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 = 1023; /*!< 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 = {
115 struct afs_cacheOps afs_MemCacheOps = {
117 afs_MemCacheTruncate,
128 int cacheDiskType; /*Type of backing disk for cache */
129 struct afs_cacheOps *afs_cacheType;
132 * Where is this vcache's entry associated dcache located/
133 * \param avc The vcache entry.
134 * \return Bucket index:
139 afs_DCGetBucket(struct vcache *avc)
144 /* This should be replaced with some sort of user configurable function */
145 if (avc->f.states & CRO) {
147 } else if (avc->f.states & CBackup) {
157 * Readjust a dcache's size.
159 * \param adc The dcache to be adjusted.
160 * \param oldSize Old size for the dcache.
161 * \param newSize The new size to be adjusted to.
165 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
167 afs_int32 adjustSize = newSize - oldSize;
175 afs_blocksUsed_0 += adjustSize;
176 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
179 afs_blocksUsed_1 += adjustSize;
180 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
183 afs_blocksUsed_2 += adjustSize;
184 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
192 * Move a dcache from one bucket to another.
194 * \param adc Operate on this dcache.
195 * \param size Size in bucket (?).
196 * \param newBucket Destination bucket.
200 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
205 /* Substract size from old bucket. */
209 afs_blocksUsed_0 -= size;
212 afs_blocksUsed_1 -= size;
215 afs_blocksUsed_2 -= size;
219 /* Set new bucket and increase destination bucket size. */
220 adc->bucket = newBucket;
225 afs_blocksUsed_0 += size;
228 afs_blocksUsed_1 += size;
231 afs_blocksUsed_2 += size;
239 * Init split caches size.
244 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
253 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
258 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
259 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
261 /* Short cut: if we don't know about it, try to kill it */
262 if (phase < 2 && afs_blocksUsed_0)
265 if (afs_pct1 > afs_tpct1)
267 if (afs_pct2 > afs_tpct2)
269 return 0; /* unlikely */
274 * Warn about failing to store a file.
276 * \param acode Associated error code.
277 * \param avolume Volume involved.
278 * \param aflags How to handle the output:
279 * aflags & 1: Print out on console
280 * aflags & 2: Print out on controlling tty
282 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
286 afs_StoreWarn(register afs_int32 acode, afs_int32 avolume,
287 register afs_int32 aflags)
289 static char problem_fmt[] =
290 "afs: failed to store file in volume %d (%s)\n";
291 static char problem_fmt_w_error[] =
292 "afs: failed to store file in volume %d (error %d)\n";
293 static char netproblems[] = "network problems";
294 static char partfull[] = "partition full";
295 static char overquota[] = "over quota";
297 AFS_STATCNT(afs_StoreWarn);
303 afs_warn(problem_fmt, avolume, netproblems);
305 afs_warnuser(problem_fmt, avolume, netproblems);
306 } else if (acode == ENOSPC) {
311 afs_warn(problem_fmt, avolume, partfull);
313 afs_warnuser(problem_fmt, avolume, partfull);
316 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
317 * Instead ENOSPC will be sent...
319 if (acode == EDQUOT) {
324 afs_warn(problem_fmt, avolume, overquota);
326 afs_warnuser(problem_fmt, avolume, overquota);
334 afs_warn(problem_fmt_w_error, avolume, acode);
336 afs_warnuser(problem_fmt_w_error, avolume, acode);
341 * Try waking up truncation daemon, if it's worth it.
344 afs_MaybeWakeupTruncateDaemon(void)
346 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
347 afs_CacheTooFull = 1;
348 if (!afs_TruncateDaemonRunning)
349 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
350 } else if (!afs_TruncateDaemonRunning
351 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
352 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
359 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
360 * struct so we need only export one symbol for AIX.
362 static struct CTD_stats {
363 osi_timeval_t CTD_beforeSleep;
364 osi_timeval_t CTD_afterSleep;
365 osi_timeval_t CTD_sleepTime;
366 osi_timeval_t CTD_runTime;
370 u_int afs_min_cache = 0;
373 * Keeps the cache clean and free by truncating uneeded files, when used.
378 afs_CacheTruncateDaemon(void)
380 osi_timeval_t CTD_tmpTime;
384 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
386 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
388 osi_GetuTime(&CTD_stats.CTD_afterSleep);
389 afs_TruncateDaemonRunning = 1;
391 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
392 MObtainWriteLock(&afs_xdcache, 266);
393 if (afs_CacheTooFull) {
394 int space_needed, slots_needed;
395 /* if we get woken up, we should try to clean something out */
396 for (counter = 0; counter < 10; counter++) {
398 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
400 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
401 afs_GetDownD(slots_needed, &space_needed, 0);
402 if ((space_needed <= 0) && (slots_needed <= 0)) {
405 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
408 if (!afs_CacheIsTooFull())
409 afs_CacheTooFull = 0;
410 } /* end of cache cleanup */
411 MReleaseWriteLock(&afs_xdcache);
414 * This is a defensive check to try to avoid starving threads
415 * that may need the global lock so thay can help free some
416 * cache space. If this thread won't be sleeping or truncating
417 * any cache files then give up the global lock so other
418 * threads get a chance to run.
420 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
421 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
422 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
426 * This is where we free the discarded cache elements.
428 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
429 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
430 afs_FreeDiscardedDCache();
433 /* See if we need to continue to run. Someone may have
434 * signalled us while we were executing.
436 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
437 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
438 /* Collect statistics on truncate daemon. */
439 CTD_stats.CTD_nSleeps++;
440 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
441 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
442 CTD_stats.CTD_beforeSleep);
443 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
445 afs_TruncateDaemonRunning = 0;
446 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
447 afs_TruncateDaemonRunning = 1;
449 osi_GetuTime(&CTD_stats.CTD_afterSleep);
450 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
451 CTD_stats.CTD_afterSleep);
452 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
454 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
456 afs_termState = AFSOP_STOP_AFSDB;
458 afs_termState = AFSOP_STOP_RXEVENT;
460 afs_osi_Wakeup(&afs_termState);
468 * Make adjustment for the new size in the disk cache entry
470 * \note Major Assumptions Here:
471 * Assumes that frag size is an integral power of two, less one,
472 * and that this is a two's complement machine. I don't
473 * know of any filesystems which violate this assumption...
475 * \param adc Ptr to dcache entry.
476 * \param anewsize New size desired.
481 afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
483 register afs_int32 oldSize;
485 AFS_STATCNT(afs_AdjustSize);
487 adc->dflags |= DFEntryMod;
488 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
489 adc->f.chunkBytes = newSize;
492 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
493 afs_DCAdjustSize(adc, oldSize, newSize);
494 if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {
496 /* We're growing the file, wakeup the daemon */
497 afs_MaybeWakeupTruncateDaemon();
499 afs_blocksUsed += (newSize - oldSize);
500 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
505 * This routine is responsible for moving at least one entry (but up
506 * to some number of them) from the LRU queue to the free queue.
508 * \param anumber Number of entries that should ideally be moved.
509 * \param aneedSpace How much space we need (1K blocks);
512 * The anumber parameter is just a hint; at least one entry MUST be
513 * moved, or we'll panic. We must be called with afs_xdcache
514 * write-locked. We should try to satisfy both anumber and aneedspace,
515 * whichever is more demanding - need to do several things:
516 * 1. only grab up to anumber victims if aneedSpace <= 0, not
517 * the whole set of MAXATONCE.
518 * 2. dynamically choose MAXATONCE to reflect severity of
519 * demand: something like (*aneedSpace >> (logChunk - 9))
521 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
522 * indicates that the cache is not properly configured/tuned or
523 * something. We should be able to automatically correct that problem.
526 #define MAXATONCE 16 /* max we can obtain at once */
528 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
532 struct VenusFid *afid;
536 register struct vcache *tvc;
537 afs_uint32 victims[MAXATONCE];
538 struct dcache *victimDCs[MAXATONCE];
539 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
540 afs_uint32 victimPtr; /* next free item in victim arrays */
541 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
542 afs_uint32 maxVictimPtr; /* where it is */
545 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
549 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
550 vfslocked = VFS_LOCK_GIANT(afs_globalVFS);
553 AFS_STATCNT(afs_GetDownD);
555 if (CheckLock(&afs_xdcache) != -1)
556 osi_Panic("getdownd nolock");
557 /* decrement anumber first for all dudes in free list */
558 /* SHOULD always decrement anumber first, even if aneedSpace >0,
559 * because we should try to free space even if anumber <=0 */
560 if (!aneedSpace || *aneedSpace <= 0) {
561 anumber -= afs_freeDCCount;
563 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
564 VFS_UNLOCK_GIANT(vfslocked);
566 return; /* enough already free */
570 /* bounds check parameter */
571 if (anumber > MAXATONCE)
572 anumber = MAXATONCE; /* all we can do */
574 /* rewrite so phases include a better eligiblity for gc test*/
576 * The phase variable manages reclaims. Set to 0, the first pass,
577 * we don't reclaim active entries, or other than target bucket.
578 * Set to 1, we reclaim even active ones in target bucket.
579 * Set to 2, we reclaim any inactive one.
580 * Set to 3, we reclaim even active ones.
588 for (i = 0; i < afs_cacheFiles; i++)
589 /* turn off all flags */
590 afs_indexFlags[i] &= ~IFFlag;
592 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
593 /* find oldest entries for reclamation */
594 maxVictimPtr = victimPtr = 0;
595 hzero(maxVictimTime);
596 curbucket = afs_DCWhichBucket(phase, buckethint);
597 /* select victims from access time array */
598 for (i = 0; i < afs_cacheFiles; i++) {
599 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
600 /* skip if dirty or already free */
603 tdc = afs_indexTable[i];
604 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
606 /* Wrong bucket; can't use it! */
609 if (tdc && (tdc->refCount != 0)) {
610 /* Referenced; can't use it! */
613 hset(vtime, afs_indexTimes[i]);
615 /* if we've already looked at this one, skip it */
616 if (afs_indexFlags[i] & IFFlag)
619 if (victimPtr < MAXATONCE) {
620 /* if there's at least one free victim slot left */
621 victims[victimPtr] = i;
622 hset(victimTimes[victimPtr], vtime);
623 if (hcmp(vtime, maxVictimTime) > 0) {
624 hset(maxVictimTime, vtime);
625 maxVictimPtr = victimPtr;
628 } else if (hcmp(vtime, maxVictimTime) < 0) {
630 * We're older than youngest victim, so we replace at
633 /* find youngest (largest LRU) victim */
636 osi_Panic("getdownd local");
638 hset(victimTimes[j], vtime);
639 /* recompute maxVictimTime */
640 hset(maxVictimTime, vtime);
641 for (j = 0; j < victimPtr; j++)
642 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
643 hset(maxVictimTime, victimTimes[j]);
649 /* now really reclaim the victims */
650 j = 0; /* flag to track if we actually got any of the victims */
651 /* first, hold all the victims, since we're going to release the lock
652 * during the truncate operation.
654 for (i = 0; i < victimPtr; i++) {
655 tdc = afs_GetDSlot(victims[i], 0);
656 /* We got tdc->tlock(R) here */
657 if (tdc->refCount == 1)
661 ReleaseReadLock(&tdc->tlock);
665 for (i = 0; i < victimPtr; i++) {
666 /* q is first elt in dcache entry */
668 /* now, since we're dropping the afs_xdcache lock below, we
669 * have to verify, before proceeding, that there are no other
670 * references to this dcache entry, even now. Note that we
671 * compare with 1, since we bumped it above when we called
672 * afs_GetDSlot to preserve the entry's identity.
674 if (tdc && tdc->refCount == 1) {
675 unsigned char chunkFlags;
676 afs_size_t tchunkoffset = 0;
678 /* xdcache is lower than the xvcache lock */
679 MReleaseWriteLock(&afs_xdcache);
680 MObtainReadLock(&afs_xvcache);
681 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
682 MReleaseReadLock(&afs_xvcache);
683 MObtainWriteLock(&afs_xdcache, 527);
685 if (tdc->refCount > 1)
688 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
689 chunkFlags = afs_indexFlags[tdc->index];
690 if (((phase & 1) == 0) && osi_Active(tvc))
692 if (((phase & 1) == 1) && osi_Active(tvc)
693 && (tvc->f.states & CDCLock)
694 && (chunkFlags & IFAnyPages))
696 if (chunkFlags & IFDataMod)
698 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
699 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
700 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
701 ICL_HANDLE_OFFSET(tchunkoffset));
703 #if defined(AFS_SUN5_ENV)
705 * Now we try to invalidate pages. We do this only for
706 * Solaris. For other platforms, it's OK to recycle a
707 * dcache entry out from under a page, because the strategy
708 * function can call afs_GetDCache().
710 if (!skip && (chunkFlags & IFAnyPages)) {
713 MReleaseWriteLock(&afs_xdcache);
714 MObtainWriteLock(&tvc->vlock, 543);
715 if (tvc->multiPage) {
719 /* block locking pages */
720 tvc->vstates |= VPageCleaning;
721 /* block getting new pages */
723 MReleaseWriteLock(&tvc->vlock);
724 /* One last recheck */
725 MObtainWriteLock(&afs_xdcache, 333);
726 chunkFlags = afs_indexFlags[tdc->index];
727 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
728 || (osi_Active(tvc) && (tvc->f.states & CDCLock)
729 && (chunkFlags & IFAnyPages))) {
731 MReleaseWriteLock(&afs_xdcache);
734 MReleaseWriteLock(&afs_xdcache);
736 code = osi_VM_GetDownD(tvc, tdc);
738 MObtainWriteLock(&afs_xdcache, 269);
739 /* we actually removed all pages, clean and dirty */
741 afs_indexFlags[tdc->index] &=
742 ~(IFDirtyPages | IFAnyPages);
745 MReleaseWriteLock(&afs_xdcache);
747 MObtainWriteLock(&tvc->vlock, 544);
748 if (--tvc->activeV == 0
749 && (tvc->vstates & VRevokeWait)) {
750 tvc->vstates &= ~VRevokeWait;
751 afs_osi_Wakeup((char *)&tvc->vstates);
754 if (tvc->vstates & VPageCleaning) {
755 tvc->vstates &= ~VPageCleaning;
756 afs_osi_Wakeup((char *)&tvc->vstates);
759 MReleaseWriteLock(&tvc->vlock);
761 #endif /* AFS_SUN5_ENV */
763 MReleaseWriteLock(&afs_xdcache);
766 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
767 MObtainWriteLock(&afs_xdcache, 528);
768 if (afs_indexFlags[tdc->index] &
769 (IFDataMod | IFDirtyPages | IFAnyPages))
771 if (tdc->refCount > 1)
774 #if defined(AFS_SUN5_ENV)
776 /* no vnode, so IFDirtyPages is spurious (we don't
777 * sweep dcaches on vnode recycling, so we can have
778 * DIRTYPAGES set even when all pages are gone). Just
780 * Hold vcache lock to prevent vnode from being
781 * created while we're clearing IFDirtyPages.
783 afs_indexFlags[tdc->index] &=
784 ~(IFDirtyPages | IFAnyPages);
788 /* skip this guy and mark him as recently used */
789 afs_indexFlags[tdc->index] |= IFFlag;
790 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
791 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
792 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
793 ICL_HANDLE_OFFSET(tchunkoffset));
795 /* flush this dude from the data cache and reclaim;
796 * first, make sure no one will care that we damage
797 * it, by removing it from all hash tables. Then,
798 * melt it down for parts. Note that any concurrent
799 * (new possibility!) calls to GetDownD won't touch
800 * this guy because his reference count is > 0. */
801 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
802 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
803 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
804 ICL_HANDLE_OFFSET(tchunkoffset));
805 AFS_STATCNT(afs_gget);
806 afs_HashOutDCache(tdc, 1);
807 if (tdc->f.chunkBytes != 0) {
811 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
816 afs_DiscardDCache(tdc);
821 j = 1; /* we reclaimed at least one victim */
825 } /* end of for victims loop */
828 /* Phase is 0 and no one was found, so try phase 1 (ignore
829 * osi_Active flag) */
832 for (i = 0; i < afs_cacheFiles; i++)
833 /* turn off all flags */
834 afs_indexFlags[i] &= ~IFFlag;
837 /* found no one in phases 0-5, we're hosed */
841 } /* big while loop */
843 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
844 VFS_UNLOCK_GIANT(vfslocked);
853 * Remove adc from any hash tables that would allow it to be located
854 * again by afs_FindDCache or afs_GetDCache.
856 * \param adc Pointer to dcache entry to remove from hash tables.
858 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
862 afs_HashOutDCache(struct dcache *adc, int zap)
866 AFS_STATCNT(afs_glink);
868 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
870 /* if this guy is in the hash table, pull him out */
871 if (adc->f.fid.Fid.Volume != 0) {
872 /* remove entry from first hash chains */
873 i = DCHash(&adc->f.fid, adc->f.chunk);
874 us = afs_dchashTbl[i];
875 if (us == adc->index) {
876 /* first dude in the list */
877 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
879 /* somewhere on the chain */
880 while (us != NULLIDX) {
881 if (afs_dcnextTbl[us] == adc->index) {
882 /* found item pointing at the one to delete */
883 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
886 us = afs_dcnextTbl[us];
889 osi_Panic("dcache hc");
891 /* remove entry from *other* hash chain */
892 i = DVHash(&adc->f.fid);
893 us = afs_dvhashTbl[i];
894 if (us == adc->index) {
895 /* first dude in the list */
896 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
898 /* somewhere on the chain */
899 while (us != NULLIDX) {
900 if (afs_dvnextTbl[us] == adc->index) {
901 /* found item pointing at the one to delete */
902 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
905 us = afs_dvnextTbl[us];
908 osi_Panic("dcache hv");
913 /* prevent entry from being found on a reboot (it is already out of
914 * the hash table, but after a crash, we just look at fid fields of
915 * stable (old) entries).
917 adc->f.fid.Fid.Volume = 0; /* invalid */
919 /* mark entry as modified */
920 adc->dflags |= DFEntryMod;
925 } /*afs_HashOutDCache */
928 * Flush the given dcache entry, pulling it from hash chains
929 * and truncating the associated cache file.
931 * \param adc Ptr to dcache entry to flush.
934 * This routine must be called with the afs_xdcache lock held
938 afs_FlushDCache(register struct dcache *adc)
940 AFS_STATCNT(afs_FlushDCache);
942 * Bump the number of cache files flushed.
944 afs_stats_cmperf.cacheFlushes++;
946 /* remove from all hash tables */
947 afs_HashOutDCache(adc, 1);
949 /* Free its space; special case null operation, since truncate operation
950 * in UFS is slow even in this case, and this allows us to pre-truncate
951 * these files at more convenient times with fewer locks set
952 * (see afs_GetDownD).
954 if (adc->f.chunkBytes != 0) {
955 afs_DiscardDCache(adc);
956 afs_MaybeWakeupTruncateDaemon();
961 if (afs_WaitForCacheDrain) {
962 if (afs_blocksUsed <=
963 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
964 afs_WaitForCacheDrain = 0;
965 afs_osi_Wakeup(&afs_WaitForCacheDrain);
968 } /*afs_FlushDCache */
972 * Put a dcache entry on the free dcache entry list.
974 * \param adc dcache entry to free.
976 * \note Environment: called with afs_xdcache lock write-locked.
979 afs_FreeDCache(register struct dcache *adc)
981 /* Thread on free list, update free list count and mark entry as
982 * freed in its indexFlags element. Also, ensure DCache entry gets
983 * written out (set DFEntryMod).
986 afs_dvnextTbl[adc->index] = afs_freeDCList;
987 afs_freeDCList = adc->index;
989 afs_indexFlags[adc->index] |= IFFree;
990 adc->dflags |= DFEntryMod;
992 if (afs_WaitForCacheDrain) {
993 if ((afs_blocksUsed - afs_blocksDiscarded) <=
994 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
995 afs_WaitForCacheDrain = 0;
996 afs_osi_Wakeup(&afs_WaitForCacheDrain);
999 } /* afs_FreeDCache */
1002 * Discard the cache element by moving it to the discardDCList.
1003 * This puts the cache element into a quasi-freed state, where
1004 * the space may be reused, but the file has not been truncated.
1006 * \note Major Assumptions Here:
1007 * Assumes that frag size is an integral power of two, less one,
1008 * and that this is a two's complement machine. I don't
1009 * know of any filesystems which violate this assumption...
1011 * \param adr Ptr to dcache entry.
1013 * \note Environment:
1014 * Must be called with afs_xdcache write-locked.
1018 afs_DiscardDCache(register struct dcache *adc)
1020 register afs_int32 size;
1022 AFS_STATCNT(afs_DiscardDCache);
1024 osi_Assert(adc->refCount == 1);
1026 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1027 afs_blocksDiscarded += size;
1028 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1030 afs_dvnextTbl[adc->index] = afs_discardDCList;
1031 afs_discardDCList = adc->index;
1032 afs_discardDCCount++;
1034 adc->f.fid.Fid.Volume = 0;
1035 adc->dflags |= DFEntryMod;
1036 afs_indexFlags[adc->index] |= IFDiscarded;
1038 if (afs_WaitForCacheDrain) {
1039 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1040 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1041 afs_WaitForCacheDrain = 0;
1042 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1046 } /*afs_DiscardDCache */
1049 * Free the next element on the list of discarded cache elements.
1052 afs_FreeDiscardedDCache(void)
1054 register struct dcache *tdc;
1055 register struct osi_file *tfile;
1056 register afs_int32 size;
1058 AFS_STATCNT(afs_FreeDiscardedDCache);
1060 MObtainWriteLock(&afs_xdcache, 510);
1061 if (!afs_blocksDiscarded) {
1062 MReleaseWriteLock(&afs_xdcache);
1067 * Get an entry from the list of discarded cache elements
1069 tdc = afs_GetDSlot(afs_discardDCList, 0);
1070 osi_Assert(tdc->refCount == 1);
1071 ReleaseReadLock(&tdc->tlock);
1073 afs_discardDCList = afs_dvnextTbl[tdc->index];
1074 afs_dvnextTbl[tdc->index] = NULLIDX;
1075 afs_discardDCCount--;
1076 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1077 afs_blocksDiscarded -= size;
1078 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1079 /* We can lock because we just took it off the free list */
1080 ObtainWriteLock(&tdc->lock, 626);
1081 MReleaseWriteLock(&afs_xdcache);
1084 * Truncate the element to reclaim its space
1086 tfile = afs_CFileOpen(&tdc->f.inode);
1087 afs_CFileTruncate(tfile, 0);
1088 afs_CFileClose(tfile);
1089 afs_AdjustSize(tdc, 0);
1090 afs_DCMoveBucket(tdc, 0, 0);
1093 * Free the element we just truncated
1095 MObtainWriteLock(&afs_xdcache, 511);
1096 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1097 afs_FreeDCache(tdc);
1098 tdc->f.states &= ~(DRO|DBackup|DRW);
1099 ReleaseWriteLock(&tdc->lock);
1101 MReleaseWriteLock(&afs_xdcache);
1105 * Free as many entries from the list of discarded cache elements
1106 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1111 afs_MaybeFreeDiscardedDCache(void)
1114 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1116 while (afs_blocksDiscarded
1117 && (afs_blocksUsed >
1118 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1119 afs_FreeDiscardedDCache();
1125 * Try to free up a certain number of disk slots.
1127 * \param anumber Targeted number of disk slots to free up.
1129 * \note Environment:
1130 * Must be called with afs_xdcache write-locked.
1134 afs_GetDownDSlot(int anumber)
1136 struct afs_q *tq, *nq;
1141 AFS_STATCNT(afs_GetDownDSlot);
1142 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1143 osi_Panic("diskless getdowndslot");
1145 if (CheckLock(&afs_xdcache) != -1)
1146 osi_Panic("getdowndslot nolock");
1148 /* decrement anumber first for all dudes in free list */
1149 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1152 return; /* enough already free */
1154 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1156 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1157 nq = QPrev(tq); /* in case we remove it */
1158 if (tdc->refCount == 0) {
1159 if ((ix = tdc->index) == NULLIDX)
1160 osi_Panic("getdowndslot");
1161 /* pull the entry out of the lruq and put it on the free list */
1162 QRemove(&tdc->lruq);
1164 /* write-through if modified */
1165 if (tdc->dflags & DFEntryMod) {
1166 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1168 * ask proxy to do this for us - we don't have the stack space
1170 while (tdc->dflags & DFEntryMod) {
1173 s = SPLOCK(afs_sgibklock);
1174 if (afs_sgibklist == NULL) {
1175 /* if slot is free, grab it. */
1176 afs_sgibklist = tdc;
1177 SV_SIGNAL(&afs_sgibksync);
1179 /* wait for daemon to (start, then) finish. */
1180 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1184 tdc->dflags &= ~DFEntryMod;
1185 afs_WriteDCache(tdc, 1);
1189 /* finally put the entry in the free list */
1190 afs_indexTable[ix] = NULL;
1191 afs_indexFlags[ix] &= ~IFEverUsed;
1192 tdc->index = NULLIDX;
1193 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1194 afs_freeDSList = tdc;
1198 } /*afs_GetDownDSlot */
1205 * Increment the reference count on a disk cache entry,
1206 * which already has a non-zero refcount. In order to
1207 * increment the refcount of a zero-reference entry, you
1208 * have to hold afs_xdcache.
1211 * adc : Pointer to the dcache entry to increment.
1214 * Nothing interesting.
1217 afs_RefDCache(struct dcache *adc)
1219 ObtainWriteLock(&adc->tlock, 627);
1220 if (adc->refCount < 0)
1221 osi_Panic("RefDCache: negative refcount");
1223 ReleaseWriteLock(&adc->tlock);
1232 * Decrement the reference count on a disk cache entry.
1235 * ad : Ptr to the dcache entry to decrement.
1238 * Nothing interesting.
1241 afs_PutDCache(register struct dcache *adc)
1243 AFS_STATCNT(afs_PutDCache);
1244 ObtainWriteLock(&adc->tlock, 276);
1245 if (adc->refCount <= 0)
1246 osi_Panic("putdcache");
1248 ReleaseWriteLock(&adc->tlock);
1257 * Try to discard all data associated with this file from the
1261 * avc : Pointer to the cache info for the file.
1264 * Both pvnLock and lock are write held.
1267 afs_TryToSmush(register struct vcache *avc, struct AFS_UCRED *acred, int sync)
1269 register struct dcache *tdc;
1272 AFS_STATCNT(afs_TryToSmush);
1273 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1274 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
1275 sync = 1; /* XX Temp testing XX */
1277 #if defined(AFS_SUN5_ENV)
1278 ObtainWriteLock(&avc->vlock, 573);
1279 avc->activeV++; /* block new getpages */
1280 ReleaseWriteLock(&avc->vlock);
1283 /* Flush VM pages */
1284 osi_VM_TryToSmush(avc, acred, sync);
1287 * Get the hash chain containing all dce's for this fid
1289 i = DVHash(&avc->f.fid);
1290 MObtainWriteLock(&afs_xdcache, 277);
1291 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1292 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1293 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1294 int releaseTlock = 1;
1295 tdc = afs_GetDSlot(index, NULL);
1296 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1298 if ((afs_indexFlags[index] & IFDataMod) == 0
1299 && tdc->refCount == 1) {
1300 ReleaseReadLock(&tdc->tlock);
1302 afs_FlushDCache(tdc);
1305 afs_indexTable[index] = 0;
1308 ReleaseReadLock(&tdc->tlock);
1312 #if defined(AFS_SUN5_ENV)
1313 ObtainWriteLock(&avc->vlock, 545);
1314 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1315 avc->vstates &= ~VRevokeWait;
1316 afs_osi_Wakeup((char *)&avc->vstates);
1318 ReleaseWriteLock(&avc->vlock);
1320 MReleaseWriteLock(&afs_xdcache);
1322 * It's treated like a callback so that when we do lookups we'll
1323 * invalidate the unique bit if any
1324 * trytoSmush occured during the lookup call
1330 * afs_DCacheMissingChunks
1333 * Given the cached info for a file, return the number of chunks that
1334 * are not available from the dcache.
1337 * avc: Pointer to the (held) vcache entry to look in.
1340 * The number of chunks which are not currently cached.
1343 * The vcache entry is held upon entry.
1347 afs_DCacheMissingChunks(struct vcache *avc)
1350 afs_size_t totalLength = 0;
1351 afs_uint32 totalChunks = 0;
1354 totalLength = avc->f.m.Length;
1355 if (avc->f.truncPos < totalLength)
1356 totalLength = avc->f.truncPos;
1358 /* Length is 0, no chunk missing. */
1359 if (totalLength == 0)
1362 /* If totalLength is a multiple of chunksize, the last byte appears
1363 * as being part of the next chunk, which does not exist.
1364 * Decrementing totalLength by one fixes that.
1367 totalChunks = (AFS_CHUNK(totalLength) + 1);
1369 /* If we're a directory, we only ever have one chunk, regardless of
1370 * the size of the dir.
1372 if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
1376 printf("Should have %d chunks for %u bytes\n",
1377 totalChunks, (totalLength + 1));
1379 i = DVHash(&avc->f.fid);
1380 MObtainWriteLock(&afs_xdcache, 1001);
1381 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1382 i = afs_dvnextTbl[index];
1383 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1384 tdc = afs_GetDSlot(index, NULL);
1385 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1388 ReleaseReadLock(&tdc->tlock);
1392 MReleaseWriteLock(&afs_xdcache);
1394 /*printf("Missing %d chunks\n", totalChunks);*/
1396 return (totalChunks);
1403 * Given the cached info for a file and a byte offset into the
1404 * file, make sure the dcache entry for that file and containing
1405 * the given byte is available, returning it to our caller.
1408 * avc : Pointer to the (held) vcache entry to look in.
1409 * abyte : Which byte we want to get to.
1412 * Pointer to the dcache entry covering the file & desired byte,
1413 * or NULL if not found.
1416 * The vcache entry is held upon entry.
1420 afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1423 register afs_int32 i, index;
1424 register struct dcache *tdc = NULL;
1426 AFS_STATCNT(afs_FindDCache);
1427 chunk = AFS_CHUNK(abyte);
1430 * Hash on the [fid, chunk] and get the corresponding dcache index
1431 * after write-locking the dcache.
1433 i = DCHash(&avc->f.fid, chunk);
1434 MObtainWriteLock(&afs_xdcache, 278);
1435 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1436 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1437 tdc = afs_GetDSlot(index, NULL);
1438 ReleaseReadLock(&tdc->tlock);
1439 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1440 break; /* leaving refCount high for caller */
1444 index = afs_dcnextTbl[index];
1446 if (index != NULLIDX) {
1447 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1448 hadd32(afs_indexCounter, 1);
1449 MReleaseWriteLock(&afs_xdcache);
1452 MReleaseWriteLock(&afs_xdcache);
1454 } /*afs_FindDCache */
1458 * Get a fresh dcache from the free or discarded list.
1460 * \param avc Who's dcache is this going to be?
1461 * \param chunk The position where it will be placed in.
1462 * \param lock How are locks held.
1463 * \param ashFid If this dcache going to be used for a shadow dir,
1466 * \note Required locks:
1468 * - avc (R if (lock & 1) set and W otherwise)
1469 * \note It write locks the new dcache. The caller must unlock it.
1471 * \return The new dcache.
1473 struct dcache *afs_AllocDCache(struct vcache *avc,
1476 struct VenusFid *ashFid)
1478 struct dcache *tdc = NULL;
1479 afs_uint32 size = 0;
1480 struct osi_file *file;
1482 if (afs_discardDCList == NULLIDX
1483 || ((lock & 2) && afs_freeDCList != NULLIDX)) {
1485 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1486 tdc = afs_GetDSlot(afs_freeDCList, 0);
1487 osi_Assert(tdc->refCount == 1);
1488 ReleaseReadLock(&tdc->tlock);
1489 ObtainWriteLock(&tdc->lock, 604);
1490 afs_freeDCList = afs_dvnextTbl[tdc->index];
1493 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1494 tdc = afs_GetDSlot(afs_discardDCList, 0);
1495 osi_Assert(tdc->refCount == 1);
1496 ReleaseReadLock(&tdc->tlock);
1497 ObtainWriteLock(&tdc->lock, 605);
1498 afs_discardDCList = afs_dvnextTbl[tdc->index];
1499 afs_discardDCCount--;
1501 ((tdc->f.chunkBytes +
1502 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1503 tdc->f.states &= ~(DRO|DBackup|DRW);
1504 afs_DCMoveBucket(tdc, size, 0);
1505 afs_blocksDiscarded -= size;
1506 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1508 /* Truncate the chunk so zeroes get filled properly */
1509 file = afs_CFileOpen(&tdc->f.inode);
1510 afs_CFileTruncate(file, 0);
1511 afs_CFileClose(file);
1512 afs_AdjustSize(tdc, 0);
1518 * avc->lock(R) if setLocks
1519 * avc->lock(W) if !setLocks
1525 * Fill in the newly-allocated dcache record.
1527 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1529 /* Use shadow fid if provided. */
1530 tdc->f.fid = *ashFid;
1532 /* Use normal vcache's fid otherwise. */
1533 tdc->f.fid = avc->f.fid;
1534 if (avc->f.states & CRO)
1535 tdc->f.states = DRO;
1536 else if (avc->f.states & CBackup)
1537 tdc->f.states = DBackup;
1539 tdc->f.states = DRW;
1540 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1541 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1543 hones(tdc->f.versionNo); /* invalid value */
1544 tdc->f.chunk = chunk;
1545 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1547 if (tdc->lruq.prev == &tdc->lruq)
1548 osi_Panic("lruq 1");
1557 * This function is called to obtain a reference to data stored in
1558 * the disk cache, locating a chunk of data containing the desired
1559 * byte and returning a reference to the disk cache entry, with its
1560 * reference count incremented.
1564 * avc : Ptr to a vcache entry (unlocked)
1565 * abyte : Byte position in the file desired
1566 * areq : Request structure identifying the requesting user.
1567 * aflags : Settings as follows:
1569 * 2 : Return after creating entry.
1570 * 4 : called from afs_vnop_write.c
1571 * *alen contains length of data to be written.
1573 * aoffset : Set to the offset within the chunk where the resident
1575 * alen : Set to the number of bytes of data after the desired
1576 * byte (including the byte itself) which can be read
1580 * The vcache entry pointed to by avc is unlocked upon entry.
1584 * Update the vnode-to-dcache hint if we can get the vnode lock
1585 * right away. Assumes dcache entry is at least read-locked.
1588 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1590 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1591 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1594 ReleaseWriteLock(&v->lock);
1598 /* avc - Write-locked unless aflags & 1 */
1600 afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1601 register struct vrequest *areq, afs_size_t * aoffset,
1602 afs_size_t * alen, int aflags)
1604 register afs_int32 i, code, shortcut;
1605 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1606 register afs_int32 adjustsize = 0;
1612 afs_size_t maxGoodLength; /* amount of good data at server */
1613 afs_size_t Position = 0;
1614 afs_int32 size, tlen; /* size of segment to transfer */
1615 struct afs_FetchOutput *tsmall = 0;
1616 register struct dcache *tdc;
1617 register struct osi_file *file;
1618 register struct afs_conn *tc;
1620 struct server *newCallback = NULL;
1621 char setNewCallback;
1622 char setVcacheStatus;
1623 char doVcacheUpdate;
1625 int doAdjustSize = 0;
1626 int doReallyAdjustSize = 0;
1627 int overWriteWholeChunk = 0;
1630 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1631 int fromReplica; /*Are we reading from a replica? */
1632 int numFetchLoops; /*# times around the fetch/analyze loop */
1633 #endif /* AFS_NOSTATS */
1635 AFS_STATCNT(afs_GetDCache);
1639 setLocks = aflags & 1;
1642 * Determine the chunk number and offset within the chunk corresponding
1643 * to the desired byte.
1645 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1648 chunk = AFS_CHUNK(abyte);
1651 /* come back to here if we waited for the cache to drain. */
1654 setNewCallback = setVcacheStatus = 0;
1658 ObtainWriteLock(&avc->lock, 616);
1660 ObtainReadLock(&avc->lock);
1665 * avc->lock(R) if setLocks && !slowPass
1666 * avc->lock(W) if !setLocks || slowPass
1671 /* check hints first! (might could use bcmp or some such...) */
1672 if ((tdc = avc->dchint)) {
1676 * The locking order between afs_xdcache and dcache lock matters.
1677 * The hint dcache entry could be anywhere, even on the free list.
1678 * Locking afs_xdcache ensures that noone is trying to pull dcache
1679 * entries from the free list, and thereby assuming them to be not
1680 * referenced and not locked.
1682 MObtainReadLock(&afs_xdcache);
1683 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1685 if (dcLocked && (tdc->index != NULLIDX)
1686 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1687 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1688 /* got the right one. It might not be the right version, and it
1689 * might be fetching, but it's the right dcache entry.
1691 /* All this code should be integrated better with what follows:
1692 * I can save a good bit more time under a write lock if I do..
1694 ObtainWriteLock(&tdc->tlock, 603);
1696 ReleaseWriteLock(&tdc->tlock);
1698 MReleaseReadLock(&afs_xdcache);
1701 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1702 && !(tdc->dflags & DFFetching)) {
1704 afs_stats_cmperf.dcacheHits++;
1705 MObtainWriteLock(&afs_xdcache, 559);
1706 QRemove(&tdc->lruq);
1707 QAdd(&afs_DLRU, &tdc->lruq);
1708 MReleaseWriteLock(&afs_xdcache);
1711 * avc->lock(R) if setLocks && !slowPass
1712 * avc->lock(W) if !setLocks || slowPass
1719 ReleaseSharedLock(&tdc->lock);
1720 MReleaseReadLock(&afs_xdcache);
1728 * avc->lock(R) if setLocks && !slowPass
1729 * avc->lock(W) if !setLocks || slowPass
1730 * tdc->lock(S) if tdc
1733 if (!tdc) { /* If the hint wasn't the right dcache entry */
1735 * Hash on the [fid, chunk] and get the corresponding dcache index
1736 * after write-locking the dcache.
1741 * avc->lock(R) if setLocks && !slowPass
1742 * avc->lock(W) if !setLocks || slowPass
1745 i = DCHash(&avc->f.fid, chunk);
1746 /* check to make sure our space is fine */
1747 afs_MaybeWakeupTruncateDaemon();
1749 MObtainWriteLock(&afs_xdcache, 280);
1751 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1752 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1753 tdc = afs_GetDSlot(index, NULL);
1754 ReleaseReadLock(&tdc->tlock);
1757 * avc->lock(R) if setLocks && !slowPass
1758 * avc->lock(W) if !setLocks || slowPass
1761 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1762 /* Move it up in the beginning of the list */
1763 if (afs_dchashTbl[i] != index) {
1764 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1765 afs_dcnextTbl[index] = afs_dchashTbl[i];
1766 afs_dchashTbl[i] = index;
1768 MReleaseWriteLock(&afs_xdcache);
1769 ObtainSharedLock(&tdc->lock, 606);
1770 break; /* leaving refCount high for caller */
1776 index = afs_dcnextTbl[index];
1780 * If we didn't find the entry, we'll create one.
1782 if (index == NULLIDX) {
1785 * avc->lock(R) if setLocks
1786 * avc->lock(W) if !setLocks
1789 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1790 avc, ICL_TYPE_INT32, chunk);
1792 /* Make sure there is a free dcache entry for us to use */
1793 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1796 avc->f.states |= CDCLock;
1797 /* just need slots */
1798 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1800 avc->f.states &= ~CDCLock;
1801 if (afs_discardDCList != NULLIDX
1802 || afs_freeDCList != NULLIDX)
1804 /* If we can't get space for 5 mins we give up and panic */
1805 if (++downDCount > 300) {
1806 #if defined(AFS_CACHE_BYPASS)
1807 afs_warn("GetDCache calling osi_Panic: No space in five minutes.\n downDCount: %d\n aoffset: %d alen: %d\n", downDCount, aoffset, alen);
1809 osi_Panic("getdcache");
1811 MReleaseWriteLock(&afs_xdcache);
1814 * avc->lock(R) if setLocks
1815 * avc->lock(W) if !setLocks
1817 afs_osi_Wait(1000, 0, 0);
1822 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1825 * Now add to the two hash chains - note that i is still set
1826 * from the above DCHash call.
1828 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1829 afs_dchashTbl[i] = tdc->index;
1830 i = DVHash(&avc->f.fid);
1831 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1832 afs_dvhashTbl[i] = tdc->index;
1833 tdc->dflags = DFEntryMod;
1835 afs_MaybeWakeupTruncateDaemon();
1836 MReleaseWriteLock(&afs_xdcache);
1837 ConvertWToSLock(&tdc->lock);
1842 /* vcache->dcache hint failed */
1845 * avc->lock(R) if setLocks && !slowPass
1846 * avc->lock(W) if !setLocks || slowPass
1849 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1850 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1851 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1852 hgetlo(avc->f.m.DataVersion));
1854 * Here we have the entry in tdc, with its refCount incremented.
1855 * Note: we don't use the S-lock on avc; it costs concurrency when
1856 * storing a file back to the server.
1860 * Not a newly created file so we need to check the file's length and
1861 * compare data versions since someone could have changed the data or we're
1862 * reading a file written elsewhere. We only want to bypass doing no-op
1863 * read rpcs on newly created files (dv of 0) since only then we guarantee
1864 * that this chunk's data hasn't been filled by another client.
1866 size = AFS_CHUNKSIZE(abyte);
1867 if (aflags & 4) /* called from write */
1869 else /* called from read */
1870 tlen = tdc->validPos - abyte;
1871 Position = AFS_CHUNKTOBASE(chunk);
1872 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1873 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1874 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1875 ICL_HANDLE_OFFSET(Position));
1876 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
1878 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
1879 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
1880 overWriteWholeChunk = 1;
1881 if (doAdjustSize || overWriteWholeChunk) {
1882 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1884 #ifdef AFS_SGI64_ENV
1887 #else /* AFS_SGI64_ENV */
1890 #endif /* AFS_SGI64_ENV */
1891 #else /* AFS_SGI_ENV */
1894 #endif /* AFS_SGI_ENV */
1895 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
1896 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1897 #if defined(AFS_SUN5_ENV) || defined(AFS_OSF_ENV)
1898 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
1900 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
1902 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1903 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
1904 doReallyAdjustSize = 1;
1906 if (doReallyAdjustSize || overWriteWholeChunk) {
1907 /* no data in file to read at this position */
1908 UpgradeSToWLock(&tdc->lock, 607);
1909 file = afs_CFileOpen(&tdc->f.inode);
1910 afs_CFileTruncate(file, 0);
1911 afs_CFileClose(file);
1912 afs_AdjustSize(tdc, 0);
1913 hset(tdc->f.versionNo, avc->f.m.DataVersion);
1914 tdc->dflags |= DFEntryMod;
1916 ConvertWToSLock(&tdc->lock);
1921 * We must read in the whole chunk if the version number doesn't
1925 /* don't need data, just a unique dcache entry */
1926 ObtainWriteLock(&afs_xdcache, 608);
1927 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1928 hadd32(afs_indexCounter, 1);
1929 ReleaseWriteLock(&afs_xdcache);
1931 updateV2DC(setLocks, avc, tdc, 553);
1932 if (vType(avc) == VDIR)
1935 *aoffset = AFS_CHUNKOFFSET(abyte);
1936 if (tdc->validPos < abyte)
1937 *alen = (afs_size_t) 0;
1939 *alen = tdc->validPos - abyte;
1940 ReleaseSharedLock(&tdc->lock);
1943 ReleaseWriteLock(&avc->lock);
1945 ReleaseReadLock(&avc->lock);
1947 return tdc; /* check if we're done */
1952 * avc->lock(R) if setLocks && !slowPass
1953 * avc->lock(W) if !setLocks || slowPass
1956 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
1958 setNewCallback = setVcacheStatus = 0;
1962 * avc->lock(R) if setLocks && !slowPass
1963 * avc->lock(W) if !setLocks || slowPass
1966 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
1968 * Version number mismatch.
1971 * If we are disconnected, then we can't do much of anything
1972 * because the data doesn't match the file.
1974 if (AFS_IS_DISCONNECTED) {
1975 ReleaseSharedLock(&tdc->lock);
1978 ReleaseWriteLock(&avc->lock);
1980 ReleaseReadLock(&avc->lock);
1982 /* Flush the Dcache */
1987 UpgradeSToWLock(&tdc->lock, 609);
1990 * If data ever existed for this vnode, and this is a text object,
1991 * do some clearing. Now, you'd think you need only do the flush
1992 * when VTEXT is on, but VTEXT is turned off when the text object
1993 * is freed, while pages are left lying around in memory marked
1994 * with this vnode. If we would reactivate (create a new text
1995 * object from) this vnode, we could easily stumble upon some of
1996 * these old pages in pagein. So, we always flush these guys.
1997 * Sun has a wonderful lack of useful invariants in this system.
1999 * avc->flushDV is the data version # of the file at the last text
2000 * flush. Clearly, at least, we don't have to flush the file more
2001 * often than it changes
2003 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2005 * By here, the cache entry is always write-locked. We can
2006 * deadlock if we call osi_Flush with the cache entry locked...
2007 * Unlock the dcache too.
2009 ReleaseWriteLock(&tdc->lock);
2010 if (setLocks && !slowPass)
2011 ReleaseReadLock(&avc->lock);
2013 ReleaseWriteLock(&avc->lock);
2017 * Call osi_FlushPages in open, read/write, and map, since it
2018 * is too hard here to figure out if we should lock the
2021 if (setLocks && !slowPass)
2022 ObtainReadLock(&avc->lock);
2024 ObtainWriteLock(&avc->lock, 66);
2025 ObtainWriteLock(&tdc->lock, 610);
2030 * avc->lock(R) if setLocks && !slowPass
2031 * avc->lock(W) if !setLocks || slowPass
2035 /* Watch for standard race condition around osi_FlushText */
2036 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2037 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2038 afs_stats_cmperf.dcacheHits++;
2039 ConvertWToSLock(&tdc->lock);
2043 /* Sleep here when cache needs to be drained. */
2044 if (setLocks && !slowPass
2045 && (afs_blocksUsed >
2046 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2047 /* Make sure truncate daemon is running */
2048 afs_MaybeWakeupTruncateDaemon();
2049 ObtainWriteLock(&tdc->tlock, 614);
2050 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2051 ReleaseWriteLock(&tdc->tlock);
2052 ReleaseWriteLock(&tdc->lock);
2053 ReleaseReadLock(&avc->lock);
2054 while ((afs_blocksUsed - afs_blocksDiscarded) >
2055 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2056 afs_WaitForCacheDrain = 1;
2057 afs_osi_Sleep(&afs_WaitForCacheDrain);
2059 afs_MaybeFreeDiscardedDCache();
2060 /* need to check if someone else got the chunk first. */
2061 goto RetryGetDCache;
2064 /* Do not fetch data beyond truncPos. */
2065 maxGoodLength = avc->f.m.Length;
2066 if (avc->f.truncPos < maxGoodLength)
2067 maxGoodLength = avc->f.truncPos;
2068 Position = AFS_CHUNKBASE(abyte);
2069 if (vType(avc) == VDIR) {
2070 size = avc->f.m.Length;
2071 if (size > tdc->f.chunkBytes) {
2072 /* pre-reserve space for file */
2073 afs_AdjustSize(tdc, size);
2075 size = 999999999; /* max size for transfer */
2077 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2078 /* don't read past end of good data on server */
2079 if (Position + size > maxGoodLength)
2080 size = maxGoodLength - Position;
2082 size = 0; /* Handle random races */
2083 if (size > tdc->f.chunkBytes) {
2084 /* pre-reserve space for file */
2085 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2086 /* max size for transfer still in size */
2089 if (afs_mariner && !tdc->f.chunk)
2090 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2092 * Right now, we only have one tool, and it's a hammer. So, we
2093 * fetch the whole file.
2095 DZap(tdc); /* pages in cache may be old */
2096 file = afs_CFileOpen(&tdc->f.inode);
2097 afs_RemoveVCB(&avc->f.fid);
2098 tdc->f.states |= DWriting;
2099 tdc->dflags |= DFFetching;
2100 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2101 if (tdc->mflags & DFFetchReq) {
2102 tdc->mflags &= ~DFFetchReq;
2103 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2104 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2105 __FILE__, ICL_TYPE_INT32, __LINE__,
2106 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2110 (struct afs_FetchOutput *)osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2111 setVcacheStatus = 0;
2114 * Remember if we are doing the reading from a replicated volume,
2115 * and how many times we've zipped around the fetch/analyze loop.
2117 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2119 accP = &(afs_stats_cmfullperf.accessinf);
2121 (accP->replicatedRefs)++;
2123 (accP->unreplicatedRefs)++;
2124 #endif /* AFS_NOSTATS */
2125 /* this is a cache miss */
2126 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2127 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2128 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2131 afs_stats_cmperf.dcacheMisses++;
2134 * Dynamic root support: fetch data from local memory.
2136 if (afs_IsDynroot(avc)) {
2140 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2142 dynrootDir += Position;
2143 dynrootLen -= Position;
2144 if (size > dynrootLen)
2148 code = afs_CFileWrite(file, 0, dynrootDir, size);
2156 tdc->validPos = Position + size;
2157 afs_CFileTruncate(file, size); /* prune it */
2158 } else if (afs_IsDynrootMount(avc)) {
2162 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2164 dynrootDir += Position;
2165 dynrootLen -= Position;
2166 if (size > dynrootLen)
2170 code = afs_CFileWrite(file, 0, dynrootDir, size);
2178 tdc->validPos = Position + size;
2179 afs_CFileTruncate(file, size); /* prune it */
2182 * Not a dynamic vnode: do the real fetch.
2187 * avc->lock(R) if setLocks && !slowPass
2188 * avc->lock(W) if !setLocks || slowPass
2192 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK);
2197 (accP->numReplicasAccessed)++;
2199 #endif /* AFS_NOSTATS */
2200 if (!setLocks || slowPass) {
2201 avc->callback = tc->srvr->server;
2203 newCallback = tc->srvr->server;
2207 code = afs_CacheFetchProc(tc, file, Position, tdc,
2213 /* callback could have been broken (or expired) in a race here,
2214 * but we return the data anyway. It's as good as we knew about
2215 * when we started. */
2217 * validPos is updated by CacheFetchProc, and can only be
2218 * modifed under a dcache write lock, which we've blocked out
2220 size = tdc->validPos - Position; /* actual segment size */
2223 afs_CFileTruncate(file, size); /* prune it */
2225 if (!setLocks || slowPass) {
2226 ObtainWriteLock(&afs_xcbhash, 453);
2227 afs_DequeueCallback(avc);
2228 avc->f.states &= ~(CStatd | CUnique);
2229 avc->callback = NULL;
2230 ReleaseWriteLock(&afs_xcbhash);
2231 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2232 osi_dnlc_purgedp(avc);
2234 /* Something lost. Forget about performance, and go
2235 * back with a vcache write lock.
2237 afs_CFileTruncate(file, 0);
2238 afs_AdjustSize(tdc, 0);
2239 afs_CFileClose(file);
2240 osi_FreeLargeSpace(tsmall);
2242 ReleaseWriteLock(&tdc->lock);
2245 ReleaseReadLock(&avc->lock);
2247 goto RetryGetDCache;
2251 } while (afs_Analyze
2252 (tc, code, &avc->f.fid, areq,
2253 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2257 * avc->lock(R) if setLocks && !slowPass
2258 * avc->lock(W) if !setLocks || slowPass
2264 * In the case of replicated access, jot down info on the number of
2265 * attempts it took before we got through or gave up.
2268 if (numFetchLoops <= 1)
2269 (accP->refFirstReplicaOK)++;
2270 if (numFetchLoops > accP->maxReplicasPerRef)
2271 accP->maxReplicasPerRef = numFetchLoops;
2273 #endif /* AFS_NOSTATS */
2275 tdc->dflags &= ~DFFetching;
2276 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2277 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2278 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2279 tdc, ICL_TYPE_INT32, tdc->dflags);
2280 if (avc->execsOrWriters == 0)
2281 tdc->f.states &= ~DWriting;
2283 /* now, if code != 0, we have an error and should punt.
2284 * note that we have the vcache write lock, either because
2285 * !setLocks or slowPass.
2288 afs_CFileTruncate(file, 0);
2289 afs_AdjustSize(tdc, 0);
2290 afs_CFileClose(file);
2291 ZapDCE(tdc); /* sets DFEntryMod */
2292 if (vType(avc) == VDIR) {
2295 tdc->f.states &= ~(DRO|DBackup|DRW);
2296 afs_DCMoveBucket(tdc, 0, 0);
2297 ReleaseWriteLock(&tdc->lock);
2299 if (!afs_IsDynroot(avc)) {
2300 ObtainWriteLock(&afs_xcbhash, 454);
2301 afs_DequeueCallback(avc);
2302 avc->f.states &= ~(CStatd | CUnique);
2303 ReleaseWriteLock(&afs_xcbhash);
2304 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2305 osi_dnlc_purgedp(avc);
2308 * avc->lock(W); assert(!setLocks || slowPass)
2310 osi_Assert(!setLocks || slowPass);
2316 /* otherwise we copy in the just-fetched info */
2317 afs_CFileClose(file);
2318 afs_AdjustSize(tdc, size); /* new size */
2320 * Copy appropriate fields into vcache. Status is
2321 * copied later where we selectively acquire the
2322 * vcache write lock.
2325 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2327 setVcacheStatus = 1;
2328 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2329 tsmall->OutStatus.DataVersion);
2330 tdc->dflags |= DFEntryMod;
2331 afs_indexFlags[tdc->index] |= IFEverUsed;
2332 ConvertWToSLock(&tdc->lock);
2333 } /*Data version numbers don't match */
2336 * Data version numbers match.
2338 afs_stats_cmperf.dcacheHits++;
2339 } /*Data version numbers match */
2341 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2345 * avc->lock(R) if setLocks && !slowPass
2346 * avc->lock(W) if !setLocks || slowPass
2347 * tdc->lock(S) if tdc
2351 * See if this was a reference to a file in the local cell.
2353 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2354 afs_stats_cmperf.dlocalAccesses++;
2356 afs_stats_cmperf.dremoteAccesses++;
2358 /* Fix up LRU info */
2361 MObtainWriteLock(&afs_xdcache, 602);
2362 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2363 hadd32(afs_indexCounter, 1);
2364 MReleaseWriteLock(&afs_xdcache);
2366 /* return the data */
2367 if (vType(avc) == VDIR)
2370 *aoffset = AFS_CHUNKOFFSET(abyte);
2371 *alen = (tdc->f.chunkBytes - *aoffset);
2372 ReleaseSharedLock(&tdc->lock);
2377 * avc->lock(R) if setLocks && !slowPass
2378 * avc->lock(W) if !setLocks || slowPass
2381 /* Fix up the callback and status values in the vcache */
2383 if (setLocks && !slowPass) {
2386 * This is our dirty little secret to parallel fetches.
2387 * We don't write-lock the vcache while doing the fetch,
2388 * but potentially we'll need to update the vcache after
2389 * the fetch is done.
2391 * Drop the read lock and try to re-obtain the write
2392 * lock. If the vcache still has the same DV, it's
2393 * ok to go ahead and install the new data.
2395 afs_hyper_t currentDV, statusDV;
2397 hset(currentDV, avc->f.m.DataVersion);
2399 if (setNewCallback && avc->callback != newCallback)
2403 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2404 tsmall->OutStatus.DataVersion);
2406 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2408 if (setVcacheStatus && !hsame(currentDV, statusDV))
2412 ReleaseReadLock(&avc->lock);
2414 if (doVcacheUpdate) {
2415 ObtainWriteLock(&avc->lock, 615);
2416 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2417 /* We lose. Someone will beat us to it. */
2419 ReleaseWriteLock(&avc->lock);
2424 /* With slow pass, we've already done all the updates */
2426 ReleaseWriteLock(&avc->lock);
2429 /* Check if we need to perform any last-minute fixes with a write-lock */
2430 if (!setLocks || doVcacheUpdate) {
2432 avc->callback = newCallback;
2433 if (tsmall && setVcacheStatus)
2434 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2436 ReleaseWriteLock(&avc->lock);
2440 osi_FreeLargeSpace(tsmall);
2443 } /*afs_GetDCache */
2447 * afs_WriteThroughDSlots
2450 * Sweep through the dcache slots and write out any modified
2451 * in-memory data back on to our caching store.
2457 * The afs_xdcache is write-locked through this whole affair.
2460 afs_WriteThroughDSlots(void)
2462 register struct dcache *tdc;
2463 register afs_int32 i, touchedit = 0;
2465 struct afs_q DirtyQ, *tq;
2467 AFS_STATCNT(afs_WriteThroughDSlots);
2470 * Because of lock ordering, we can't grab dcache locks while
2471 * holding afs_xdcache. So we enter xdcache, get a reference
2472 * for every dcache entry, and exit xdcache.
2474 MObtainWriteLock(&afs_xdcache, 283);
2476 for (i = 0; i < afs_cacheFiles; i++) {
2477 tdc = afs_indexTable[i];
2479 /* Grab tlock in case the existing refcount isn't zero */
2480 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2481 ObtainWriteLock(&tdc->tlock, 623);
2483 ReleaseWriteLock(&tdc->tlock);
2485 QAdd(&DirtyQ, &tdc->dirty);
2488 MReleaseWriteLock(&afs_xdcache);
2491 * Now, for each dcache entry we found, check if it's dirty.
2492 * If so, get write-lock, get afs_xdcache, which protects
2493 * afs_cacheInodep, and flush it. Don't forget to put back
2497 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2499 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2501 if (tdc->dflags & DFEntryMod) {
2504 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2506 /* Now that we have the write lock, double-check */
2507 if (wrLock && (tdc->dflags & DFEntryMod)) {
2508 tdc->dflags &= ~DFEntryMod;
2509 MObtainWriteLock(&afs_xdcache, 620);
2510 afs_WriteDCache(tdc, 1);
2511 MReleaseWriteLock(&afs_xdcache);
2515 ReleaseWriteLock(&tdc->lock);
2521 MObtainWriteLock(&afs_xdcache, 617);
2522 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2523 /* Touch the file to make sure that the mtime on the file is kept
2524 * up-to-date to avoid losing cached files on cold starts because
2525 * their mtime seems old...
2527 struct afs_fheader theader;
2529 theader.magic = AFS_FHMAGIC;
2530 theader.firstCSize = AFS_FIRSTCSIZE;
2531 theader.otherCSize = AFS_OTHERCSIZE;
2532 theader.version = AFS_CI_VERSION;
2533 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2535 MReleaseWriteLock(&afs_xdcache);
2542 * Return a pointer to an freshly initialized dcache entry using
2543 * a memory-based cache. The tlock will be read-locked.
2546 * aslot : Dcache slot to look at.
2547 * tmpdc : Ptr to dcache entry.
2550 * Must be called with afs_xdcache write-locked.
2554 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2556 register struct dcache *tdc;
2559 AFS_STATCNT(afs_MemGetDSlot);
2560 if (CheckLock(&afs_xdcache) != -1)
2561 osi_Panic("getdslot nolock");
2562 if (aslot < 0 || aslot >= afs_cacheFiles)
2563 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2564 tdc = afs_indexTable[aslot];
2566 QRemove(&tdc->lruq); /* move to queue head */
2567 QAdd(&afs_DLRU, &tdc->lruq);
2568 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2569 ObtainWriteLock(&tdc->tlock, 624);
2571 ConvertWToRLock(&tdc->tlock);
2574 if (tmpdc == NULL) {
2575 if (!afs_freeDSList)
2576 afs_GetDownDSlot(4);
2577 if (!afs_freeDSList) {
2578 /* none free, making one is better than a panic */
2579 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2580 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2581 #ifdef KERNEL_HAVE_PIN
2582 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2585 tdc = afs_freeDSList;
2586 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2589 tdc->dflags = 0; /* up-to-date, not in free q */
2591 QAdd(&afs_DLRU, &tdc->lruq);
2592 if (tdc->lruq.prev == &tdc->lruq)
2593 osi_Panic("lruq 3");
2599 /* initialize entry */
2600 tdc->f.fid.Cell = 0;
2601 tdc->f.fid.Fid.Volume = 0;
2603 hones(tdc->f.versionNo);
2604 tdc->f.inode.mem = aslot;
2605 tdc->dflags |= DFEntryMod;
2608 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2611 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2612 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2613 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2616 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2617 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2618 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2619 ObtainReadLock(&tdc->tlock);
2622 afs_indexTable[aslot] = tdc;
2625 } /*afs_MemGetDSlot */
2627 unsigned int last_error = 0, lasterrtime = 0;
2633 * Return a pointer to an freshly initialized dcache entry using
2634 * a UFS-based disk cache. The dcache tlock will be read-locked.
2637 * aslot : Dcache slot to look at.
2638 * tmpdc : Ptr to dcache entry.
2641 * afs_xdcache lock write-locked.
2644 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2646 register afs_int32 code;
2647 register struct dcache *tdc;
2651 AFS_STATCNT(afs_UFSGetDSlot);
2652 if (CheckLock(&afs_xdcache) != -1)
2653 osi_Panic("getdslot nolock");
2654 if (aslot < 0 || aslot >= afs_cacheFiles)
2655 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2656 tdc = afs_indexTable[aslot];
2658 QRemove(&tdc->lruq); /* move to queue head */
2659 QAdd(&afs_DLRU, &tdc->lruq);
2660 /* Grab tlock in case refCount != 0 */
2661 ObtainWriteLock(&tdc->tlock, 625);
2663 ConvertWToRLock(&tdc->tlock);
2666 /* otherwise we should read it in from the cache file */
2668 * If we weren't passed an in-memory region to place the file info,
2669 * we have to allocate one.
2671 if (tmpdc == NULL) {
2672 if (!afs_freeDSList)
2673 afs_GetDownDSlot(4);
2674 if (!afs_freeDSList) {
2675 /* none free, making one is better than a panic */
2676 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2677 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2678 #ifdef KERNEL_HAVE_PIN
2679 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2682 tdc = afs_freeDSList;
2683 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2686 tdc->dflags = 0; /* up-to-date, not in free q */
2688 QAdd(&afs_DLRU, &tdc->lruq);
2689 if (tdc->lruq.prev == &tdc->lruq)
2690 osi_Panic("lruq 3");
2697 * Seek to the aslot'th entry and read it in.
2700 afs_osi_Read(afs_cacheInodep,
2701 sizeof(struct fcache) * aslot +
2702 sizeof(struct afs_fheader), (char *)(&tdc->f),
2703 sizeof(struct fcache));
2705 if (code != sizeof(struct fcache))
2707 if (!afs_CellNumValid(tdc->f.fid.Cell))
2711 tdc->f.fid.Cell = 0;
2712 tdc->f.fid.Fid.Volume = 0;
2714 hones(tdc->f.versionNo);
2715 tdc->dflags |= DFEntryMod;
2716 #if defined(KERNEL_HAVE_UERROR)
2717 last_error = getuerror();
2719 lasterrtime = osi_Time();
2720 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2721 tdc->f.states &= ~(DRO|DBackup|DRW);
2722 afs_DCMoveBucket(tdc, 0, 0);
2725 if (tdc->f.states & DRO) {
2726 afs_DCMoveBucket(tdc, 0, 2);
2727 } else if (tdc->f.states & DBackup) {
2728 afs_DCMoveBucket(tdc, 0, 1);
2730 afs_DCMoveBucket(tdc, 0, 1);
2736 if (tdc->f.chunk >= 0)
2737 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2742 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2743 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2744 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2747 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2748 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2749 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2750 ObtainReadLock(&tdc->tlock);
2753 * If we didn't read into a temporary dcache region, update the
2754 * slot pointer table.
2757 afs_indexTable[aslot] = tdc;
2760 } /*afs_UFSGetDSlot */
2765 * Write a particular dcache entry back to its home in the
2768 * \param adc Pointer to the dcache entry to write.
2769 * \param atime If true, set the modtime on the file to the current time.
2771 * \note Environment:
2772 * Must be called with the afs_xdcache lock at least read-locked,
2773 * and dcache entry at least read-locked.
2774 * The reference count is not changed.
2778 afs_WriteDCache(register struct dcache *adc, int atime)
2780 register afs_int32 code;
2782 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2784 AFS_STATCNT(afs_WriteDCache);
2785 osi_Assert(WriteLocked(&afs_xdcache));
2787 adc->f.modTime = osi_Time();
2789 * Seek to the right dcache slot and write the in-memory image out to disk.
2791 afs_cellname_write();
2793 afs_osi_Write(afs_cacheInodep,
2794 sizeof(struct fcache) * adc->index +
2795 sizeof(struct afs_fheader), (char *)(&adc->f),
2796 sizeof(struct fcache));
2797 if (code != sizeof(struct fcache))
2805 * Wake up users of a particular file waiting for stores to take
2808 * \param avc Ptr to related vcache entry.
2810 * \note Environment:
2811 * Nothing interesting.
2814 afs_wakeup(register struct vcache *avc)
2817 register struct brequest *tb;
2819 AFS_STATCNT(afs_wakeup);
2820 for (i = 0; i < NBRS; i++, tb++) {
2821 /* if request is valid and for this file, we've found it */
2822 if (tb->refCount > 0 && avc == tb->vc) {
2825 * If CSafeStore is on, then we don't awaken the guy
2826 * waiting for the store until the whole store has finished.
2827 * Otherwise, we do it now. Note that if CSafeStore is on,
2828 * the BStore routine actually wakes up the user, instead
2830 * I think this is redundant now because this sort of thing
2831 * is already being handled by the higher-level code.
2833 if ((avc->f.states & CSafeStore) == 0) {
2835 tb->flags |= BUVALID;
2836 if (tb->flags & BUWAIT) {
2837 tb->flags &= ~BUWAIT;
2849 * Given a file name and inode, set up that file to be an
2850 * active member in the AFS cache. This also involves checking
2851 * the usability of its data.
2853 * \param afile Name of the cache file to initialize.
2854 * \param ainode Inode of the file.
2856 * \note Environment:
2857 * This function is called only during initialization.
2860 afs_InitCacheFile(char *afile, ino_t ainode)
2862 register afs_int32 code;
2865 struct osi_file *tfile;
2866 struct osi_stat tstat;
2867 register struct dcache *tdc;
2869 AFS_STATCNT(afs_InitCacheFile);
2870 index = afs_stats_cmperf.cacheNumEntries;
2871 if (index >= afs_cacheFiles)
2874 MObtainWriteLock(&afs_xdcache, 282);
2875 tdc = afs_GetDSlot(index, NULL);
2876 ReleaseReadLock(&tdc->tlock);
2877 MReleaseWriteLock(&afs_xdcache);
2879 ObtainWriteLock(&tdc->lock, 621);
2880 MObtainWriteLock(&afs_xdcache, 622);
2882 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
2884 ReleaseWriteLock(&afs_xdcache);
2885 ReleaseWriteLock(&tdc->lock);
2890 /* Add any other 'complex' inode types here ... */
2891 #if defined(UKERNEL) || !defined(LINUX_USE_FH)
2892 tdc->f.inode.ufs = ainode;
2894 osi_Panic("Can't init cache with inode numbers when complex inodes are "
2899 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
2901 tfile = osi_UFSOpen(&tdc->f.inode);
2902 code = afs_osi_Stat(tfile, &tstat);
2904 osi_Panic("initcachefile stat");
2907 * If file size doesn't match the cache info file, it's probably bad.
2909 if (tdc->f.chunkBytes != tstat.size)
2911 tdc->f.chunkBytes = 0;
2914 * If file changed within T (120?) seconds of cache info file, it's
2915 * probably bad. In addition, if slot changed within last T seconds,
2916 * the cache info file may be incorrectly identified, and so slot
2919 if (cacheInfoModTime < tstat.mtime + 120)
2921 if (cacheInfoModTime < tdc->f.modTime + 120)
2923 /* In case write through is behind, make sure cache items entry is
2924 * at least as new as the chunk.
2926 if (tdc->f.modTime < tstat.mtime)
2929 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
2930 if (tstat.size != 0)
2931 osi_UFSTruncate(tfile, 0);
2932 tdc->f.states &= ~(DRO|DBackup|DRW);
2933 afs_DCMoveBucket(tdc, 0, 0);
2934 /* put entry in free cache slot list */
2935 afs_dvnextTbl[tdc->index] = afs_freeDCList;
2936 afs_freeDCList = index;
2938 afs_indexFlags[index] |= IFFree;
2939 afs_indexUnique[index] = 0;
2942 * We must put this entry in the appropriate hash tables.
2943 * Note that i is still set from the above DCHash call
2945 code = DCHash(&tdc->f.fid, tdc->f.chunk);
2946 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
2947 afs_dchashTbl[code] = tdc->index;
2948 code = DVHash(&tdc->f.fid);
2949 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
2950 afs_dvhashTbl[code] = tdc->index;
2951 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
2953 /* has nontrivial amt of data */
2954 afs_indexFlags[index] |= IFEverUsed;
2955 afs_stats_cmperf.cacheFilesReused++;
2957 * Initialize index times to file's mod times; init indexCounter
2960 hset32(afs_indexTimes[index], tstat.atime);
2961 if (hgetlo(afs_indexCounter) < tstat.atime) {
2962 hset32(afs_indexCounter, tstat.atime);
2964 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
2965 } /*File is not bad */
2967 osi_UFSClose(tfile);
2968 tdc->f.states &= ~DWriting;
2969 tdc->dflags &= ~DFEntryMod;
2970 /* don't set f.modTime; we're just cleaning up */
2971 afs_WriteDCache(tdc, 0);
2972 ReleaseWriteLock(&afs_xdcache);
2973 ReleaseWriteLock(&tdc->lock);
2975 afs_stats_cmperf.cacheNumEntries++;
2980 /*Max # of struct dcache's resident at any time*/
2982 * If 'dchint' is enabled then in-memory dcache min is increased because of
2988 * Initialize dcache related variables.
2998 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3000 register struct dcache *tdp;
3004 afs_freeDCList = NULLIDX;
3005 afs_discardDCList = NULLIDX;
3006 afs_freeDCCount = 0;
3007 afs_freeDSList = NULL;
3008 hzero(afs_indexCounter);
3010 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3016 if (achunk < 0 || achunk > 30)
3017 achunk = 13; /* Use default */
3018 AFS_SETCHUNKSIZE(achunk);
3024 if (aflags & AFSCALL_INIT_MEMCACHE) {
3026 * Use a memory cache instead of a disk cache
3028 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3029 afs_cacheType = &afs_MemCacheOps;
3030 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3031 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3032 /* ablocks is reported in 1K blocks */
3033 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3035 printf("afsd: memory cache too large for available memory.\n");
3036 printf("afsd: AFS files cannot be accessed.\n\n");
3038 afiles = ablocks = 0;
3040 printf("Memory cache: Allocating %d dcache entries...",
3043 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3044 afs_cacheType = &afs_UfsCacheOps;
3047 if (aDentries > 512)
3048 afs_dhashsize = 2048;
3049 /* initialize hash tables */
3051 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3053 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3054 for (i = 0; i < afs_dhashsize; i++) {
3055 afs_dvhashTbl[i] = NULLIDX;
3056 afs_dchashTbl[i] = NULLIDX;
3058 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3059 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3060 for (i = 0; i < afiles; i++) {
3061 afs_dvnextTbl[i] = NULLIDX;
3062 afs_dcnextTbl[i] = NULLIDX;
3065 /* Allocate and zero the pointer array to the dcache entries */
3066 afs_indexTable = (struct dcache **)
3067 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3068 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3070 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3071 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3073 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3074 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3075 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3076 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3078 /* Allocate and thread the struct dcache entries themselves */
3079 tdp = afs_Initial_freeDSList =
3080 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3081 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3082 #ifdef KERNEL_HAVE_PIN
3083 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3084 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3085 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3086 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3087 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3088 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3089 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3090 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3091 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3094 afs_freeDSList = &tdp[0];
3095 for (i = 0; i < aDentries - 1; i++) {
3096 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3097 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3098 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3099 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3101 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3102 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3103 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3104 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3106 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3107 afs_cacheBlocks = ablocks;
3108 afs_ComputeCacheParms(); /* compute parms based on cache size */
3110 afs_dcentries = aDentries;
3112 afs_stats_cmperf.cacheBucket0_Discarded =
3113 afs_stats_cmperf.cacheBucket1_Discarded =
3114 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3120 * Shuts down the cache.
3124 shutdown_dcache(void)
3128 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3129 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3130 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3131 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3132 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3133 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3134 afs_osi_Free(afs_Initial_freeDSList,
3135 afs_dcentries * sizeof(struct dcache));
3136 #ifdef KERNEL_HAVE_PIN
3137 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3138 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3139 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3140 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3141 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3142 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3143 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3147 for (i = 0; i < afs_dhashsize; i++) {
3148 afs_dvhashTbl[i] = NULLIDX;
3149 afs_dchashTbl[i] = NULLIDX;
3152 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3153 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3155 afs_blocksUsed = afs_dcentries = 0;
3156 afs_stats_cmperf.cacheBucket0_Discarded =
3157 afs_stats_cmperf.cacheBucket1_Discarded =
3158 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3159 hzero(afs_indexCounter);
3161 afs_freeDCCount = 0;
3162 afs_freeDCList = NULLIDX;
3163 afs_discardDCList = NULLIDX;
3164 afs_freeDSList = afs_Initial_freeDSList = 0;
3166 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3172 * Get a dcache ready for writing, respecting the current cache size limits
3174 * len is required because afs_GetDCache with flag == 4 expects the length
3175 * field to be filled. It decides from this whether it's necessary to fetch
3176 * data into the chunk before writing or not (when the whole chunk is
3179 * \param avc The vcache to fetch a dcache for
3180 * \param filePos The start of the section to be written
3181 * \param len The length of the section to be written
3185 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3186 * must be released and afs_PutDCache() called to free dcache.
3189 * \note avc->lock must be held on entry. Function may release and reobtain
3190 * avc->lock and GLOCK.
3194 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3195 afs_size_t len, struct vrequest *areq,
3197 struct dcache *tdc = NULL;
3200 /* read the cached info */
3202 tdc = afs_FindDCache(avc, filePos);
3204 ObtainWriteLock(&tdc->lock, 657);
3205 } else if (afs_blocksUsed >
3206 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3207 tdc = afs_FindDCache(avc, filePos);
3209 ObtainWriteLock(&tdc->lock, 658);
3210 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3211 || (tdc->dflags & DFFetching)) {
3212 ReleaseWriteLock(&tdc->lock);
3218 afs_MaybeWakeupTruncateDaemon();
3219 while (afs_blocksUsed >
3220 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3221 ReleaseWriteLock(&avc->lock);
3222 if (afs_blocksUsed - afs_blocksDiscarded >
3223 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3224 afs_WaitForCacheDrain = 1;
3225 afs_osi_Sleep(&afs_WaitForCacheDrain);
3227 afs_MaybeFreeDiscardedDCache();
3228 afs_MaybeWakeupTruncateDaemon();
3229 ObtainWriteLock(&avc->lock, 509);
3231 avc->f.states |= CDirty;
3232 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3234 ObtainWriteLock(&tdc->lock, 659);
3237 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3239 ObtainWriteLock(&tdc->lock, 660);
3242 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3243 afs_stats_cmperf.cacheCurrDirtyChunks++;
3244 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3246 if (!(tdc->f.states & DWriting)) {
3247 /* don't mark entry as mod if we don't have to */
3248 tdc->f.states |= DWriting;
3249 tdc->dflags |= DFEntryMod;
3255 #if defined(AFS_DISCON_ENV)
3258 * Make a shadow copy of a dir's dcache. It's used for disconnected
3259 * operations like remove/create/rename to keep the original directory data.
3260 * On reconnection, we can diff the original data with the server and get the
3261 * server changes and with the local data to get the local changes.
3263 * \param avc The dir vnode.
3264 * \param adc The dir dcache.
3266 * \return 0 for success.
3268 * \note The vcache entry must be write locked.
3269 * \note The dcache entry must be read locked.
3271 int afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3273 int i, code, ret_code = 0, written, trans_size;
3274 struct dcache *new_dc = NULL;
3275 struct osi_file *tfile_src, *tfile_dst;
3276 struct VenusFid shadow_fid;
3279 /* Is this a dir? */
3280 if (vType(avc) != VDIR)
3283 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3286 /* Generate a fid for the shadow dir. */
3287 shadow_fid.Cell = avc->f.fid.Cell;
3288 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3289 afs_GenShadowFid(&shadow_fid);
3291 ObtainWriteLock(&afs_xdcache, 716);
3293 /* Get a fresh dcache. */
3294 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3296 ObtainReadLock(&adc->mflock);
3298 /* Set up the new fid. */
3299 /* Copy interesting data from original dir dcache. */
3300 new_dc->mflags = adc->mflags;
3301 new_dc->dflags = adc->dflags;
3302 new_dc->f.modTime = adc->f.modTime;
3303 new_dc->f.versionNo = adc->f.versionNo;
3304 new_dc->f.states = adc->f.states;
3305 new_dc->f.chunk= adc->f.chunk;
3306 new_dc->f.chunkBytes = adc->f.chunkBytes;
3308 ReleaseReadLock(&adc->mflock);
3310 /* Now add to the two hash chains */
3311 i = DCHash(&shadow_fid, 0);
3312 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3313 afs_dchashTbl[i] = new_dc->index;
3315 i = DVHash(&shadow_fid);
3316 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3317 afs_dvhashTbl[i] = new_dc->index;
3319 ReleaseWriteLock(&afs_xdcache);
3321 /* Alloc a 4k block. */
3322 data = (char *) afs_osi_Alloc(4096);
3324 printf("afs_MakeShadowDir: could not alloc data\n");
3329 /* Open the files. */
3330 tfile_src = afs_CFileOpen(&adc->f.inode);
3331 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3333 /* And now copy dir dcache data into this dcache,
3337 while (written < adc->f.chunkBytes) {
3338 trans_size = adc->f.chunkBytes - written;
3339 if (trans_size > 4096)
3342 /* Read a chunk from the dcache. */
3343 code = afs_CFileRead(tfile_src, written, data, trans_size);
3344 if (code < trans_size) {
3349 /* Write it to the new dcache. */
3350 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3351 if (code < trans_size) {
3356 written+=trans_size;
3359 afs_CFileClose(tfile_dst);
3360 afs_CFileClose(tfile_src);
3362 afs_osi_Free(data, 4096);
3364 ReleaseWriteLock(&new_dc->lock);
3365 afs_PutDCache(new_dc);
3368 ObtainWriteLock(&afs_xvcache, 763);
3369 ObtainWriteLock(&afs_disconDirtyLock, 765);
3370 QAdd(&afs_disconShadow, &avc->shadowq);
3372 ReleaseWriteLock(&afs_disconDirtyLock);
3373 ReleaseWriteLock(&afs_xvcache);
3375 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3376 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3384 * Delete the dcaches of a shadow dir.
3386 * \param avc The vcache containing the shadow fid.
3388 * \note avc must be write locked.
3390 void afs_DeleteShadowDir(struct vcache *avc)
3393 struct VenusFid shadow_fid;
3395 shadow_fid.Cell = avc->f.fid.Cell;
3396 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3397 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3398 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3400 tdc = afs_FindDCacheByFid(&shadow_fid);
3402 afs_HashOutDCache(tdc, 1);
3403 afs_DiscardDCache(tdc);
3406 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3407 ObtainWriteLock(&afs_disconDirtyLock, 708);
3408 QRemove(&avc->shadowq);
3409 ReleaseWriteLock(&afs_disconDirtyLock);
3410 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3414 * Populate a dcache with empty chunks up to a given file size,
3415 * used before extending a file in order to avoid 'holes' which
3416 * we can't access in disconnected mode.
3418 * \param avc The vcache which is being extended (locked)
3419 * \param alen The new length of the file
3422 void afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq) {
3424 afs_size_t len, offset;
3425 afs_int32 start, end;
3427 /* We're doing this to deal with the situation where we extend
3428 * by writing after lseek()ing past the end of the file . If that
3429 * extension skips chunks, then those chunks won't be created, and
3430 * GetDCache will assume that they have to be fetched from the server.
3431 * So, for each chunk between the current file position, and the new
3432 * length we GetDCache for that chunk.
3435 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3438 if (avc->f.m.Length == 0)
3441 start = AFS_CHUNK(avc->f.m.Length)+1;
3443 end = AFS_CHUNK(apos);
3446 len = AFS_CHUNKTOSIZE(start);
3447 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);