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 * Get a dcache entry from the discard or free list
1072 * @param[in] indexp A pointer to the head of the dcache free list or discard
1073 * list (afs_freeDCList, or afs_discardDCList)
1075 * @return A dcache from that list, or NULL if none could be retrieved.
1077 * @pre afs_xdcache is write-locked
1079 static struct dcache *
1080 afs_GetDSlotFromList(afs_int32 *indexp)
1084 for ( ; *indexp != NULLIDX; indexp = &afs_dvnextTbl[*indexp]) {
1085 tdc = afs_GetUnusedDSlot(*indexp);
1087 osi_Assert(tdc->refCount == 1);
1088 ReleaseReadLock(&tdc->tlock);
1089 *indexp = afs_dvnextTbl[tdc->index];
1090 afs_dvnextTbl[tdc->index] = NULLIDX;
1098 * Free the next element on the list of discarded cache elements.
1101 afs_FreeDiscardedDCache(void)
1104 struct osi_file *tfile;
1107 AFS_STATCNT(afs_FreeDiscardedDCache);
1109 ObtainWriteLock(&afs_xdcache, 510);
1110 if (!afs_blocksDiscarded) {
1111 ReleaseWriteLock(&afs_xdcache);
1116 * Get an entry from the list of discarded cache elements
1118 tdc = afs_GetDSlotFromList(&afs_discardDCList);
1121 afs_discardDCCount--;
1122 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1123 afs_blocksDiscarded -= size;
1124 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1125 /* We can lock because we just took it off the free list */
1126 ObtainWriteLock(&tdc->lock, 626);
1127 ReleaseWriteLock(&afs_xdcache);
1130 * Truncate the element to reclaim its space
1132 tfile = afs_CFileOpen(&tdc->f.inode);
1133 afs_CFileTruncate(tfile, 0);
1134 afs_CFileClose(tfile);
1135 afs_AdjustSize(tdc, 0);
1136 afs_DCMoveBucket(tdc, 0, 0);
1139 * Free the element we just truncated
1141 ObtainWriteLock(&afs_xdcache, 511);
1142 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1143 afs_FreeDCache(tdc);
1144 tdc->f.states &= ~(DRO|DBackup|DRW);
1145 ReleaseWriteLock(&tdc->lock);
1147 ReleaseWriteLock(&afs_xdcache);
1151 * Free as many entries from the list of discarded cache elements
1152 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1157 afs_MaybeFreeDiscardedDCache(void)
1160 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1162 while (afs_blocksDiscarded
1163 && (afs_blocksUsed >
1164 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1165 afs_FreeDiscardedDCache();
1171 * Try to free up a certain number of disk slots.
1173 * \param anumber Targeted number of disk slots to free up.
1175 * \note Environment:
1176 * Must be called with afs_xdcache write-locked.
1180 afs_GetDownDSlot(int anumber)
1182 struct afs_q *tq, *nq;
1187 AFS_STATCNT(afs_GetDownDSlot);
1188 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1189 osi_Panic("diskless getdowndslot");
1191 if (CheckLock(&afs_xdcache) != -1)
1192 osi_Panic("getdowndslot nolock");
1194 /* decrement anumber first for all dudes in free list */
1195 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1198 return; /* enough already free */
1200 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1202 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1203 nq = QPrev(tq); /* in case we remove it */
1204 if (tdc->refCount == 0) {
1205 if ((ix = tdc->index) == NULLIDX)
1206 osi_Panic("getdowndslot");
1207 /* pull the entry out of the lruq and put it on the free list */
1208 QRemove(&tdc->lruq);
1210 /* write-through if modified */
1211 if (tdc->dflags & DFEntryMod) {
1212 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1214 * ask proxy to do this for us - we don't have the stack space
1216 while (tdc->dflags & DFEntryMod) {
1219 s = SPLOCK(afs_sgibklock);
1220 if (afs_sgibklist == NULL) {
1221 /* if slot is free, grab it. */
1222 afs_sgibklist = tdc;
1223 SV_SIGNAL(&afs_sgibksync);
1225 /* wait for daemon to (start, then) finish. */
1226 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1230 tdc->dflags &= ~DFEntryMod;
1231 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
1235 /* finally put the entry in the free list */
1236 afs_indexTable[ix] = NULL;
1237 afs_indexFlags[ix] &= ~IFEverUsed;
1238 tdc->index = NULLIDX;
1239 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1240 afs_freeDSList = tdc;
1244 } /*afs_GetDownDSlot */
1251 * Increment the reference count on a disk cache entry,
1252 * which already has a non-zero refcount. In order to
1253 * increment the refcount of a zero-reference entry, you
1254 * have to hold afs_xdcache.
1257 * adc : Pointer to the dcache entry to increment.
1260 * Nothing interesting.
1263 afs_RefDCache(struct dcache *adc)
1265 ObtainWriteLock(&adc->tlock, 627);
1266 if (adc->refCount < 0)
1267 osi_Panic("RefDCache: negative refcount");
1269 ReleaseWriteLock(&adc->tlock);
1278 * Decrement the reference count on a disk cache entry.
1281 * ad : Ptr to the dcache entry to decrement.
1284 * Nothing interesting.
1287 afs_PutDCache(struct dcache *adc)
1289 AFS_STATCNT(afs_PutDCache);
1290 ObtainWriteLock(&adc->tlock, 276);
1291 if (adc->refCount <= 0)
1292 osi_Panic("putdcache");
1294 ReleaseWriteLock(&adc->tlock);
1303 * Try to discard all data associated with this file from the
1307 * avc : Pointer to the cache info for the file.
1310 * Both pvnLock and lock are write held.
1313 afs_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
1318 AFS_STATCNT(afs_TryToSmush);
1319 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1320 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
1321 sync = 1; /* XX Temp testing XX */
1323 #if defined(AFS_SUN5_ENV)
1324 ObtainWriteLock(&avc->vlock, 573);
1325 avc->activeV++; /* block new getpages */
1326 ReleaseWriteLock(&avc->vlock);
1329 /* Flush VM pages */
1330 osi_VM_TryToSmush(avc, acred, sync);
1333 * Get the hash chain containing all dce's for this fid
1335 i = DVHash(&avc->f.fid);
1336 ObtainWriteLock(&afs_xdcache, 277);
1337 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1338 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1339 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1340 int releaseTlock = 1;
1341 tdc = afs_GetValidDSlot(index);
1343 /* afs_TryToSmush is best-effort; we may not actually discard
1344 * everything, so failure to discard a dcache due to an i/o
1348 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1350 if ((afs_indexFlags[index] & IFDataMod) == 0
1351 && tdc->refCount == 1) {
1352 ReleaseReadLock(&tdc->tlock);
1354 afs_FlushDCache(tdc);
1357 afs_indexTable[index] = 0;
1360 ReleaseReadLock(&tdc->tlock);
1364 #if defined(AFS_SUN5_ENV)
1365 ObtainWriteLock(&avc->vlock, 545);
1366 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1367 avc->vstates &= ~VRevokeWait;
1368 afs_osi_Wakeup((char *)&avc->vstates);
1370 ReleaseWriteLock(&avc->vlock);
1372 ReleaseWriteLock(&afs_xdcache);
1374 * It's treated like a callback so that when we do lookups we'll
1375 * invalidate the unique bit if any
1376 * trytoSmush occured during the lookup call
1382 * afs_DCacheMissingChunks
1385 * Given the cached info for a file, return the number of chunks that
1386 * are not available from the dcache.
1389 * avc: Pointer to the (held) vcache entry to look in.
1392 * The number of chunks which are not currently cached.
1395 * The vcache entry is held upon entry.
1399 afs_DCacheMissingChunks(struct vcache *avc)
1402 afs_size_t totalLength = 0;
1403 afs_uint32 totalChunks = 0;
1406 totalLength = avc->f.m.Length;
1407 if (avc->f.truncPos < totalLength)
1408 totalLength = avc->f.truncPos;
1410 /* Length is 0, no chunk missing. */
1411 if (totalLength == 0)
1414 /* If totalLength is a multiple of chunksize, the last byte appears
1415 * as being part of the next chunk, which does not exist.
1416 * Decrementing totalLength by one fixes that.
1419 totalChunks = (AFS_CHUNK(totalLength) + 1);
1421 /* If we're a directory, we only ever have one chunk, regardless of
1422 * the size of the dir.
1424 if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
1428 printf("Should have %d chunks for %u bytes\n",
1429 totalChunks, (totalLength + 1));
1431 i = DVHash(&avc->f.fid);
1432 ObtainWriteLock(&afs_xdcache, 1001);
1433 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1434 i = afs_dvnextTbl[index];
1435 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1436 tdc = afs_GetValidDSlot(index);
1438 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1441 ReleaseReadLock(&tdc->tlock);
1446 ReleaseWriteLock(&afs_xdcache);
1448 /*printf("Missing %d chunks\n", totalChunks);*/
1450 return (totalChunks);
1457 * Given the cached info for a file and a byte offset into the
1458 * file, make sure the dcache entry for that file and containing
1459 * the given byte is available, returning it to our caller.
1462 * avc : Pointer to the (held) vcache entry to look in.
1463 * abyte : Which byte we want to get to.
1466 * Pointer to the dcache entry covering the file & desired byte,
1467 * or NULL if not found.
1470 * The vcache entry is held upon entry.
1474 afs_FindDCache(struct vcache *avc, afs_size_t abyte)
1478 struct dcache *tdc = NULL;
1480 AFS_STATCNT(afs_FindDCache);
1481 chunk = AFS_CHUNK(abyte);
1484 * Hash on the [fid, chunk] and get the corresponding dcache index
1485 * after write-locking the dcache.
1487 i = DCHash(&avc->f.fid, chunk);
1488 ObtainWriteLock(&afs_xdcache, 278);
1489 for (index = afs_dchashTbl[i]; index != NULLIDX; index = afs_dcnextTbl[index]) {
1490 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1491 tdc = afs_GetValidDSlot(index);
1493 /* afs_FindDCache is best-effort; we may not find the given
1494 * file/offset, so if we cannot find the given dcache due to
1495 * i/o errors, that is okay. */
1498 ReleaseReadLock(&tdc->tlock);
1499 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1500 break; /* leaving refCount high for caller */
1505 if (index != NULLIDX) {
1506 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1507 hadd32(afs_indexCounter, 1);
1508 ReleaseWriteLock(&afs_xdcache);
1511 ReleaseWriteLock(&afs_xdcache);
1513 } /*afs_FindDCache */
1517 * Get a fresh dcache from the free or discarded list.
1519 * \param avc Who's dcache is this going to be?
1520 * \param chunk The position where it will be placed in.
1521 * \param lock How are locks held.
1522 * \param ashFid If this dcache going to be used for a shadow dir,
1525 * \note Required locks:
1527 * - avc (R if (lock & 1) set and W otherwise)
1528 * \note It write locks the new dcache. The caller must unlock it.
1530 * \return The new dcache.
1533 afs_AllocDCache(struct vcache *avc, afs_int32 chunk, afs_int32 lock,
1534 struct VenusFid *ashFid)
1536 struct dcache *tdc = NULL;
1537 afs_uint32 size = 0;
1538 struct osi_file *file;
1540 if (afs_discardDCList == NULLIDX
1541 || ((lock & 2) && afs_freeDCList != NULLIDX)) {
1543 tdc = afs_GetDSlotFromList(&afs_freeDCList);
1545 afs_indexFlags[tdc->index] &= ~IFFree;
1546 ObtainWriteLock(&tdc->lock, 604);
1549 tdc = afs_GetDSlotFromList(&afs_discardDCList);
1551 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1552 ObtainWriteLock(&tdc->lock, 605);
1553 afs_discardDCCount--;
1555 ((tdc->f.chunkBytes +
1556 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1557 tdc->f.states &= ~(DRO|DBackup|DRW);
1558 afs_DCMoveBucket(tdc, size, 0);
1559 afs_blocksDiscarded -= size;
1560 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1562 /* Truncate the chunk so zeroes get filled properly */
1563 file = afs_CFileOpen(&tdc->f.inode);
1564 afs_CFileTruncate(file, 0);
1565 afs_CFileClose(file);
1566 afs_AdjustSize(tdc, 0);
1572 * avc->lock(R) if setLocks
1573 * avc->lock(W) if !setLocks
1579 * Fill in the newly-allocated dcache record.
1581 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1583 /* Use shadow fid if provided. */
1584 tdc->f.fid = *ashFid;
1586 /* Use normal vcache's fid otherwise. */
1587 tdc->f.fid = avc->f.fid;
1588 if (avc->f.states & CRO)
1589 tdc->f.states = DRO;
1590 else if (avc->f.states & CBackup)
1591 tdc->f.states = DBackup;
1593 tdc->f.states = DRW;
1594 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1595 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1597 hones(tdc->f.versionNo); /* invalid value */
1598 tdc->f.chunk = chunk;
1599 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1601 if (tdc->lruq.prev == &tdc->lruq)
1602 osi_Panic("lruq 1");
1611 * This function is called to obtain a reference to data stored in
1612 * the disk cache, locating a chunk of data containing the desired
1613 * byte and returning a reference to the disk cache entry, with its
1614 * reference count incremented.
1618 * avc : Ptr to a vcache entry (unlocked)
1619 * abyte : Byte position in the file desired
1620 * areq : Request structure identifying the requesting user.
1621 * aflags : Settings as follows:
1623 * 2 : Return after creating entry.
1624 * 4 : called from afs_vnop_write.c
1625 * *alen contains length of data to be written.
1627 * aoffset : Set to the offset within the chunk where the resident
1629 * alen : Set to the number of bytes of data after the desired
1630 * byte (including the byte itself) which can be read
1634 * The vcache entry pointed to by avc is unlocked upon entry.
1638 * Update the vnode-to-dcache hint if we can get the vnode lock
1639 * right away. Assumes dcache entry is at least read-locked.
1642 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1644 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1645 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1648 ReleaseWriteLock(&v->lock);
1652 /* avc - Write-locked unless aflags & 1 */
1654 afs_GetDCache(struct vcache *avc, afs_size_t abyte,
1655 struct vrequest *areq, afs_size_t * aoffset,
1656 afs_size_t * alen, int aflags)
1658 afs_int32 i, code, shortcut;
1659 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1660 afs_int32 adjustsize = 0;
1666 afs_size_t Position = 0;
1667 afs_int32 size, tlen; /* size of segment to transfer */
1668 struct afs_FetchOutput *tsmall = 0;
1670 struct osi_file *file;
1671 struct afs_conn *tc;
1673 struct server *newCallback = NULL;
1674 char setNewCallback;
1675 char setVcacheStatus;
1676 char doVcacheUpdate;
1678 int doAdjustSize = 0;
1679 int doReallyAdjustSize = 0;
1680 int overWriteWholeChunk = 0;
1681 struct rx_connection *rxconn;
1684 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1685 int fromReplica; /*Are we reading from a replica? */
1686 int numFetchLoops; /*# times around the fetch/analyze loop */
1687 #endif /* AFS_NOSTATS */
1689 AFS_STATCNT(afs_GetDCache);
1693 setLocks = aflags & 1;
1696 * Determine the chunk number and offset within the chunk corresponding
1697 * to the desired byte.
1699 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1702 chunk = AFS_CHUNK(abyte);
1705 /* come back to here if we waited for the cache to drain. */
1708 setNewCallback = setVcacheStatus = 0;
1712 ObtainWriteLock(&avc->lock, 616);
1714 ObtainReadLock(&avc->lock);
1719 * avc->lock(R) if setLocks && !slowPass
1720 * avc->lock(W) if !setLocks || slowPass
1725 /* check hints first! (might could use bcmp or some such...) */
1726 if ((tdc = avc->dchint)) {
1730 * The locking order between afs_xdcache and dcache lock matters.
1731 * The hint dcache entry could be anywhere, even on the free list.
1732 * Locking afs_xdcache ensures that noone is trying to pull dcache
1733 * entries from the free list, and thereby assuming them to be not
1734 * referenced and not locked.
1736 ObtainReadLock(&afs_xdcache);
1737 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1739 if (dcLocked && (tdc->index != NULLIDX)
1740 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1741 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1742 /* got the right one. It might not be the right version, and it
1743 * might be fetching, but it's the right dcache entry.
1745 /* All this code should be integrated better with what follows:
1746 * I can save a good bit more time under a write lock if I do..
1748 ObtainWriteLock(&tdc->tlock, 603);
1750 ReleaseWriteLock(&tdc->tlock);
1752 ReleaseReadLock(&afs_xdcache);
1755 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1756 && !(tdc->dflags & DFFetching)) {
1758 afs_stats_cmperf.dcacheHits++;
1759 ObtainWriteLock(&afs_xdcache, 559);
1760 QRemove(&tdc->lruq);
1761 QAdd(&afs_DLRU, &tdc->lruq);
1762 ReleaseWriteLock(&afs_xdcache);
1765 * avc->lock(R) if setLocks && !slowPass
1766 * avc->lock(W) if !setLocks || slowPass
1773 ReleaseSharedLock(&tdc->lock);
1774 ReleaseReadLock(&afs_xdcache);
1782 * avc->lock(R) if setLocks && !slowPass
1783 * avc->lock(W) if !setLocks || slowPass
1784 * tdc->lock(S) if tdc
1787 if (!tdc) { /* If the hint wasn't the right dcache entry */
1788 int dslot_error = 0;
1790 * Hash on the [fid, chunk] and get the corresponding dcache index
1791 * after write-locking the dcache.
1796 * avc->lock(R) if setLocks && !slowPass
1797 * avc->lock(W) if !setLocks || slowPass
1800 i = DCHash(&avc->f.fid, chunk);
1801 /* check to make sure our space is fine */
1802 afs_MaybeWakeupTruncateDaemon();
1804 ObtainWriteLock(&afs_xdcache, 280);
1806 for (index = afs_dchashTbl[i]; index != NULLIDX; us = index, index = afs_dcnextTbl[index]) {
1807 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1808 tdc = afs_GetValidDSlot(index);
1810 /* we got an i/o error when trying to get the given dslot,
1811 * but do not bail out just yet; it is possible the dcache
1812 * we're looking for is elsewhere, so it doesn't matter if
1813 * we can't load this one. */
1817 ReleaseReadLock(&tdc->tlock);
1820 * avc->lock(R) if setLocks && !slowPass
1821 * avc->lock(W) if !setLocks || slowPass
1824 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1825 /* Move it up in the beginning of the list */
1826 if (afs_dchashTbl[i] != index) {
1827 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1828 afs_dcnextTbl[index] = afs_dchashTbl[i];
1829 afs_dchashTbl[i] = index;
1831 ReleaseWriteLock(&afs_xdcache);
1832 ObtainSharedLock(&tdc->lock, 606);
1833 break; /* leaving refCount high for caller */
1841 * If we didn't find the entry, we'll create one.
1843 if (index == NULLIDX) {
1846 * avc->lock(R) if setLocks
1847 * avc->lock(W) if !setLocks
1850 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1851 avc, ICL_TYPE_INT32, chunk);
1854 /* We couldn't find the dcache we want, but we hit some i/o
1855 * errors when trying to find it, so we're not sure if the
1856 * dcache we want is in the cache or not. Error out, so we
1857 * don't try to possibly create 2 separate dcaches for the
1858 * same exact data. */
1859 ReleaseWriteLock(&afs_xdcache);
1863 /* Make sure there is a free dcache entry for us to use */
1864 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1867 avc->f.states |= CDCLock;
1868 /* just need slots */
1869 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1871 avc->f.states &= ~CDCLock;
1872 if (afs_discardDCList != NULLIDX
1873 || afs_freeDCList != NULLIDX)
1875 /* If we can't get space for 5 mins we give up and panic */
1876 if (++downDCount > 300) {
1877 osi_Panic("getdcache");
1879 ReleaseWriteLock(&afs_xdcache);
1882 * avc->lock(R) if setLocks
1883 * avc->lock(W) if !setLocks
1885 afs_osi_Wait(1000, 0, 0);
1890 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1893 * Now add to the two hash chains - note that i is still set
1894 * from the above DCHash call.
1896 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1897 afs_dchashTbl[i] = tdc->index;
1898 i = DVHash(&avc->f.fid);
1899 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1900 afs_dvhashTbl[i] = tdc->index;
1901 tdc->dflags = DFEntryMod;
1903 afs_MaybeWakeupTruncateDaemon();
1904 ReleaseWriteLock(&afs_xdcache);
1905 ConvertWToSLock(&tdc->lock);
1910 /* vcache->dcache hint failed */
1913 * avc->lock(R) if setLocks && !slowPass
1914 * avc->lock(W) if !setLocks || slowPass
1917 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1918 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1919 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1920 hgetlo(avc->f.m.DataVersion));
1922 * Here we have the entry in tdc, with its refCount incremented.
1923 * Note: we don't use the S-lock on avc; it costs concurrency when
1924 * storing a file back to the server.
1928 * Not a newly created file so we need to check the file's length and
1929 * compare data versions since someone could have changed the data or we're
1930 * reading a file written elsewhere. We only want to bypass doing no-op
1931 * read rpcs on newly created files (dv of 0) since only then we guarantee
1932 * that this chunk's data hasn't been filled by another client.
1934 size = AFS_CHUNKSIZE(abyte);
1935 if (aflags & 4) /* called from write */
1937 else /* called from read */
1938 tlen = tdc->validPos - abyte;
1939 Position = AFS_CHUNKTOBASE(chunk);
1940 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1941 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1942 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1943 ICL_HANDLE_OFFSET(Position));
1944 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
1946 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
1947 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
1948 overWriteWholeChunk = 1;
1949 if (doAdjustSize || overWriteWholeChunk) {
1950 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1952 #ifdef AFS_SGI64_ENV
1955 #else /* AFS_SGI64_ENV */
1958 #endif /* AFS_SGI64_ENV */
1959 #else /* AFS_SGI_ENV */
1962 #endif /* AFS_SGI_ENV */
1963 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
1964 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1965 #if defined(AFS_SUN5_ENV)
1966 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
1968 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
1970 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1971 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
1972 doReallyAdjustSize = 1;
1974 if (doReallyAdjustSize || overWriteWholeChunk) {
1975 /* no data in file to read at this position */
1976 UpgradeSToWLock(&tdc->lock, 607);
1977 file = afs_CFileOpen(&tdc->f.inode);
1978 afs_CFileTruncate(file, 0);
1979 afs_CFileClose(file);
1980 afs_AdjustSize(tdc, 0);
1981 hset(tdc->f.versionNo, avc->f.m.DataVersion);
1982 tdc->dflags |= DFEntryMod;
1984 ConvertWToSLock(&tdc->lock);
1989 * We must read in the whole chunk if the version number doesn't
1993 /* don't need data, just a unique dcache entry */
1994 ObtainWriteLock(&afs_xdcache, 608);
1995 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1996 hadd32(afs_indexCounter, 1);
1997 ReleaseWriteLock(&afs_xdcache);
1999 updateV2DC(setLocks, avc, tdc, 553);
2000 if (vType(avc) == VDIR)
2003 *aoffset = AFS_CHUNKOFFSET(abyte);
2004 if (tdc->validPos < abyte)
2005 *alen = (afs_size_t) 0;
2007 *alen = tdc->validPos - abyte;
2008 ReleaseSharedLock(&tdc->lock);
2011 ReleaseWriteLock(&avc->lock);
2013 ReleaseReadLock(&avc->lock);
2015 return tdc; /* check if we're done */
2020 * avc->lock(R) if setLocks && !slowPass
2021 * avc->lock(W) if !setLocks || slowPass
2024 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2026 setNewCallback = setVcacheStatus = 0;
2030 * avc->lock(R) if setLocks && !slowPass
2031 * avc->lock(W) if !setLocks || slowPass
2034 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2036 * Version number mismatch.
2039 * If we are disconnected, then we can't do much of anything
2040 * because the data doesn't match the file.
2042 if (AFS_IS_DISCONNECTED) {
2043 ReleaseSharedLock(&tdc->lock);
2046 ReleaseWriteLock(&avc->lock);
2048 ReleaseReadLock(&avc->lock);
2050 /* Flush the Dcache */
2055 UpgradeSToWLock(&tdc->lock, 609);
2058 * If data ever existed for this vnode, and this is a text object,
2059 * do some clearing. Now, you'd think you need only do the flush
2060 * when VTEXT is on, but VTEXT is turned off when the text object
2061 * is freed, while pages are left lying around in memory marked
2062 * with this vnode. If we would reactivate (create a new text
2063 * object from) this vnode, we could easily stumble upon some of
2064 * these old pages in pagein. So, we always flush these guys.
2065 * Sun has a wonderful lack of useful invariants in this system.
2067 * avc->flushDV is the data version # of the file at the last text
2068 * flush. Clearly, at least, we don't have to flush the file more
2069 * often than it changes
2071 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2073 * By here, the cache entry is always write-locked. We can
2074 * deadlock if we call osi_Flush with the cache entry locked...
2075 * Unlock the dcache too.
2077 ReleaseWriteLock(&tdc->lock);
2078 if (setLocks && !slowPass)
2079 ReleaseReadLock(&avc->lock);
2081 ReleaseWriteLock(&avc->lock);
2085 * Call osi_FlushPages in open, read/write, and map, since it
2086 * is too hard here to figure out if we should lock the
2089 if (setLocks && !slowPass)
2090 ObtainReadLock(&avc->lock);
2092 ObtainWriteLock(&avc->lock, 66);
2093 ObtainWriteLock(&tdc->lock, 610);
2098 * avc->lock(R) if setLocks && !slowPass
2099 * avc->lock(W) if !setLocks || slowPass
2103 /* Watch for standard race condition around osi_FlushText */
2104 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2105 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2106 afs_stats_cmperf.dcacheHits++;
2107 ConvertWToSLock(&tdc->lock);
2111 /* Sleep here when cache needs to be drained. */
2112 if (setLocks && !slowPass
2113 && (afs_blocksUsed >
2114 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2115 /* Make sure truncate daemon is running */
2116 afs_MaybeWakeupTruncateDaemon();
2117 ObtainWriteLock(&tdc->tlock, 614);
2118 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2119 ReleaseWriteLock(&tdc->tlock);
2120 ReleaseWriteLock(&tdc->lock);
2121 ReleaseReadLock(&avc->lock);
2122 while ((afs_blocksUsed - afs_blocksDiscarded) >
2123 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2124 afs_WaitForCacheDrain = 1;
2125 afs_osi_Sleep(&afs_WaitForCacheDrain);
2127 afs_MaybeFreeDiscardedDCache();
2128 /* need to check if someone else got the chunk first. */
2129 goto RetryGetDCache;
2132 Position = AFS_CHUNKBASE(abyte);
2133 if (vType(avc) == VDIR) {
2134 size = avc->f.m.Length;
2135 if (size > tdc->f.chunkBytes) {
2136 /* pre-reserve space for file */
2137 afs_AdjustSize(tdc, size);
2139 size = 999999999; /* max size for transfer */
2141 afs_size_t maxGoodLength;
2143 /* estimate how much data we're expecting back from the server,
2144 * and reserve space in the dcache entry for it */
2146 maxGoodLength = avc->f.m.Length;
2147 if (avc->f.truncPos < maxGoodLength)
2148 maxGoodLength = avc->f.truncPos;
2150 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2151 if (Position + size > maxGoodLength)
2152 size = maxGoodLength - Position;
2154 size = 0; /* Handle random races */
2155 if (size > tdc->f.chunkBytes) {
2156 /* pre-reserve estimated space for file */
2157 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2161 /* For the actual fetch, do not limit the request to the
2162 * length of the file. If this results in a read past EOF on
2163 * the server, the server will just reply with less data than
2164 * requested. If we limit ourselves to only requesting data up
2165 * to the avc file length, we open ourselves up to races if the
2166 * file is extended on the server at about the same time.
2168 * However, we must restrict ourselves to the avc->f.truncPos
2169 * length, since this represents an outstanding local
2170 * truncation of the file that will be committed to the
2171 * fileserver when we actually write the fileserver contents.
2172 * If we do not restrict the fetch length based on
2173 * avc->f.truncPos, a different truncate operation extending
2174 * the file length could cause the old data after
2175 * avc->f.truncPos to reappear, instead of extending the file
2176 * with NUL bytes. */
2177 size = AFS_CHUNKSIZE(abyte);
2178 if (Position + size > avc->f.truncPos) {
2179 size = avc->f.truncPos - Position;
2186 if (afs_mariner && !tdc->f.chunk)
2187 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2189 * Right now, we only have one tool, and it's a hammer. So, we
2190 * fetch the whole file.
2192 DZap(tdc); /* pages in cache may be old */
2193 file = afs_CFileOpen(&tdc->f.inode);
2194 afs_RemoveVCB(&avc->f.fid);
2195 tdc->f.states |= DWriting;
2196 tdc->dflags |= DFFetching;
2197 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2198 if (tdc->mflags & DFFetchReq) {
2199 tdc->mflags &= ~DFFetchReq;
2200 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2201 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2202 __FILE__, ICL_TYPE_INT32, __LINE__,
2203 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2206 tsmall = osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2207 setVcacheStatus = 0;
2210 * Remember if we are doing the reading from a replicated volume,
2211 * and how many times we've zipped around the fetch/analyze loop.
2213 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2215 accP = &(afs_stats_cmfullperf.accessinf);
2217 (accP->replicatedRefs)++;
2219 (accP->unreplicatedRefs)++;
2220 #endif /* AFS_NOSTATS */
2221 /* this is a cache miss */
2222 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2223 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2224 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2227 afs_stats_cmperf.dcacheMisses++;
2230 * Dynamic root support: fetch data from local memory.
2232 if (afs_IsDynroot(avc)) {
2236 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2238 dynrootDir += Position;
2239 dynrootLen -= Position;
2240 if (size > dynrootLen)
2244 code = afs_CFileWrite(file, 0, dynrootDir, size);
2252 tdc->validPos = Position + size;
2253 afs_CFileTruncate(file, size); /* prune it */
2254 } else if (afs_IsDynrootMount(avc)) {
2258 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2260 dynrootDir += Position;
2261 dynrootLen -= Position;
2262 if (size > dynrootLen)
2266 code = afs_CFileWrite(file, 0, dynrootDir, size);
2274 tdc->validPos = Position + size;
2275 afs_CFileTruncate(file, size); /* prune it */
2278 * Not a dynamic vnode: do the real fetch.
2283 * avc->lock(R) if setLocks && !slowPass
2284 * avc->lock(W) if !setLocks || slowPass
2288 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
2293 (accP->numReplicasAccessed)++;
2295 #endif /* AFS_NOSTATS */
2296 if (!setLocks || slowPass) {
2297 avc->callback = tc->parent->srvr->server;
2299 newCallback = tc->parent->srvr->server;
2303 code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
2309 /* callback could have been broken (or expired) in a race here,
2310 * but we return the data anyway. It's as good as we knew about
2311 * when we started. */
2313 * validPos is updated by CacheFetchProc, and can only be
2314 * modifed under a dcache write lock, which we've blocked out
2316 size = tdc->validPos - Position; /* actual segment size */
2319 afs_CFileTruncate(file, size); /* prune it */
2321 if (!setLocks || slowPass) {
2322 ObtainWriteLock(&afs_xcbhash, 453);
2323 afs_DequeueCallback(avc);
2324 avc->f.states &= ~(CStatd | CUnique);
2325 avc->callback = NULL;
2326 ReleaseWriteLock(&afs_xcbhash);
2327 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2328 osi_dnlc_purgedp(avc);
2330 /* Something lost. Forget about performance, and go
2331 * back with a vcache write lock.
2333 afs_CFileTruncate(file, 0);
2334 afs_AdjustSize(tdc, 0);
2335 afs_CFileClose(file);
2336 osi_FreeLargeSpace(tsmall);
2338 ReleaseWriteLock(&tdc->lock);
2341 ReleaseReadLock(&avc->lock);
2343 goto RetryGetDCache;
2347 } while (afs_Analyze
2348 (tc, rxconn, code, &avc->f.fid, areq,
2349 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2353 * avc->lock(R) if setLocks && !slowPass
2354 * avc->lock(W) if !setLocks || slowPass
2360 * In the case of replicated access, jot down info on the number of
2361 * attempts it took before we got through or gave up.
2364 if (numFetchLoops <= 1)
2365 (accP->refFirstReplicaOK)++;
2366 if (numFetchLoops > accP->maxReplicasPerRef)
2367 accP->maxReplicasPerRef = numFetchLoops;
2369 #endif /* AFS_NOSTATS */
2371 tdc->dflags &= ~DFFetching;
2372 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2373 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2374 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2375 tdc, ICL_TYPE_INT32, tdc->dflags);
2376 if (avc->execsOrWriters == 0)
2377 tdc->f.states &= ~DWriting;
2379 /* now, if code != 0, we have an error and should punt.
2380 * note that we have the vcache write lock, either because
2381 * !setLocks or slowPass.
2384 afs_CFileTruncate(file, 0);
2385 afs_AdjustSize(tdc, 0);
2386 afs_CFileClose(file);
2387 ZapDCE(tdc); /* sets DFEntryMod */
2388 if (vType(avc) == VDIR) {
2391 tdc->f.states &= ~(DRO|DBackup|DRW);
2392 afs_DCMoveBucket(tdc, 0, 0);
2393 ReleaseWriteLock(&tdc->lock);
2395 if (!afs_IsDynroot(avc)) {
2396 ObtainWriteLock(&afs_xcbhash, 454);
2397 afs_DequeueCallback(avc);
2398 avc->f.states &= ~(CStatd | CUnique);
2399 ReleaseWriteLock(&afs_xcbhash);
2400 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2401 osi_dnlc_purgedp(avc);
2404 * avc->lock(W); assert(!setLocks || slowPass)
2406 osi_Assert(!setLocks || slowPass);
2412 /* otherwise we copy in the just-fetched info */
2413 afs_CFileClose(file);
2414 afs_AdjustSize(tdc, size); /* new size */
2416 * Copy appropriate fields into vcache. Status is
2417 * copied later where we selectively acquire the
2418 * vcache write lock.
2421 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2423 setVcacheStatus = 1;
2424 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2425 tsmall->OutStatus.DataVersion);
2426 tdc->dflags |= DFEntryMod;
2427 afs_indexFlags[tdc->index] |= IFEverUsed;
2428 ConvertWToSLock(&tdc->lock);
2429 } /*Data version numbers don't match */
2432 * Data version numbers match.
2434 afs_stats_cmperf.dcacheHits++;
2435 } /*Data version numbers match */
2437 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2441 * avc->lock(R) if setLocks && !slowPass
2442 * avc->lock(W) if !setLocks || slowPass
2443 * tdc->lock(S) if tdc
2447 * See if this was a reference to a file in the local cell.
2449 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2450 afs_stats_cmperf.dlocalAccesses++;
2452 afs_stats_cmperf.dremoteAccesses++;
2454 /* Fix up LRU info */
2457 ObtainWriteLock(&afs_xdcache, 602);
2458 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2459 hadd32(afs_indexCounter, 1);
2460 ReleaseWriteLock(&afs_xdcache);
2462 /* return the data */
2463 if (vType(avc) == VDIR)
2466 *aoffset = AFS_CHUNKOFFSET(abyte);
2467 *alen = (tdc->f.chunkBytes - *aoffset);
2468 ReleaseSharedLock(&tdc->lock);
2473 * avc->lock(R) if setLocks && !slowPass
2474 * avc->lock(W) if !setLocks || slowPass
2477 /* Fix up the callback and status values in the vcache */
2479 if (setLocks && !slowPass) {
2482 * This is our dirty little secret to parallel fetches.
2483 * We don't write-lock the vcache while doing the fetch,
2484 * but potentially we'll need to update the vcache after
2485 * the fetch is done.
2487 * Drop the read lock and try to re-obtain the write
2488 * lock. If the vcache still has the same DV, it's
2489 * ok to go ahead and install the new data.
2491 afs_hyper_t currentDV, statusDV;
2493 hset(currentDV, avc->f.m.DataVersion);
2495 if (setNewCallback && avc->callback != newCallback)
2499 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2500 tsmall->OutStatus.DataVersion);
2502 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2504 if (setVcacheStatus && !hsame(currentDV, statusDV))
2508 ReleaseReadLock(&avc->lock);
2510 if (doVcacheUpdate) {
2511 ObtainWriteLock(&avc->lock, 615);
2512 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2513 /* We lose. Someone will beat us to it. */
2515 ReleaseWriteLock(&avc->lock);
2520 /* With slow pass, we've already done all the updates */
2522 ReleaseWriteLock(&avc->lock);
2525 /* Check if we need to perform any last-minute fixes with a write-lock */
2526 if (!setLocks || doVcacheUpdate) {
2528 avc->callback = newCallback;
2529 if (tsmall && setVcacheStatus)
2530 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2532 ReleaseWriteLock(&avc->lock);
2536 osi_FreeLargeSpace(tsmall);
2539 } /*afs_GetDCache */
2543 * afs_WriteThroughDSlots
2546 * Sweep through the dcache slots and write out any modified
2547 * in-memory data back on to our caching store.
2553 * The afs_xdcache is write-locked through this whole affair.
2556 afs_WriteThroughDSlots(void)
2559 afs_int32 i, touchedit = 0;
2561 struct afs_q DirtyQ, *tq;
2563 AFS_STATCNT(afs_WriteThroughDSlots);
2566 * Because of lock ordering, we can't grab dcache locks while
2567 * holding afs_xdcache. So we enter xdcache, get a reference
2568 * for every dcache entry, and exit xdcache.
2570 ObtainWriteLock(&afs_xdcache, 283);
2572 for (i = 0; i < afs_cacheFiles; i++) {
2573 tdc = afs_indexTable[i];
2575 /* Grab tlock in case the existing refcount isn't zero */
2576 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2577 ObtainWriteLock(&tdc->tlock, 623);
2579 ReleaseWriteLock(&tdc->tlock);
2581 QAdd(&DirtyQ, &tdc->dirty);
2584 ReleaseWriteLock(&afs_xdcache);
2587 * Now, for each dcache entry we found, check if it's dirty.
2588 * If so, get write-lock, get afs_xdcache, which protects
2589 * afs_cacheInodep, and flush it. Don't forget to put back
2593 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2595 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2597 if (tdc->dflags & DFEntryMod) {
2600 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2602 /* Now that we have the write lock, double-check */
2603 if (wrLock && (tdc->dflags & DFEntryMod)) {
2604 tdc->dflags &= ~DFEntryMod;
2605 ObtainWriteLock(&afs_xdcache, 620);
2606 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
2607 ReleaseWriteLock(&afs_xdcache);
2611 ReleaseWriteLock(&tdc->lock);
2617 ObtainWriteLock(&afs_xdcache, 617);
2618 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2619 /* Touch the file to make sure that the mtime on the file is kept
2620 * up-to-date to avoid losing cached files on cold starts because
2621 * their mtime seems old...
2623 struct afs_fheader theader;
2625 theader.magic = AFS_FHMAGIC;
2626 theader.firstCSize = AFS_FIRSTCSIZE;
2627 theader.otherCSize = AFS_OTHERCSIZE;
2628 theader.version = AFS_CI_VERSION;
2629 theader.dataSize = sizeof(struct fcache);
2630 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2632 ReleaseWriteLock(&afs_xdcache);
2639 * Return a pointer to an freshly initialized dcache entry using
2640 * a memory-based cache. The tlock will be read-locked.
2643 * aslot : Dcache slot to look at.
2644 * needvalid : Whether the specified slot should already exist
2647 * Must be called with afs_xdcache write-locked.
2651 afs_MemGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
2656 AFS_STATCNT(afs_MemGetDSlot);
2657 if (CheckLock(&afs_xdcache) != -1)
2658 osi_Panic("getdslot nolock");
2659 if (aslot < 0 || aslot >= afs_cacheFiles)
2660 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2661 tdc = afs_indexTable[aslot];
2663 QRemove(&tdc->lruq); /* move to queue head */
2664 QAdd(&afs_DLRU, &tdc->lruq);
2665 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2666 ObtainWriteLock(&tdc->tlock, 624);
2668 ConvertWToRLock(&tdc->tlock);
2672 /* if 'indexvalid' is true, the slot must already exist and be populated
2673 * somewhere. for memcache, the only place that dcache entries exist is
2674 * in memory, so if we did not find it above, something is very wrong. */
2675 osi_Assert(!indexvalid);
2677 if (!afs_freeDSList)
2678 afs_GetDownDSlot(4);
2679 if (!afs_freeDSList) {
2680 /* none free, making one is better than a panic */
2681 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2682 tdc = afs_osi_Alloc(sizeof(struct dcache));
2683 osi_Assert(tdc != NULL);
2684 #ifdef KERNEL_HAVE_PIN
2685 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2688 tdc = afs_freeDSList;
2689 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2692 tdc->dflags = 0; /* up-to-date, not in free q */
2694 QAdd(&afs_DLRU, &tdc->lruq);
2695 if (tdc->lruq.prev == &tdc->lruq)
2696 osi_Panic("lruq 3");
2698 /* initialize entry */
2699 tdc->f.fid.Cell = 0;
2700 tdc->f.fid.Fid.Volume = 0;
2702 hones(tdc->f.versionNo);
2703 tdc->f.inode.mem = aslot;
2704 tdc->dflags |= DFEntryMod;
2707 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2710 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2711 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2712 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2715 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2716 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2717 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2718 ObtainReadLock(&tdc->tlock);
2720 afs_indexTable[aslot] = tdc;
2723 } /*afs_MemGetDSlot */
2725 unsigned int last_error = 0, lasterrtime = 0;
2731 * Return a pointer to an freshly initialized dcache entry using
2732 * a UFS-based disk cache. The dcache tlock will be read-locked.
2735 * aslot : Dcache slot to look at.
2736 * indexvalid : 1 if we know the slot we're giving is valid, and thus
2737 * reading the dcache from the disk index should succeed. 0
2738 * if we are initializing a new dcache, and so reading from
2739 * the disk index may fail.
2740 * datavalid : 0 if we are loading a dcache entry from the free or
2741 * discard list, so we know the data in the given dcache is
2742 * not valid. 1 if we are loading a known used dcache, so the
2743 * data in the dcache must be valid.
2746 * afs_xdcache lock write-locked.
2749 afs_UFSGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
2757 AFS_STATCNT(afs_UFSGetDSlot);
2758 if (CheckLock(&afs_xdcache) != -1)
2759 osi_Panic("getdslot nolock");
2760 if (aslot < 0 || aslot >= afs_cacheFiles)
2761 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2762 tdc = afs_indexTable[aslot];
2764 QRemove(&tdc->lruq); /* move to queue head */
2765 QAdd(&afs_DLRU, &tdc->lruq);
2766 /* Grab tlock in case refCount != 0 */
2767 ObtainWriteLock(&tdc->tlock, 625);
2769 ConvertWToRLock(&tdc->tlock);
2773 /* otherwise we should read it in from the cache file */
2774 if (!afs_freeDSList)
2775 afs_GetDownDSlot(4);
2776 if (!afs_freeDSList) {
2777 /* none free, making one is better than a panic */
2778 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2779 tdc = afs_osi_Alloc(sizeof(struct dcache));
2780 osi_Assert(tdc != NULL);
2781 #ifdef KERNEL_HAVE_PIN
2782 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2785 tdc = afs_freeDSList;
2786 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2789 tdc->dflags = 0; /* up-to-date, not in free q */
2791 QAdd(&afs_DLRU, &tdc->lruq);
2792 if (tdc->lruq.prev == &tdc->lruq)
2793 osi_Panic("lruq 3");
2796 * Seek to the aslot'th entry and read it in.
2798 off = sizeof(struct fcache)*aslot + sizeof(struct afs_fheader);
2800 afs_osi_Read(afs_cacheInodep,
2801 off, (char *)(&tdc->f),
2802 sizeof(struct fcache));
2804 if (code != sizeof(struct fcache)) {
2806 #if defined(KERNEL_HAVE_UERROR)
2807 last_error = getuerror();
2811 lasterrtime = osi_Time();
2813 struct osi_stat tstat;
2814 if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
2817 afs_warn("afs: disk cache read error in CacheItems slot %d "
2818 "off %d/%d code %d/%d\n",
2820 off, (int)tstat.size,
2821 (int)code, (int)sizeof(struct fcache));
2822 /* put tdc back on the free dslot list */
2823 QRemove(&tdc->lruq);
2824 tdc->index = NULLIDX;
2825 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
2826 afs_freeDSList = tdc;
2830 if (!afs_CellNumValid(tdc->f.fid.Cell)) {
2833 osi_Panic("afs: needed valid dcache but index %d off %d has "
2834 "invalid cell num %d\n",
2835 (int)aslot, off, (int)tdc->f.fid.Cell);
2839 if (datavalid && tdc->f.fid.Fid.Volume == 0) {
2840 osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
2844 if (!entryok || !datavalid) {
2845 tdc->f.fid.Cell = 0;
2846 tdc->f.fid.Fid.Volume = 0;
2848 hones(tdc->f.versionNo);
2849 tdc->dflags |= DFEntryMod;
2850 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2851 tdc->f.states &= ~(DRO|DBackup|DRW);
2852 afs_DCMoveBucket(tdc, 0, 0);
2855 if (tdc->f.states & DRO) {
2856 afs_DCMoveBucket(tdc, 0, 2);
2857 } else if (tdc->f.states & DBackup) {
2858 afs_DCMoveBucket(tdc, 0, 1);
2860 afs_DCMoveBucket(tdc, 0, 1);
2866 if (tdc->f.chunk >= 0)
2867 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2872 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2873 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2874 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2877 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2878 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2879 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2880 ObtainReadLock(&tdc->tlock);
2883 * If we didn't read into a temporary dcache region, update the
2884 * slot pointer table.
2886 afs_indexTable[aslot] = tdc;
2889 } /*afs_UFSGetDSlot */
2894 * Write a particular dcache entry back to its home in the
2897 * \param adc Pointer to the dcache entry to write.
2898 * \param atime If true, set the modtime on the file to the current time.
2900 * \note Environment:
2901 * Must be called with the afs_xdcache lock at least read-locked,
2902 * and dcache entry at least read-locked.
2903 * The reference count is not changed.
2907 afs_WriteDCache(struct dcache *adc, int atime)
2911 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2913 AFS_STATCNT(afs_WriteDCache);
2914 osi_Assert(WriteLocked(&afs_xdcache));
2916 adc->f.modTime = osi_Time();
2918 if ((afs_indexFlags[adc->index] & (IFFree | IFDiscarded)) == 0 &&
2919 adc->f.fid.Fid.Volume == 0) {
2920 /* If a dcache slot is not on the free or discard list, it must be
2921 * in the hash table. Thus, the volume must be non-zero, since that
2922 * is how we determine whether or not to unhash the entry when kicking
2923 * it out of the cache. Do this check now, since otherwise this can
2924 * cause hash table corruption and a panic later on after we read the
2926 osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
2927 adc->index, (unsigned)afs_indexFlags[adc->index]);
2931 * Seek to the right dcache slot and write the in-memory image out to disk.
2933 afs_cellname_write();
2935 afs_osi_Write(afs_cacheInodep,
2936 sizeof(struct fcache) * adc->index +
2937 sizeof(struct afs_fheader), (char *)(&adc->f),
2938 sizeof(struct fcache));
2939 if (code != sizeof(struct fcache)) {
2940 afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
2941 (long)(sizeof(struct fcache) * adc->index + sizeof(struct afs_fheader)),
2942 (int)code, (int)sizeof(struct fcache));
2951 * Wake up users of a particular file waiting for stores to take
2954 * \param avc Ptr to related vcache entry.
2956 * \note Environment:
2957 * Nothing interesting.
2960 afs_wakeup(struct vcache *avc)
2963 struct brequest *tb;
2965 AFS_STATCNT(afs_wakeup);
2966 for (i = 0; i < NBRS; i++, tb++) {
2967 /* if request is valid and for this file, we've found it */
2968 if (tb->refCount > 0 && avc == tb->vc) {
2971 * If CSafeStore is on, then we don't awaken the guy
2972 * waiting for the store until the whole store has finished.
2973 * Otherwise, we do it now. Note that if CSafeStore is on,
2974 * the BStore routine actually wakes up the user, instead
2976 * I think this is redundant now because this sort of thing
2977 * is already being handled by the higher-level code.
2979 if ((avc->f.states & CSafeStore) == 0) {
2981 tb->flags |= BUVALID;
2982 if (tb->flags & BUWAIT) {
2983 tb->flags &= ~BUWAIT;
2994 * Given a file name and inode, set up that file to be an
2995 * active member in the AFS cache. This also involves checking
2996 * the usability of its data.
2998 * \param afile Name of the cache file to initialize.
2999 * \param ainode Inode of the file.
3001 * \note Environment:
3002 * This function is called only during initialization.
3005 afs_InitCacheFile(char *afile, ino_t ainode)
3010 struct osi_file *tfile;
3011 struct osi_stat tstat;
3014 AFS_STATCNT(afs_InitCacheFile);
3015 index = afs_stats_cmperf.cacheNumEntries;
3016 if (index >= afs_cacheFiles)
3019 ObtainWriteLock(&afs_xdcache, 282);
3020 tdc = afs_GetNewDSlot(index);
3021 ReleaseReadLock(&tdc->tlock);
3022 ReleaseWriteLock(&afs_xdcache);
3024 ObtainWriteLock(&tdc->lock, 621);
3025 ObtainWriteLock(&afs_xdcache, 622);
3026 if (!afile && !ainode) {
3031 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
3033 ReleaseWriteLock(&afs_xdcache);
3034 ReleaseWriteLock(&tdc->lock);
3039 /* Add any other 'complex' inode types here ... */
3040 #if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
3041 tdc->f.inode.ufs = ainode;
3043 osi_Panic("Can't init cache with inode numbers when complex inodes are "
3048 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3050 tfile = osi_UFSOpen(&tdc->f.inode);
3051 code = afs_osi_Stat(tfile, &tstat);
3053 osi_Panic("initcachefile stat");
3056 * If file size doesn't match the cache info file, it's probably bad.
3058 if (tdc->f.chunkBytes != tstat.size)
3061 * If file changed within T (120?) seconds of cache info file, it's
3062 * probably bad. In addition, if slot changed within last T seconds,
3063 * the cache info file may be incorrectly identified, and so slot
3066 if (cacheInfoModTime < tstat.mtime + 120)
3068 if (cacheInfoModTime < tdc->f.modTime + 120)
3070 /* In case write through is behind, make sure cache items entry is
3071 * at least as new as the chunk.
3073 if (tdc->f.modTime < tstat.mtime)
3076 tdc->f.chunkBytes = 0;
3079 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3080 if (tfile && tstat.size != 0)
3081 osi_UFSTruncate(tfile, 0);
3082 tdc->f.states &= ~(DRO|DBackup|DRW);
3083 afs_DCMoveBucket(tdc, 0, 0);
3084 /* put entry in free cache slot list */
3085 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3086 afs_freeDCList = index;
3088 afs_indexFlags[index] |= IFFree;
3089 afs_indexUnique[index] = 0;
3092 * We must put this entry in the appropriate hash tables.
3093 * Note that i is still set from the above DCHash call
3095 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3096 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3097 afs_dchashTbl[code] = tdc->index;
3098 code = DVHash(&tdc->f.fid);
3099 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3100 afs_dvhashTbl[code] = tdc->index;
3101 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3103 /* has nontrivial amt of data */
3104 afs_indexFlags[index] |= IFEverUsed;
3105 afs_stats_cmperf.cacheFilesReused++;
3107 * Initialize index times to file's mod times; init indexCounter
3110 hset32(afs_indexTimes[index], tstat.atime);
3111 if (hgetlo(afs_indexCounter) < tstat.atime) {
3112 hset32(afs_indexCounter, tstat.atime);
3114 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3115 } /*File is not bad */
3118 osi_UFSClose(tfile);
3119 tdc->f.states &= ~DWriting;
3120 tdc->dflags &= ~DFEntryMod;
3121 /* don't set f.modTime; we're just cleaning up */
3122 osi_Assert(afs_WriteDCache(tdc, 0) == 0);
3123 ReleaseWriteLock(&afs_xdcache);
3124 ReleaseWriteLock(&tdc->lock);
3126 afs_stats_cmperf.cacheNumEntries++;
3131 /*Max # of struct dcache's resident at any time*/
3133 * If 'dchint' is enabled then in-memory dcache min is increased because of
3139 * Initialize dcache related variables.
3149 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3155 afs_freeDCList = NULLIDX;
3156 afs_discardDCList = NULLIDX;
3157 afs_freeDCCount = 0;
3158 afs_freeDSList = NULL;
3159 hzero(afs_indexCounter);
3161 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3167 if (achunk < 0 || achunk > 30)
3168 achunk = 13; /* Use default */
3169 AFS_SETCHUNKSIZE(achunk);
3175 if (aDentries > 512)
3176 afs_dhashsize = 2048;
3177 /* initialize hash tables */
3178 afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3179 osi_Assert(afs_dvhashTbl != NULL);
3180 afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3181 osi_Assert(afs_dchashTbl != NULL);
3182 for (i = 0; i < afs_dhashsize; i++) {
3183 afs_dvhashTbl[i] = NULLIDX;
3184 afs_dchashTbl[i] = NULLIDX;
3186 afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3187 osi_Assert(afs_dvnextTbl != NULL);
3188 afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3189 osi_Assert(afs_dcnextTbl != NULL);
3190 for (i = 0; i < afiles; i++) {
3191 afs_dvnextTbl[i] = NULLIDX;
3192 afs_dcnextTbl[i] = NULLIDX;
3195 /* Allocate and zero the pointer array to the dcache entries */
3196 afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3197 osi_Assert(afs_indexTable != NULL);
3198 memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3199 afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3200 osi_Assert(afs_indexTimes != NULL);
3201 memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3202 afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
3203 osi_Assert(afs_indexUnique != NULL);
3204 memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3205 afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
3206 osi_Assert(afs_indexFlags != NULL);
3207 memset(afs_indexFlags, 0, afiles * sizeof(char));
3209 /* Allocate and thread the struct dcache entries themselves */
3210 tdp = afs_Initial_freeDSList =
3211 afs_osi_Alloc(aDentries * sizeof(struct dcache));
3212 osi_Assert(tdp != NULL);
3213 memset(tdp, 0, aDentries * sizeof(struct dcache));
3214 #ifdef KERNEL_HAVE_PIN
3215 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3216 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3217 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3218 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3219 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3220 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3221 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3222 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3223 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3226 afs_freeDSList = &tdp[0];
3227 for (i = 0; i < aDentries - 1; i++) {
3228 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3229 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3230 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3231 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3233 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3234 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3235 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3236 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3238 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3239 afs_cacheBlocks = ablocks;
3240 afs_ComputeCacheParms(); /* compute parms based on cache size */
3242 afs_dcentries = aDentries;
3244 afs_stats_cmperf.cacheBucket0_Discarded =
3245 afs_stats_cmperf.cacheBucket1_Discarded =
3246 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3250 if (aflags & AFSCALL_INIT_MEMCACHE) {
3252 * Use a memory cache instead of a disk cache
3254 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3255 afs_cacheType = &afs_MemCacheOps;
3256 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3257 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3258 /* ablocks is reported in 1K blocks */
3259 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3261 afs_warn("afsd: memory cache too large for available memory.\n");
3262 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3264 afiles = ablocks = 0;
3266 afs_warn("Memory cache: Allocating %d dcache entries...",
3269 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3270 afs_cacheType = &afs_UfsCacheOps;
3275 * Shuts down the cache.
3279 shutdown_dcache(void)
3283 #ifdef AFS_CACHE_VNODE_PATH
3284 if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
3286 for (i = 0; i < afs_cacheFiles; i++) {
3287 tdc = afs_indexTable[i];
3289 afs_osi_FreeStr(tdc->f.inode.ufs);
3295 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3296 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3297 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3298 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3299 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3300 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3301 afs_osi_Free(afs_Initial_freeDSList,
3302 afs_dcentries * sizeof(struct dcache));
3303 #ifdef KERNEL_HAVE_PIN
3304 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3305 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3306 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3307 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3308 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3309 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3310 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3314 for (i = 0; i < afs_dhashsize; i++) {
3315 afs_dvhashTbl[i] = NULLIDX;
3316 afs_dchashTbl[i] = NULLIDX;
3319 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3320 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3322 afs_blocksUsed = afs_dcentries = 0;
3323 afs_stats_cmperf.cacheBucket0_Discarded =
3324 afs_stats_cmperf.cacheBucket1_Discarded =
3325 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3326 hzero(afs_indexCounter);
3328 afs_freeDCCount = 0;
3329 afs_freeDCList = NULLIDX;
3330 afs_discardDCList = NULLIDX;
3331 afs_freeDSList = afs_Initial_freeDSList = 0;
3333 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3339 * Get a dcache ready for writing, respecting the current cache size limits
3341 * len is required because afs_GetDCache with flag == 4 expects the length
3342 * field to be filled. It decides from this whether it's necessary to fetch
3343 * data into the chunk before writing or not (when the whole chunk is
3346 * \param avc The vcache to fetch a dcache for
3347 * \param filePos The start of the section to be written
3348 * \param len The length of the section to be written
3352 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3353 * must be released and afs_PutDCache() called to free dcache.
3356 * \note avc->lock must be held on entry. Function may release and reobtain
3357 * avc->lock and GLOCK.
3361 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3362 afs_size_t len, struct vrequest *areq,
3365 struct dcache *tdc = NULL;
3368 /* read the cached info */
3370 tdc = afs_FindDCache(avc, filePos);
3372 ObtainWriteLock(&tdc->lock, 657);
3373 } else if (afs_blocksUsed >
3374 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3375 tdc = afs_FindDCache(avc, filePos);
3377 ObtainWriteLock(&tdc->lock, 658);
3378 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3379 || (tdc->dflags & DFFetching)) {
3380 ReleaseWriteLock(&tdc->lock);
3386 afs_MaybeWakeupTruncateDaemon();
3387 while (afs_blocksUsed >
3388 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3389 ReleaseWriteLock(&avc->lock);
3390 if (afs_blocksUsed - afs_blocksDiscarded >
3391 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3392 afs_WaitForCacheDrain = 1;
3393 afs_osi_Sleep(&afs_WaitForCacheDrain);
3395 afs_MaybeFreeDiscardedDCache();
3396 afs_MaybeWakeupTruncateDaemon();
3397 ObtainWriteLock(&avc->lock, 509);
3399 avc->f.states |= CDirty;
3400 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3402 ObtainWriteLock(&tdc->lock, 659);
3405 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3407 ObtainWriteLock(&tdc->lock, 660);
3410 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3411 afs_stats_cmperf.cacheCurrDirtyChunks++;
3412 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3414 if (!(tdc->f.states & DWriting)) {
3415 /* don't mark entry as mod if we don't have to */
3416 tdc->f.states |= DWriting;
3417 tdc->dflags |= DFEntryMod;
3424 * Make a shadow copy of a dir's dcache. It's used for disconnected
3425 * operations like remove/create/rename to keep the original directory data.
3426 * On reconnection, we can diff the original data with the server and get the
3427 * server changes and with the local data to get the local changes.
3429 * \param avc The dir vnode.
3430 * \param adc The dir dcache.
3432 * \return 0 for success.
3434 * \note The vcache entry must be write locked.
3435 * \note The dcache entry must be read locked.
3438 afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3440 int i, code, ret_code = 0, written, trans_size;
3441 struct dcache *new_dc = NULL;
3442 struct osi_file *tfile_src, *tfile_dst;
3443 struct VenusFid shadow_fid;
3446 /* Is this a dir? */
3447 if (vType(avc) != VDIR)
3450 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3453 /* Generate a fid for the shadow dir. */
3454 shadow_fid.Cell = avc->f.fid.Cell;
3455 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3456 afs_GenShadowFid(&shadow_fid);
3458 ObtainWriteLock(&afs_xdcache, 716);
3460 /* Get a fresh dcache. */
3461 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3463 ObtainReadLock(&adc->mflock);
3465 /* Set up the new fid. */
3466 /* Copy interesting data from original dir dcache. */
3467 new_dc->mflags = adc->mflags;
3468 new_dc->dflags = adc->dflags;
3469 new_dc->f.modTime = adc->f.modTime;
3470 new_dc->f.versionNo = adc->f.versionNo;
3471 new_dc->f.states = adc->f.states;
3472 new_dc->f.chunk= adc->f.chunk;
3473 new_dc->f.chunkBytes = adc->f.chunkBytes;
3475 ReleaseReadLock(&adc->mflock);
3477 /* Now add to the two hash chains */
3478 i = DCHash(&shadow_fid, 0);
3479 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3480 afs_dchashTbl[i] = new_dc->index;
3482 i = DVHash(&shadow_fid);
3483 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3484 afs_dvhashTbl[i] = new_dc->index;
3486 ReleaseWriteLock(&afs_xdcache);
3488 /* Alloc a 4k block. */
3489 data = afs_osi_Alloc(4096);
3491 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3496 /* Open the files. */
3497 tfile_src = afs_CFileOpen(&adc->f.inode);
3498 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3500 /* And now copy dir dcache data into this dcache,
3504 while (written < adc->f.chunkBytes) {
3505 trans_size = adc->f.chunkBytes - written;
3506 if (trans_size > 4096)
3509 /* Read a chunk from the dcache. */
3510 code = afs_CFileRead(tfile_src, written, data, trans_size);
3511 if (code < trans_size) {
3516 /* Write it to the new dcache. */
3517 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3518 if (code < trans_size) {
3523 written+=trans_size;
3526 afs_CFileClose(tfile_dst);
3527 afs_CFileClose(tfile_src);
3529 afs_osi_Free(data, 4096);
3531 ReleaseWriteLock(&new_dc->lock);
3532 afs_PutDCache(new_dc);
3535 ObtainWriteLock(&afs_xvcache, 763);
3536 ObtainWriteLock(&afs_disconDirtyLock, 765);
3537 QAdd(&afs_disconShadow, &avc->shadowq);
3538 osi_Assert((afs_RefVCache(avc) == 0));
3539 ReleaseWriteLock(&afs_disconDirtyLock);
3540 ReleaseWriteLock(&afs_xvcache);
3542 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3543 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3551 * Delete the dcaches of a shadow dir.
3553 * \param avc The vcache containing the shadow fid.
3555 * \note avc must be write locked.
3558 afs_DeleteShadowDir(struct vcache *avc)
3561 struct VenusFid shadow_fid;
3563 shadow_fid.Cell = avc->f.fid.Cell;
3564 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3565 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3566 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3568 tdc = afs_FindDCacheByFid(&shadow_fid);
3570 afs_HashOutDCache(tdc, 1);
3571 afs_DiscardDCache(tdc);
3574 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3575 ObtainWriteLock(&afs_disconDirtyLock, 708);
3576 QRemove(&avc->shadowq);
3577 ReleaseWriteLock(&afs_disconDirtyLock);
3578 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3582 * Populate a dcache with empty chunks up to a given file size,
3583 * used before extending a file in order to avoid 'holes' which
3584 * we can't access in disconnected mode.
3586 * \param avc The vcache which is being extended (locked)
3587 * \param alen The new length of the file
3591 afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
3594 afs_size_t len, offset;
3595 afs_int32 start, end;
3597 /* We're doing this to deal with the situation where we extend
3598 * by writing after lseek()ing past the end of the file . If that
3599 * extension skips chunks, then those chunks won't be created, and
3600 * GetDCache will assume that they have to be fetched from the server.
3601 * So, for each chunk between the current file position, and the new
3602 * length we GetDCache for that chunk.
3605 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3608 if (avc->f.m.Length == 0)
3611 start = AFS_CHUNK(avc->f.m.Length)+1;
3613 end = AFS_CHUNK(apos);
3616 len = AFS_CHUNKTOSIZE(start);
3617 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);