2504 ReleaseWriteLock(&avc->lock);
2507 /* Check if we need to perform any last-minute fixes with a write-lock */
2508 if (!setLocks || doVcacheUpdate) {
2510 avc->callback = newCallback;
2511 if (tsmall && setVcacheStatus)
2512 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2514 ReleaseWriteLock(&avc->lock);
2518 osi_FreeLargeSpace(tsmall);
2521 } /*afs_GetDCache */
2525 * afs_WriteThroughDSlots
2528 * Sweep through the dcache slots and write out any modified
2529 * in-memory data back on to our caching store.
2535 * The afs_xdcache is write-locked through this whole affair.
2538 afs_WriteThroughDSlots(void)
2541 afs_int32 i, touchedit = 0;
2543 struct afs_q DirtyQ, *tq;
2545 AFS_STATCNT(afs_WriteThroughDSlots);
2548 * Because of lock ordering, we can't grab dcache locks while
2549 * holding afs_xdcache. So we enter xdcache, get a reference
2550 * for every dcache entry, and exit xdcache.
2552 ObtainWriteLock(&afs_xdcache, 283);
2554 for (i = 0; i < afs_cacheFiles; i++) {
2555 tdc = afs_indexTable[i];
2557 /* Grab tlock in case the existing refcount isn't zero */
2558 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2559 ObtainWriteLock(&tdc->tlock, 623);
2561 ReleaseWriteLock(&tdc->tlock);
2563 QAdd(&DirtyQ, &tdc->dirty);
2566 ReleaseWriteLock(&afs_xdcache);
2569 * Now, for each dcache entry we found, check if it's dirty.
2570 * If so, get write-lock, get afs_xdcache, which protects
2571 * afs_cacheInodep, and flush it. Don't forget to put back
2575 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2577 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2579 if (tdc->dflags & DFEntryMod) {
2582 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2584 /* Now that we have the write lock, double-check */
2585 if (wrLock && (tdc->dflags & DFEntryMod)) {
2586 tdc->dflags &= ~DFEntryMod;
2587 ObtainWriteLock(&afs_xdcache, 620);
2588 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
2589 ReleaseWriteLock(&afs_xdcache);
2593 ReleaseWriteLock(&tdc->lock);
2599 ObtainWriteLock(&afs_xdcache, 617);
2600 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2601 /* Touch the file to make sure that the mtime on the file is kept
2602 * up-to-date to avoid losing cached files on cold starts because
2603 * their mtime seems old...
2605 struct afs_fheader theader;
2607 theader.magic = AFS_FHMAGIC;
2608 theader.firstCSize = AFS_FIRSTCSIZE;
2609 theader.otherCSize = AFS_OTHERCSIZE;
2610 theader.version = AFS_CI_VERSION;
2611 theader.dataSize = sizeof(struct fcache);
2612 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2614 ReleaseWriteLock(&afs_xdcache);
2621 * Return a pointer to an freshly initialized dcache entry using
2622 * a memory-based cache. The tlock will be read-locked.
2625 * aslot : Dcache slot to look at.
2626 * needvalid : Whether the specified slot should already exist
2629 * Must be called with afs_xdcache write-locked.
2633 afs_MemGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
2638 AFS_STATCNT(afs_MemGetDSlot);
2639 if (CheckLock(&afs_xdcache) != -1)
2640 osi_Panic("getdslot nolock");
2641 if (aslot < 0 || aslot >= afs_cacheFiles)
2642 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2643 tdc = afs_indexTable[aslot];
2645 QRemove(&tdc->lruq); /* move to queue head */
2646 QAdd(&afs_DLRU, &tdc->lruq);
2647 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2648 ObtainWriteLock(&tdc->tlock, 624);
2650 ConvertWToRLock(&tdc->tlock);
2654 /* if 'indexvalid' is true, the slot must already exist and be populated
2655 * somewhere. for memcache, the only place that dcache entries exist is
2656 * in memory, so if we did not find it above, something is very wrong. */
2657 osi_Assert(!indexvalid);
2659 if (!afs_freeDSList)
2660 afs_GetDownDSlot(4);
2661 if (!afs_freeDSList) {
2662 /* none free, making one is better than a panic */
2663 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2664 tdc = afs_osi_Alloc(sizeof(struct dcache));
2665 osi_Assert(tdc != NULL);
2666 #ifdef KERNEL_HAVE_PIN
2667 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2670 tdc = afs_freeDSList;
2671 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2674 tdc->dflags = 0; /* up-to-date, not in free q */
2676 QAdd(&afs_DLRU, &tdc->lruq);
2677 if (tdc->lruq.prev == &tdc->lruq)
2678 osi_Panic("lruq 3");
2680 /* initialize entry */
2681 tdc->f.fid.Cell = 0;
2682 tdc->f.fid.Fid.Volume = 0;
2684 hones(tdc->f.versionNo);
2685 tdc->f.inode.mem = aslot;
2686 tdc->dflags |= DFEntryMod;
2689 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2692 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2693 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2694 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2697 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2698 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2699 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2700 ObtainReadLock(&tdc->tlock);
2702 afs_indexTable[aslot] = tdc;
2705 } /*afs_MemGetDSlot */
2707 unsigned int last_error = 0, lasterrtime = 0;
2713 * Return a pointer to an freshly initialized dcache entry using
2714 * a UFS-based disk cache. The dcache tlock will be read-locked.
2717 * aslot : Dcache slot to look at.
2718 * indexvalid : 1 if we know the slot we're giving is valid, and thus
2719 * reading the dcache from the disk index should succeed. 0
2720 * if we are initializing a new dcache, and so reading from
2721 * the disk index may fail.
2722 * datavalid : 0 if we are loading a dcache entry from the free or
2723 * discard list, so we know the data in the given dcache is
2724 * not valid. 1 if we are loading a known used dcache, so the
2725 * data in the dcache must be valid.
2728 * afs_xdcache lock write-locked.
2731 afs_UFSGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
2739 AFS_STATCNT(afs_UFSGetDSlot);
2740 if (CheckLock(&afs_xdcache) != -1)
2741 osi_Panic("getdslot nolock");
2742 if (aslot < 0 || aslot >= afs_cacheFiles)
2743 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2744 tdc = afs_indexTable[aslot];
2746 QRemove(&tdc->lruq); /* move to queue head */
2747 QAdd(&afs_DLRU, &tdc->lruq);
2748 /* Grab tlock in case refCount != 0 */
2749 ObtainWriteLock(&tdc->tlock, 625);
2751 ConvertWToRLock(&tdc->tlock);
2755 /* otherwise we should read it in from the cache file */
2756 if (!afs_freeDSList)
2757 afs_GetDownDSlot(4);
2758 if (!afs_freeDSList) {
2759 /* none free, making one is better than a panic */
2760 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2761 tdc = afs_osi_Alloc(sizeof(struct dcache));
2762 osi_Assert(tdc != NULL);
2763 #ifdef KERNEL_HAVE_PIN
2764 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2767 tdc = afs_freeDSList;
2768 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2771 tdc->dflags = 0; /* up-to-date, not in free q */
2773 QAdd(&afs_DLRU, &tdc->lruq);
2774 if (tdc->lruq.prev == &tdc->lruq)
2775 osi_Panic("lruq 3");
2778 * Seek to the aslot'th entry and read it in.
2780 off = sizeof(struct fcache)*aslot + sizeof(struct afs_fheader);
2782 afs_osi_Read(afs_cacheInodep,
2783 off, (char *)(&tdc->f),
2784 sizeof(struct fcache));
2786 if (code != sizeof(struct fcache)) {
2788 #if defined(KERNEL_HAVE_UERROR)
2789 last_error = getuerror();
2793 lasterrtime = osi_Time();
2795 struct osi_stat tstat;
2796 if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
2799 afs_warn("afs: disk cache read error in CacheItems slot %d "
2800 "off %d/%d code %d/%d\n",
2802 off, (int)tstat.size,
2803 (int)code, (int)sizeof(struct fcache));
2804 /* put tdc back on the free dslot list */
2805 QRemove(&tdc->lruq);
2806 tdc->index = NULLIDX;
2807 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
2808 afs_freeDSList = tdc;
2812 if (!afs_CellNumValid(tdc->f.fid.Cell)) {
2815 osi_Panic("afs: needed valid dcache but index %d off %d has "
2816 "invalid cell num %d\n",
2817 (int)aslot, off, (int)tdc->f.fid.Cell);
2821 if (datavalid && tdc->f.fid.Fid.Volume == 0) {
2822 osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
2826 if (!entryok || !datavalid) {
2827 tdc->f.fid.Cell = 0;
2828 tdc->f.fid.Fid.Volume = 0;
2830 hones(tdc->f.versionNo);
2831 tdc->dflags |= DFEntryMod;
2832 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2833 tdc->f.states &= ~(DRO|DBackup|DRW);
2834 afs_DCMoveBucket(tdc, 0, 0);
2837 if (tdc->f.states & DRO) {
2838 afs_DCMoveBucket(tdc, 0, 2);
2839 } else if (tdc->f.states & DBackup) {
2840 afs_DCMoveBucket(tdc, 0, 1);
2842 afs_DCMoveBucket(tdc, 0, 1);
2848 if (tdc->f.chunk >= 0)
2849 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2854 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2855 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2856 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2859 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2860 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2861 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2862 ObtainReadLock(&tdc->tlock);
2865 * If we didn't read into a temporary dcache region, update the
2866 * slot pointer table.
2868 afs_indexTable[aslot] = tdc;
2871 } /*afs_UFSGetDSlot */
2876 * Write a particular dcache entry back to its home in the
2879 * \param adc Pointer to the dcache entry to write.
2880 * \param atime If true, set the modtime on the file to the current time.
2882 * \note Environment:
2883 * Must be called with the afs_xdcache lock at least read-locked,
2884 * and dcache entry at least read-locked.
2885 * The reference count is not changed.
2889 afs_WriteDCache(struct dcache *adc, int atime)
2893 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2895 AFS_STATCNT(afs_WriteDCache);
2896 osi_Assert(WriteLocked(&afs_xdcache));
2898 adc->f.modTime = osi_Time();
2900 if ((afs_indexFlags[adc->index] & (IFFree | IFDiscarded)) == 0 &&
2901 adc->f.fid.Fid.Volume == 0) {
2902 /* If a dcache slot is not on the free or discard list, it must be
2903 * in the hash table. Thus, the volume must be non-zero, since that
2904 * is how we determine whether or not to unhash the entry when kicking
2905 * it out of the cache. Do this check now, since otherwise this can
2906 * cause hash table corruption and a panic later on after we read the
2908 osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
2909 adc->index, (unsigned)afs_indexFlags[adc->index]);
2913 * Seek to the right dcache slot and write the in-memory image out to disk.
2915 afs_cellname_write();
2917 afs_osi_Write(afs_cacheInodep,
2918 sizeof(struct fcache) * adc->index +
2919 sizeof(struct afs_fheader), (char *)(&adc->f),
2920 sizeof(struct fcache));
2921 if (code != sizeof(struct fcache)) {
2922 afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
2923 (long)(sizeof(struct fcache) * adc->index + sizeof(struct afs_fheader)),
2924 (int)code, (int)sizeof(struct fcache));
2933 * Wake up users of a particular file waiting for stores to take
2936 * \param avc Ptr to related vcache entry.
2938 * \note Environment:
2939 * Nothing interesting.
2942 afs_wakeup(struct vcache *avc)
2945 struct brequest *tb;
2947 AFS_STATCNT(afs_wakeup);
2948 for (i = 0; i < NBRS; i++, tb++) {
2949 /* if request is valid and for this file, we've found it */
2950 if (tb->refCount > 0 && avc == tb->vc) {
2953 * If CSafeStore is on, then we don't awaken the guy
2954 * waiting for the store until the whole store has finished.
2955 * Otherwise, we do it now. Note that if CSafeStore is on,
2956 * the BStore routine actually wakes up the user, instead
2958 * I think this is redundant now because this sort of thing
2959 * is already being handled by the higher-level code.
2961 if ((avc->f.states & CSafeStore) == 0) {
2963 tb->flags |= BUVALID;
2964 if (tb->flags & BUWAIT) {
2965 tb->flags &= ~BUWAIT;
2976 * Given a file name and inode, set up that file to be an
2977 * active member in the AFS cache. This also involves checking
2978 * the usability of its data.
2980 * \param afile Name of the cache file to initialize.
2981 * \param ainode Inode of the file.
2983 * \note Environment:
2984 * This function is called only during initialization.
2987 afs_InitCacheFile(char *afile, ino_t ainode)
2992 struct osi_file *tfile;
2993 struct osi_stat tstat;
2996 AFS_STATCNT(afs_InitCacheFile);
2997 index = afs_stats_cmperf.cacheNumEntries;
2998 if (index >= afs_cacheFiles)
3001 ObtainWriteLock(&afs_xdcache, 282);
3002 tdc = afs_GetNewDSlot(index);
3003 ReleaseReadLock(&tdc->tlock);
3004 ReleaseWriteLock(&afs_xdcache);
3006 ObtainWriteLock(&tdc->lock, 621);
3007 ObtainWriteLock(&afs_xdcache, 622);
3008 if (!afile && !ainode) {
3013 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
3015 ReleaseWriteLock(&afs_xdcache);
3016 ReleaseWriteLock(&tdc->lock);
3021 /* Add any other 'complex' inode types here ... */
3022 #if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
3023 tdc->f.inode.ufs = ainode;
3025 osi_Panic("Can't init cache with inode numbers when complex inodes are "
3030 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3032 tfile = osi_UFSOpen(&tdc->f.inode);
3033 code = afs_osi_Stat(tfile, &tstat);
3035 osi_Panic("initcachefile stat");
3038 * If file size doesn't match the cache info file, it's probably bad.
3040 if (tdc->f.chunkBytes != tstat.size)
3043 * If file changed within T (120?) seconds of cache info file, it's
3044 * probably bad. In addition, if slot changed within last T seconds,
3045 * the cache info file may be incorrectly identified, and so slot
3048 if (cacheInfoModTime < tstat.mtime + 120)
3050 if (cacheInfoModTime < tdc->f.modTime + 120)
3052 /* In case write through is behind, make sure cache items entry is
3053 * at least as new as the chunk.
3055 if (tdc->f.modTime < tstat.mtime)
3058 tdc->f.chunkBytes = 0;
3061 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3062 if (tfile && tstat.size != 0)
3063 osi_UFSTruncate(tfile, 0);
3064 tdc->f.states &= ~(DRO|DBackup|DRW);
3065 afs_DCMoveBucket(tdc, 0, 0);
3066 /* put entry in free cache slot list */
3067 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3068 afs_freeDCList = index;
3070 afs_indexFlags[index] |= IFFree;
3071 afs_indexUnique[index] = 0;
3074 * We must put this entry in the appropriate hash tables.
3075 * Note that i is still set from the above DCHash call
3077 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3078 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3079 afs_dchashTbl[code] = tdc->index;
3080 code = DVHash(&tdc->f.fid);
3081 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3082 afs_dvhashTbl[code] = tdc->index;
3083 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3085 /* has nontrivial amt of data */
3086 afs_indexFlags[index] |= IFEverUsed;
3087 afs_stats_cmperf.cacheFilesReused++;
3089 * Initialize index times to file's mod times; init indexCounter
3092 hset32(afs_indexTimes[index], tstat.atime);
3093 if (hgetlo(afs_indexCounter) < tstat.atime) {
3094 hset32(afs_indexCounter, tstat.atime);
3096 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3097 } /*File is not bad */
3100 osi_UFSClose(tfile);
3101 tdc->f.states &= ~DWriting;
3102 tdc->dflags &= ~DFEntryMod;
3103 /* don't set f.modTime; we're just cleaning up */
3104 osi_Assert(afs_WriteDCache(tdc, 0) == 0);
3105 ReleaseWriteLock(&afs_xdcache);
3106 ReleaseWriteLock(&tdc->lock);
3108 afs_stats_cmperf.cacheNumEntries++;
3113 /*Max # of struct dcache's resident at any time*/
3115 * If 'dchint' is enabled then in-memory dcache min is increased because of
3121 * Initialize dcache related variables.
3131 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3137 afs_freeDCList = NULLIDX;
3138 afs_discardDCList = NULLIDX;
3139 afs_freeDCCount = 0;
3140 afs_freeDSList = NULL;
3141 hzero(afs_indexCounter);
3143 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3149 if (achunk < 0 || achunk > 30)
3150 achunk = 13; /* Use default */
3151 AFS_SETCHUNKSIZE(achunk);
3157 if (aDentries > 512)
3158 afs_dhashsize = 2048;
3159 /* initialize hash tables */
3160 afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3161 osi_Assert(afs_dvhashTbl != NULL);
3162 afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3163 osi_Assert(afs_dchashTbl != NULL);
3164 for (i = 0; i < afs_dhashsize; i++) {
3165 afs_dvhashTbl[i] = NULLIDX;
3166 afs_dchashTbl[i] = NULLIDX;
3168 afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3169 osi_Assert(afs_dvnextTbl != NULL);
3170 afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3171 osi_Assert(afs_dcnextTbl != NULL);
3172 for (i = 0; i < afiles; i++) {
3173 afs_dvnextTbl[i] = NULLIDX;
3174 afs_dcnextTbl[i] = NULLIDX;
3177 /* Allocate and zero the pointer array to the dcache entries */
3178 afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3179 osi_Assert(afs_indexTable != NULL);
3180 memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3181 afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3182 osi_Assert(afs_indexTimes != NULL);
3183 memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3184 afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
3185 osi_Assert(afs_indexUnique != NULL);
3186 memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3187 afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
3188 osi_Assert(afs_indexFlags != NULL);
3189 memset(afs_indexFlags, 0, afiles * sizeof(char));
3191 /* Allocate and thread the struct dcache entries themselves */
3192 tdp = afs_Initial_freeDSList =
3193 afs_osi_Alloc(aDentries * sizeof(struct dcache));
3194 osi_Assert(tdp != NULL);
3195 memset(tdp, 0, aDentries * sizeof(struct dcache));
3196 #ifdef KERNEL_HAVE_PIN
3197 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3198 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3199 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3200 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3201 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3202 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3203 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3204 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3205 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3208 afs_freeDSList = &tdp[0];
3209 for (i = 0; i < aDentries - 1; i++) {
3210 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3211 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3212 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3213 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3215 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3216 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3217 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3218 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3220 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3221 afs_cacheBlocks = ablocks;
3222 afs_ComputeCacheParms(); /* compute parms based on cache size */
3224 afs_dcentries = aDentries;
3226 afs_stats_cmperf.cacheBucket0_Discarded =
3227 afs_stats_cmperf.cacheBucket1_Discarded =
3228 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3232 if (aflags & AFSCALL_INIT_MEMCACHE) {
3234 * Use a memory cache instead of a disk cache
3236 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3237 afs_cacheType = &afs_MemCacheOps;
3238 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3239 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3240 /* ablocks is reported in 1K blocks */
3241 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3243 afs_warn("afsd: memory cache too large for available memory.\n");
3244 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3246 afiles = ablocks = 0;
3248 afs_warn("Memory cache: Allocating %d dcache entries...",
3251 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3252 afs_cacheType = &afs_UfsCacheOps;
3257 * Shuts down the cache.
3261 shutdown_dcache(void)
3265 #ifdef AFS_CACHE_VNODE_PATH
3266 if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
3268 for (i = 0; i < afs_cacheFiles; i++) {
3269 tdc = afs_indexTable[i];
3271 afs_osi_FreeStr(tdc->f.inode.ufs);
3277 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3278 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3279 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3280 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3281 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3282 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3283 afs_osi_Free(afs_Initial_freeDSList,
3284 afs_dcentries * sizeof(struct dcache));
3285 #ifdef KERNEL_HAVE_PIN
3286 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3287 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3288 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3289 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3290 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3291 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3292 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3296 for (i = 0; i < afs_dhashsize; i++) {
3297 afs_dvhashTbl[i] = NULLIDX;
3298 afs_dchashTbl[i] = NULLIDX;
3301 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3302 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3304 afs_blocksUsed = afs_dcentries = 0;
3305 afs_stats_cmperf.cacheBucket0_Discarded =
3306 afs_stats_cmperf.cacheBucket1_Discarded =
3307 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3308 hzero(afs_indexCounter);
3310 afs_freeDCCount = 0;
3311 afs_freeDCList = NULLIDX;
3312 afs_discardDCList = NULLIDX;
3313 afs_freeDSList = afs_Initial_freeDSList = 0;
3315 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3321 * Get a dcache ready for writing, respecting the current cache size limits
3323 * len is required because afs_GetDCache with flag == 4 expects the length
3324 * field to be filled. It decides from this whether it's necessary to fetch
3325 * data into the chunk before writing or not (when the whole chunk is
3328 * \param avc The vcache to fetch a dcache for
3329 * \param filePos The start of the section to be written
3330 * \param len The length of the section to be written
3334 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3335 * must be released and afs_PutDCache() called to free dcache.
3338 * \note avc->lock must be held on entry. Function may release and reobtain
3339 * avc->lock and GLOCK.
3343 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3344 afs_size_t len, struct vrequest *areq,
3347 struct dcache *tdc = NULL;
3350 /* read the cached info */
3352 tdc = afs_FindDCache(avc, filePos);
3354 ObtainWriteLock(&tdc->lock, 657);
3355 } else if (afs_blocksUsed >
3356 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3357 tdc = afs_FindDCache(avc, filePos);
3359 ObtainWriteLock(&tdc->lock, 658);
3360 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3361 || (tdc->dflags & DFFetching)) {
3362 ReleaseWriteLock(&tdc->lock);
3368 afs_MaybeWakeupTruncateDaemon();
3369 while (afs_blocksUsed >
3370 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3371 ReleaseWriteLock(&avc->lock);
3372 if (afs_blocksUsed - afs_blocksDiscarded >
3373 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3374 afs_WaitForCacheDrain = 1;
3375 afs_osi_Sleep(&afs_WaitForCacheDrain);
3377 afs_MaybeFreeDiscardedDCache();
3378 afs_MaybeWakeupTruncateDaemon();
3379 ObtainWriteLock(&avc->lock, 509);
3381 avc->f.states |= CDirty;
3382 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3384 ObtainWriteLock(&tdc->lock, 659);
3387 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3389 ObtainWriteLock(&tdc->lock, 660);
3392 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3393 afs_stats_cmperf.cacheCurrDirtyChunks++;
3394 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3396 if (!(tdc->f.states & DWriting)) {
3397 /* don't mark entry as mod if we don't have to */
3398 tdc->f.states |= DWriting;
3399 tdc->dflags |= DFEntryMod;
3406 * Make a shadow copy of a dir's dcache. It's used for disconnected
3407 * operations like remove/create/rename to keep the original directory data.
3408 * On reconnection, we can diff the original data with the server and get the
3409 * server changes and with the local data to get the local changes.
3411 * \param avc The dir vnode.
3412 * \param adc The dir dcache.
3414 * \return 0 for success.
3416 * \note The vcache entry must be write locked.
3417 * \note The dcache entry must be read locked.
3420 afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3422 int i, code, ret_code = 0, written, trans_size;
3423 struct dcache *new_dc = NULL;
3424 struct osi_file *tfile_src, *tfile_dst;
3425 struct VenusFid shadow_fid;
3428 /* Is this a dir? */
3429 if (vType(avc) != VDIR)
3432 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3435 /* Generate a fid for the shadow dir. */
3436 shadow_fid.Cell = avc->f.fid.Cell;
3437 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3438 afs_GenShadowFid(&shadow_fid);
3440 ObtainWriteLock(&afs_xdcache, 716);
3442 /* Get a fresh dcache. */
3443 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3445 ObtainReadLock(&adc->mflock);
3447 /* Set up the new fid. */
3448 /* Copy interesting data from original dir dcache. */
3449 new_dc->mflags = adc->mflags;
3450 new_dc->dflags = adc->dflags;
3451 new_dc->f.modTime = adc->f.modTime;
3452 new_dc->f.versionNo = adc->f.versionNo;
3453 new_dc->f.states = adc->f.states;
3454 new_dc->f.chunk= adc->f.chunk;
3455 new_dc->f.chunkBytes = adc->f.chunkBytes;
3457 ReleaseReadLock(&adc->mflock);
3459 /* Now add to the two hash chains */
3460 i = DCHash(&shadow_fid, 0);
3461 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3462 afs_dchashTbl[i] = new_dc->index;
3464 i = DVHash(&shadow_fid);
3465 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3466 afs_dvhashTbl[i] = new_dc->index;
3468 ReleaseWriteLock(&afs_xdcache);
3470 /* Alloc a 4k block. */
3471 data = afs_osi_Alloc(4096);
3473 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3478 /* Open the files. */
3479 tfile_src = afs_CFileOpen(&adc->f.inode);
3480 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3482 /* And now copy dir dcache data into this dcache,
3486 while (written < adc->f.chunkBytes) {
3487 trans_size = adc->f.chunkBytes - written;
3488 if (trans_size > 4096)
3491 /* Read a chunk from the dcache. */
3492 code = afs_CFileRead(tfile_src, written, data, trans_size);
3493 if (code < trans_size) {
3498 /* Write it to the new dcache. */
3499 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3500 if (code < trans_size) {
3505 written+=trans_size;
3508 afs_CFileClose(tfile_dst);
3509 afs_CFileClose(tfile_src);
3511 afs_osi_Free(data, 4096);
3513 ReleaseWriteLock(&new_dc->lock);
3514 afs_PutDCache(new_dc);
3517 ObtainWriteLock(&afs_xvcache, 763);
3518 ObtainWriteLock(&afs_disconDirtyLock, 765);
3519 QAdd(&afs_disconShadow, &avc->shadowq);
3520 osi_Assert((afs_RefVCache(avc) == 0));
3521 ReleaseWriteLock(&afs_disconDirtyLock);
3522 ReleaseWriteLock(&afs_xvcache);
3524 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3525 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3533 * Delete the dcaches of a shadow dir.
3535 * \param avc The vcache containing the shadow fid.
3537 * \note avc must be write locked.
3540 afs_DeleteShadowDir(struct vcache *avc)
3543 struct VenusFid shadow_fid;
3545 shadow_fid.Cell = avc->f.fid.Cell;
3546 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3547 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3548 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3550 tdc = afs_FindDCacheByFid(&shadow_fid);
3552 afs_HashOutDCache(tdc, 1);
3553 afs_DiscardDCache(tdc);
3556 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3557 ObtainWriteLock(&afs_disconDirtyLock, 708);
3558 QRemove(&avc->shadowq);
3559 ReleaseWriteLock(&afs_disconDirtyLock);
3560 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3564 * Populate a dcache with empty chunks up to a given file size,
3565 * used before extending a file in order to avoid 'holes' which
3566 * we can't access in disconnected mode.
3568 * \param avc The vcache which is being extended (locked)
3569 * \param alen The new length of the file
3573 afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
3576 afs_size_t len, offset;
3577 afs_int32 start, end;
3579 /* We're doing this to deal with the situation where we extend
3580 * by writing after lseek()ing past the end of the file . If that
3581 * extension skips chunks, then those chunks won't be created, and
3582 * GetDCache will assume that they have to be fetched from the server.
3583 * So, for each chunk between the current file position, and the new
3584 * length we GetDCache for that chunk.
3587 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3590 if (avc->f.m.Length == 0)
3593 start = AFS_CHUNK(avc->f.m.Length)+1;
3595 end = AFS_CHUNK(apos);
3598 len = AFS_CHUNKTOSIZE(start);
3599 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);