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;
1631 struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
1632 osi_timeval_t xferStartTime, /*FS xfer start time */
1633 xferStopTime; /*FS xfer stop time */
1634 afs_size_t bytesToXfer; /* # bytes to xfer */
1635 afs_size_t bytesXferred; /* # bytes actually xferred */
1636 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1637 int fromReplica; /*Are we reading from a replica? */
1638 int numFetchLoops; /*# times around the fetch/analyze loop */
1639 #endif /* AFS_NOSTATS */
1641 AFS_STATCNT(afs_GetDCache);
1645 setLocks = aflags & 1;
1648 * Determine the chunk number and offset within the chunk corresponding
1649 * to the desired byte.
1651 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1654 chunk = AFS_CHUNK(abyte);
1657 /* come back to here if we waited for the cache to drain. */
1660 setNewCallback = setVcacheStatus = 0;
1664 ObtainWriteLock(&avc->lock, 616);
1666 ObtainReadLock(&avc->lock);
1671 * avc->lock(R) if setLocks && !slowPass
1672 * avc->lock(W) if !setLocks || slowPass
1677 /* check hints first! (might could use bcmp or some such...) */
1678 if ((tdc = avc->dchint)) {
1682 * The locking order between afs_xdcache and dcache lock matters.
1683 * The hint dcache entry could be anywhere, even on the free list.
1684 * Locking afs_xdcache ensures that noone is trying to pull dcache
1685 * entries from the free list, and thereby assuming them to be not
1686 * referenced and not locked.
1688 MObtainReadLock(&afs_xdcache);
1689 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1691 if (dcLocked && (tdc->index != NULLIDX)
1692 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1693 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1694 /* got the right one. It might not be the right version, and it
1695 * might be fetching, but it's the right dcache entry.
1697 /* All this code should be integrated better with what follows:
1698 * I can save a good bit more time under a write lock if I do..
1700 ObtainWriteLock(&tdc->tlock, 603);
1702 ReleaseWriteLock(&tdc->tlock);
1704 MReleaseReadLock(&afs_xdcache);
1707 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1708 && !(tdc->dflags & DFFetching)) {
1710 afs_stats_cmperf.dcacheHits++;
1711 MObtainWriteLock(&afs_xdcache, 559);
1712 QRemove(&tdc->lruq);
1713 QAdd(&afs_DLRU, &tdc->lruq);
1714 MReleaseWriteLock(&afs_xdcache);
1717 * avc->lock(R) if setLocks && !slowPass
1718 * avc->lock(W) if !setLocks || slowPass
1725 ReleaseSharedLock(&tdc->lock);
1726 MReleaseReadLock(&afs_xdcache);
1734 * avc->lock(R) if setLocks && !slowPass
1735 * avc->lock(W) if !setLocks || slowPass
1736 * tdc->lock(S) if tdc
1739 if (!tdc) { /* If the hint wasn't the right dcache entry */
1741 * Hash on the [fid, chunk] and get the corresponding dcache index
1742 * after write-locking the dcache.
1747 * avc->lock(R) if setLocks && !slowPass
1748 * avc->lock(W) if !setLocks || slowPass
1751 i = DCHash(&avc->f.fid, chunk);
1752 /* check to make sure our space is fine */
1753 afs_MaybeWakeupTruncateDaemon();
1755 MObtainWriteLock(&afs_xdcache, 280);
1757 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1758 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1759 tdc = afs_GetDSlot(index, NULL);
1760 ReleaseReadLock(&tdc->tlock);
1763 * avc->lock(R) if setLocks && !slowPass
1764 * avc->lock(W) if !setLocks || slowPass
1767 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1768 /* Move it up in the beginning of the list */
1769 if (afs_dchashTbl[i] != index) {
1770 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1771 afs_dcnextTbl[index] = afs_dchashTbl[i];
1772 afs_dchashTbl[i] = index;
1774 MReleaseWriteLock(&afs_xdcache);
1775 ObtainSharedLock(&tdc->lock, 606);
1776 break; /* leaving refCount high for caller */
1782 index = afs_dcnextTbl[index];
1786 * If we didn't find the entry, we'll create one.
1788 if (index == NULLIDX) {
1791 * avc->lock(R) if setLocks
1792 * avc->lock(W) if !setLocks
1795 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1796 avc, ICL_TYPE_INT32, chunk);
1798 /* Make sure there is a free dcache entry for us to use */
1799 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1802 avc->f.states |= CDCLock;
1803 /* just need slots */
1804 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1806 avc->f.states &= ~CDCLock;
1807 if (afs_discardDCList != NULLIDX
1808 || afs_freeDCList != NULLIDX)
1810 /* If we can't get space for 5 mins we give up and panic */
1811 if (++downDCount > 300) {
1812 #if defined(AFS_CACHE_BYPASS)
1813 afs_warn("GetDCache calling osi_Panic: No space in five minutes.\n downDCount: %d\n aoffset: %d alen: %d\n", downDCount, aoffset, alen);
1815 osi_Panic("getdcache");
1817 MReleaseWriteLock(&afs_xdcache);
1820 * avc->lock(R) if setLocks
1821 * avc->lock(W) if !setLocks
1823 afs_osi_Wait(1000, 0, 0);
1828 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1831 * Now add to the two hash chains - note that i is still set
1832 * from the above DCHash call.
1834 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1835 afs_dchashTbl[i] = tdc->index;
1836 i = DVHash(&avc->f.fid);
1837 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1838 afs_dvhashTbl[i] = tdc->index;
1839 tdc->dflags = DFEntryMod;
1841 afs_MaybeWakeupTruncateDaemon();
1842 MReleaseWriteLock(&afs_xdcache);
1843 ConvertWToSLock(&tdc->lock);
1848 /* vcache->dcache hint failed */
1851 * avc->lock(R) if setLocks && !slowPass
1852 * avc->lock(W) if !setLocks || slowPass
1855 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1856 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1857 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1858 hgetlo(avc->f.m.DataVersion));
1860 * Here we have the entry in tdc, with its refCount incremented.
1861 * Note: we don't use the S-lock on avc; it costs concurrency when
1862 * storing a file back to the server.
1866 * Not a newly created file so we need to check the file's length and
1867 * compare data versions since someone could have changed the data or we're
1868 * reading a file written elsewhere. We only want to bypass doing no-op
1869 * read rpcs on newly created files (dv of 0) since only then we guarantee
1870 * that this chunk's data hasn't been filled by another client.
1872 size = AFS_CHUNKSIZE(abyte);
1873 if (aflags & 4) /* called from write */
1875 else /* called from read */
1876 tlen = tdc->validPos - abyte;
1877 Position = AFS_CHUNKTOBASE(chunk);
1878 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1879 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1880 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1881 ICL_HANDLE_OFFSET(Position));
1882 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
1884 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
1885 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
1886 overWriteWholeChunk = 1;
1887 if (doAdjustSize || overWriteWholeChunk) {
1888 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1890 #ifdef AFS_SGI64_ENV
1893 #else /* AFS_SGI64_ENV */
1896 #endif /* AFS_SGI64_ENV */
1897 #else /* AFS_SGI_ENV */
1900 #endif /* AFS_SGI_ENV */
1901 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
1902 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1903 #if defined(AFS_SUN5_ENV) || defined(AFS_OSF_ENV)
1904 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
1906 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
1908 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1909 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
1910 doReallyAdjustSize = 1;
1912 if (doReallyAdjustSize || overWriteWholeChunk) {
1913 /* no data in file to read at this position */
1914 UpgradeSToWLock(&tdc->lock, 607);
1915 file = afs_CFileOpen(&tdc->f.inode);
1916 afs_CFileTruncate(file, 0);
1917 afs_CFileClose(file);
1918 afs_AdjustSize(tdc, 0);
1919 hset(tdc->f.versionNo, avc->f.m.DataVersion);
1920 tdc->dflags |= DFEntryMod;
1922 ConvertWToSLock(&tdc->lock);
1927 * We must read in the whole chunk if the version number doesn't
1931 /* don't need data, just a unique dcache entry */
1932 ObtainWriteLock(&afs_xdcache, 608);
1933 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1934 hadd32(afs_indexCounter, 1);
1935 ReleaseWriteLock(&afs_xdcache);
1937 updateV2DC(setLocks, avc, tdc, 553);
1938 if (vType(avc) == VDIR)
1941 *aoffset = AFS_CHUNKOFFSET(abyte);
1942 if (tdc->validPos < abyte)
1943 *alen = (afs_size_t) 0;
1945 *alen = tdc->validPos - abyte;
1946 ReleaseSharedLock(&tdc->lock);
1949 ReleaseWriteLock(&avc->lock);
1951 ReleaseReadLock(&avc->lock);
1953 return tdc; /* check if we're done */
1958 * avc->lock(R) if setLocks && !slowPass
1959 * avc->lock(W) if !setLocks || slowPass
1962 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
1964 setNewCallback = setVcacheStatus = 0;
1968 * avc->lock(R) if setLocks && !slowPass
1969 * avc->lock(W) if !setLocks || slowPass
1972 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
1974 * Version number mismatch.
1977 * If we are disconnected, then we can't do much of anything
1978 * because the data doesn't match the file.
1980 if (AFS_IS_DISCONNECTED) {
1981 ReleaseSharedLock(&tdc->lock);
1984 ReleaseWriteLock(&avc->lock);
1986 ReleaseReadLock(&avc->lock);
1988 /* Flush the Dcache */
1993 UpgradeSToWLock(&tdc->lock, 609);
1996 * If data ever existed for this vnode, and this is a text object,
1997 * do some clearing. Now, you'd think you need only do the flush
1998 * when VTEXT is on, but VTEXT is turned off when the text object
1999 * is freed, while pages are left lying around in memory marked
2000 * with this vnode. If we would reactivate (create a new text
2001 * object from) this vnode, we could easily stumble upon some of
2002 * these old pages in pagein. So, we always flush these guys.
2003 * Sun has a wonderful lack of useful invariants in this system.
2005 * avc->flushDV is the data version # of the file at the last text
2006 * flush. Clearly, at least, we don't have to flush the file more
2007 * often than it changes
2009 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2011 * By here, the cache entry is always write-locked. We can
2012 * deadlock if we call osi_Flush with the cache entry locked...
2013 * Unlock the dcache too.
2015 ReleaseWriteLock(&tdc->lock);
2016 if (setLocks && !slowPass)
2017 ReleaseReadLock(&avc->lock);
2019 ReleaseWriteLock(&avc->lock);
2023 * Call osi_FlushPages in open, read/write, and map, since it
2024 * is too hard here to figure out if we should lock the
2027 if (setLocks && !slowPass)
2028 ObtainReadLock(&avc->lock);
2030 ObtainWriteLock(&avc->lock, 66);
2031 ObtainWriteLock(&tdc->lock, 610);
2036 * avc->lock(R) if setLocks && !slowPass
2037 * avc->lock(W) if !setLocks || slowPass
2041 /* Watch for standard race condition around osi_FlushText */
2042 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2043 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2044 afs_stats_cmperf.dcacheHits++;
2045 ConvertWToSLock(&tdc->lock);
2049 /* Sleep here when cache needs to be drained. */
2050 if (setLocks && !slowPass
2051 && (afs_blocksUsed >
2052 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2053 /* Make sure truncate daemon is running */
2054 afs_MaybeWakeupTruncateDaemon();
2055 ObtainWriteLock(&tdc->tlock, 614);
2056 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2057 ReleaseWriteLock(&tdc->tlock);
2058 ReleaseWriteLock(&tdc->lock);
2059 ReleaseReadLock(&avc->lock);
2060 while ((afs_blocksUsed - afs_blocksDiscarded) >
2061 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2062 afs_WaitForCacheDrain = 1;
2063 afs_osi_Sleep(&afs_WaitForCacheDrain);
2065 afs_MaybeFreeDiscardedDCache();
2066 /* need to check if someone else got the chunk first. */
2067 goto RetryGetDCache;
2070 /* Do not fetch data beyond truncPos. */
2071 maxGoodLength = avc->f.m.Length;
2072 if (avc->f.truncPos < maxGoodLength)
2073 maxGoodLength = avc->f.truncPos;
2074 Position = AFS_CHUNKBASE(abyte);
2075 if (vType(avc) == VDIR) {
2076 size = avc->f.m.Length;
2077 if (size > tdc->f.chunkBytes) {
2078 /* pre-reserve space for file */
2079 afs_AdjustSize(tdc, size);
2081 size = 999999999; /* max size for transfer */
2083 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2084 /* don't read past end of good data on server */
2085 if (Position + size > maxGoodLength)
2086 size = maxGoodLength - Position;
2088 size = 0; /* Handle random races */
2089 if (size > tdc->f.chunkBytes) {
2090 /* pre-reserve space for file */
2091 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2092 /* max size for transfer still in size */
2095 if (afs_mariner && !tdc->f.chunk)
2096 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2098 * Right now, we only have one tool, and it's a hammer. So, we
2099 * fetch the whole file.
2101 DZap(tdc); /* pages in cache may be old */
2102 file = afs_CFileOpen(&tdc->f.inode);
2103 afs_RemoveVCB(&avc->f.fid);
2104 tdc->f.states |= DWriting;
2105 tdc->dflags |= DFFetching;
2106 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2107 if (tdc->mflags & DFFetchReq) {
2108 tdc->mflags &= ~DFFetchReq;
2109 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2110 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2111 __FILE__, ICL_TYPE_INT32, __LINE__,
2112 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2116 (struct afs_FetchOutput *)osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2117 setVcacheStatus = 0;
2120 * Remember if we are doing the reading from a replicated volume,
2121 * and how many times we've zipped around the fetch/analyze loop.
2123 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2125 accP = &(afs_stats_cmfullperf.accessinf);
2127 (accP->replicatedRefs)++;
2129 (accP->unreplicatedRefs)++;
2130 #endif /* AFS_NOSTATS */
2131 /* this is a cache miss */
2132 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2133 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2134 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2137 afs_stats_cmperf.dcacheMisses++;
2140 * Dynamic root support: fetch data from local memory.
2142 if (afs_IsDynroot(avc)) {
2146 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2148 dynrootDir += Position;
2149 dynrootLen -= Position;
2150 if (size > dynrootLen)
2154 code = afs_CFileWrite(file, 0, dynrootDir, size);
2162 tdc->validPos = Position + size;
2163 afs_CFileTruncate(file, size); /* prune it */
2164 } else if (afs_IsDynrootMount(avc)) {
2168 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2170 dynrootDir += Position;
2171 dynrootLen -= Position;
2172 if (size > dynrootLen)
2176 code = afs_CFileWrite(file, 0, dynrootDir, size);
2184 tdc->validPos = Position + size;
2185 afs_CFileTruncate(file, size); /* prune it */
2188 * Not a dynamic vnode: do the real fetch.
2193 * avc->lock(R) if setLocks && !slowPass
2194 * avc->lock(W) if !setLocks || slowPass
2198 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK);
2203 (accP->numReplicasAccessed)++;
2205 #endif /* AFS_NOSTATS */
2206 if (!setLocks || slowPass) {
2207 avc->callback = tc->srvr->server;
2209 newCallback = tc->srvr->server;
2213 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
2217 &(afs_stats_cmfullperf.rpc.
2218 fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
2219 osi_GetuTime(&xferStartTime);
2222 afs_CacheFetchProc(tc, file,
2223 (afs_size_t) Position, tdc,
2225 &bytesXferred, size, tsmall);
2227 osi_GetuTime(&xferStopTime);
2228 (xferP->numXfers)++;
2230 (xferP->numSuccesses)++;
2231 afs_stats_XferSumBytes
2232 [AFS_STATS_FS_XFERIDX_FETCHDATA] +=
2234 (xferP->sumBytes) +=
2235 (afs_stats_XferSumBytes
2236 [AFS_STATS_FS_XFERIDX_FETCHDATA] >> 10);
2237 afs_stats_XferSumBytes
2238 [AFS_STATS_FS_XFERIDX_FETCHDATA] &= 0x3FF;
2239 if (bytesXferred < xferP->minBytes)
2240 xferP->minBytes = bytesXferred;
2241 if (bytesXferred > xferP->maxBytes)
2242 xferP->maxBytes = bytesXferred;
2245 * Tally the size of the object. Note: we tally the actual size,
2246 * NOT the number of bytes that made it out over the wire.
2248 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
2249 (xferP->count[0])++;
2250 else if (bytesToXfer <=
2251 AFS_STATS_MAXBYTES_BUCKET1)
2252 (xferP->count[1])++;
2253 else if (bytesToXfer <=
2254 AFS_STATS_MAXBYTES_BUCKET2)
2255 (xferP->count[2])++;
2256 else if (bytesToXfer <=
2257 AFS_STATS_MAXBYTES_BUCKET3)
2258 (xferP->count[3])++;
2259 else if (bytesToXfer <=
2260 AFS_STATS_MAXBYTES_BUCKET4)
2261 (xferP->count[4])++;
2262 else if (bytesToXfer <=
2263 AFS_STATS_MAXBYTES_BUCKET5)
2264 (xferP->count[5])++;
2265 else if (bytesToXfer <=
2266 AFS_STATS_MAXBYTES_BUCKET6)
2267 (xferP->count[6])++;
2268 else if (bytesToXfer <=
2269 AFS_STATS_MAXBYTES_BUCKET7)
2270 (xferP->count[7])++;
2272 (xferP->count[8])++;
2274 afs_stats_GetDiff(elapsedTime, xferStartTime,
2276 afs_stats_AddTo((xferP->sumTime), elapsedTime);
2277 afs_stats_SquareAddTo((xferP->sqrTime),
2279 if (afs_stats_TimeLessThan
2280 (elapsedTime, (xferP->minTime))) {
2281 afs_stats_TimeAssign((xferP->minTime),
2284 if (afs_stats_TimeGreaterThan
2285 (elapsedTime, (xferP->maxTime))) {
2286 afs_stats_TimeAssign((xferP->maxTime),
2292 afs_CacheFetchProc(tc, file, Position, tdc,
2293 avc, 0, 0, size, tsmall);
2294 #endif /* AFS_NOSTATS */
2301 /* callback could have been broken (or expired) in a race here,
2302 * but we return the data anyway. It's as good as we knew about
2303 * when we started. */
2305 * validPos is updated by CacheFetchProc, and can only be
2306 * modifed under a dcache write lock, which we've blocked out
2308 size = tdc->validPos - Position; /* actual segment size */
2311 afs_CFileTruncate(file, size); /* prune it */
2313 if (!setLocks || slowPass) {
2314 ObtainWriteLock(&afs_xcbhash, 453);
2315 afs_DequeueCallback(avc);
2316 avc->f.states &= ~(CStatd | CUnique);
2317 avc->callback = NULL;
2318 ReleaseWriteLock(&afs_xcbhash);
2319 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2320 osi_dnlc_purgedp(avc);
2322 /* Something lost. Forget about performance, and go
2323 * back with a vcache write lock.
2325 afs_CFileTruncate(file, 0);
2326 afs_AdjustSize(tdc, 0);
2327 afs_CFileClose(file);
2328 osi_FreeLargeSpace(tsmall);
2330 ReleaseWriteLock(&tdc->lock);
2333 ReleaseReadLock(&avc->lock);
2335 goto RetryGetDCache;
2339 } while (afs_Analyze
2340 (tc, code, &avc->f.fid, areq,
2341 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2345 * avc->lock(R) if setLocks && !slowPass
2346 * avc->lock(W) if !setLocks || slowPass
2352 * In the case of replicated access, jot down info on the number of
2353 * attempts it took before we got through or gave up.
2356 if (numFetchLoops <= 1)
2357 (accP->refFirstReplicaOK)++;
2358 if (numFetchLoops > accP->maxReplicasPerRef)
2359 accP->maxReplicasPerRef = numFetchLoops;
2361 #endif /* AFS_NOSTATS */
2363 tdc->dflags &= ~DFFetching;
2364 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2365 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2366 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2367 tdc, ICL_TYPE_INT32, tdc->dflags);
2368 if (avc->execsOrWriters == 0)
2369 tdc->f.states &= ~DWriting;
2371 /* now, if code != 0, we have an error and should punt.
2372 * note that we have the vcache write lock, either because
2373 * !setLocks or slowPass.
2376 afs_CFileTruncate(file, 0);
2377 afs_AdjustSize(tdc, 0);
2378 afs_CFileClose(file);
2379 ZapDCE(tdc); /* sets DFEntryMod */
2380 if (vType(avc) == VDIR) {
2383 tdc->f.states &= ~(DRO|DBackup|DRW);
2384 afs_DCMoveBucket(tdc, 0, 0);
2385 ReleaseWriteLock(&tdc->lock);
2387 if (!afs_IsDynroot(avc)) {
2388 ObtainWriteLock(&afs_xcbhash, 454);
2389 afs_DequeueCallback(avc);
2390 avc->f.states &= ~(CStatd | CUnique);
2391 ReleaseWriteLock(&afs_xcbhash);
2392 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2393 osi_dnlc_purgedp(avc);
2396 * avc->lock(W); assert(!setLocks || slowPass)
2398 osi_Assert(!setLocks || slowPass);
2404 /* otherwise we copy in the just-fetched info */
2405 afs_CFileClose(file);
2406 afs_AdjustSize(tdc, size); /* new size */
2408 * Copy appropriate fields into vcache. Status is
2409 * copied later where we selectively acquire the
2410 * vcache write lock.
2413 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2415 setVcacheStatus = 1;
2416 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2417 tsmall->OutStatus.DataVersion);
2418 tdc->dflags |= DFEntryMod;
2419 afs_indexFlags[tdc->index] |= IFEverUsed;
2420 ConvertWToSLock(&tdc->lock);
2421 } /*Data version numbers don't match */
2424 * Data version numbers match.
2426 afs_stats_cmperf.dcacheHits++;
2427 } /*Data version numbers match */
2429 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2433 * avc->lock(R) if setLocks && !slowPass
2434 * avc->lock(W) if !setLocks || slowPass
2435 * tdc->lock(S) if tdc
2439 * See if this was a reference to a file in the local cell.
2441 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2442 afs_stats_cmperf.dlocalAccesses++;
2444 afs_stats_cmperf.dremoteAccesses++;
2446 /* Fix up LRU info */
2449 MObtainWriteLock(&afs_xdcache, 602);
2450 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2451 hadd32(afs_indexCounter, 1);
2452 MReleaseWriteLock(&afs_xdcache);
2454 /* return the data */
2455 if (vType(avc) == VDIR)
2458 *aoffset = AFS_CHUNKOFFSET(abyte);
2459 *alen = (tdc->f.chunkBytes - *aoffset);
2460 ReleaseSharedLock(&tdc->lock);
2465 * avc->lock(R) if setLocks && !slowPass
2466 * avc->lock(W) if !setLocks || slowPass
2469 /* Fix up the callback and status values in the vcache */
2471 if (setLocks && !slowPass) {
2474 * This is our dirty little secret to parallel fetches.
2475 * We don't write-lock the vcache while doing the fetch,
2476 * but potentially we'll need to update the vcache after
2477 * the fetch is done.
2479 * Drop the read lock and try to re-obtain the write
2480 * lock. If the vcache still has the same DV, it's
2481 * ok to go ahead and install the new data.
2483 afs_hyper_t currentDV, statusDV;
2485 hset(currentDV, avc->f.m.DataVersion);
2487 if (setNewCallback && avc->callback != newCallback)
2491 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2492 tsmall->OutStatus.DataVersion);
2494 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2496 if (setVcacheStatus && !hsame(currentDV, statusDV))
2500 ReleaseReadLock(&avc->lock);
2502 if (doVcacheUpdate) {
2503 ObtainWriteLock(&avc->lock, 615);
2504 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2505 /* We lose. Someone will beat us to it. */
2507 ReleaseWriteLock(&avc->lock);
2512 /* With slow pass, we've already done all the updates */
2514 ReleaseWriteLock(&avc->lock);
2517 /* Check if we need to perform any last-minute fixes with a write-lock */
2518 if (!setLocks || doVcacheUpdate) {
2520 avc->callback = newCallback;
2521 if (tsmall && setVcacheStatus)
2522 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2524 ReleaseWriteLock(&avc->lock);
2528 osi_FreeLargeSpace(tsmall);
2531 } /*afs_GetDCache */
2535 * afs_WriteThroughDSlots
2538 * Sweep through the dcache slots and write out any modified
2539 * in-memory data back on to our caching store.
2545 * The afs_xdcache is write-locked through this whole affair.
2548 afs_WriteThroughDSlots(void)
2550 register struct dcache *tdc;
2551 register afs_int32 i, touchedit = 0;
2553 struct afs_q DirtyQ, *tq;
2555 AFS_STATCNT(afs_WriteThroughDSlots);
2558 * Because of lock ordering, we can't grab dcache locks while
2559 * holding afs_xdcache. So we enter xdcache, get a reference
2560 * for every dcache entry, and exit xdcache.
2562 MObtainWriteLock(&afs_xdcache, 283);
2564 for (i = 0; i < afs_cacheFiles; i++) {
2565 tdc = afs_indexTable[i];
2567 /* Grab tlock in case the existing refcount isn't zero */
2568 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2569 ObtainWriteLock(&tdc->tlock, 623);
2571 ReleaseWriteLock(&tdc->tlock);
2573 QAdd(&DirtyQ, &tdc->dirty);
2576 MReleaseWriteLock(&afs_xdcache);
2579 * Now, for each dcache entry we found, check if it's dirty.
2580 * If so, get write-lock, get afs_xdcache, which protects
2581 * afs_cacheInodep, and flush it. Don't forget to put back
2585 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2587 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2589 if (tdc->dflags & DFEntryMod) {
2592 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2594 /* Now that we have the write lock, double-check */
2595 if (wrLock && (tdc->dflags & DFEntryMod)) {
2596 tdc->dflags &= ~DFEntryMod;
2597 MObtainWriteLock(&afs_xdcache, 620);
2598 afs_WriteDCache(tdc, 1);
2599 MReleaseWriteLock(&afs_xdcache);
2603 ReleaseWriteLock(&tdc->lock);
2609 MObtainWriteLock(&afs_xdcache, 617);
2610 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2611 /* Touch the file to make sure that the mtime on the file is kept
2612 * up-to-date to avoid losing cached files on cold starts because
2613 * their mtime seems old...
2615 struct afs_fheader theader;
2617 theader.magic = AFS_FHMAGIC;
2618 theader.firstCSize = AFS_FIRSTCSIZE;
2619 theader.otherCSize = AFS_OTHERCSIZE;
2620 theader.version = AFS_CI_VERSION;
2621 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2623 MReleaseWriteLock(&afs_xdcache);
2630 * Return a pointer to an freshly initialized dcache entry using
2631 * a memory-based cache. The tlock will be read-locked.
2634 * aslot : Dcache slot to look at.
2635 * tmpdc : Ptr to dcache entry.
2638 * Must be called with afs_xdcache write-locked.
2642 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2644 register struct dcache *tdc;
2647 AFS_STATCNT(afs_MemGetDSlot);
2648 if (CheckLock(&afs_xdcache) != -1)
2649 osi_Panic("getdslot nolock");
2650 if (aslot < 0 || aslot >= afs_cacheFiles)
2651 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2652 tdc = afs_indexTable[aslot];
2654 QRemove(&tdc->lruq); /* move to queue head */
2655 QAdd(&afs_DLRU, &tdc->lruq);
2656 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2657 ObtainWriteLock(&tdc->tlock, 624);
2659 ConvertWToRLock(&tdc->tlock);
2662 if (tmpdc == NULL) {
2663 if (!afs_freeDSList)
2664 afs_GetDownDSlot(4);
2665 if (!afs_freeDSList) {
2666 /* none free, making one is better than a panic */
2667 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2668 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2669 #ifdef KERNEL_HAVE_PIN
2670 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2673 tdc = afs_freeDSList;
2674 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2677 tdc->dflags = 0; /* up-to-date, not in free q */
2679 QAdd(&afs_DLRU, &tdc->lruq);
2680 if (tdc->lruq.prev == &tdc->lruq)
2681 osi_Panic("lruq 3");
2687 /* initialize entry */
2688 tdc->f.fid.Cell = 0;
2689 tdc->f.fid.Fid.Volume = 0;
2691 hones(tdc->f.versionNo);
2692 tdc->f.inode.mem = aslot;
2693 tdc->dflags |= DFEntryMod;
2696 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2699 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2700 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2701 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2704 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2705 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2706 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2707 ObtainReadLock(&tdc->tlock);
2710 afs_indexTable[aslot] = tdc;
2713 } /*afs_MemGetDSlot */
2715 unsigned int last_error = 0, lasterrtime = 0;
2721 * Return a pointer to an freshly initialized dcache entry using
2722 * a UFS-based disk cache. The dcache tlock will be read-locked.
2725 * aslot : Dcache slot to look at.
2726 * tmpdc : Ptr to dcache entry.
2729 * afs_xdcache lock write-locked.
2732 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2734 register afs_int32 code;
2735 register struct dcache *tdc;
2739 AFS_STATCNT(afs_UFSGetDSlot);
2740 if (CheckLock(&afs_xdcache) != -1)
2741 osi_Panic("getdslot nolock");
2742 if (aslot < 0 || aslot >= afs_cacheFiles)
2743 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2744 tdc = afs_indexTable[aslot];
2746 QRemove(&tdc->lruq); /* move to queue head */
2747 QAdd(&afs_DLRU, &tdc->lruq);
2748 /* Grab tlock in case refCount != 0 */
2749 ObtainWriteLock(&tdc->tlock, 625);
2751 ConvertWToRLock(&tdc->tlock);
2754 /* otherwise we should read it in from the cache file */
2756 * If we weren't passed an in-memory region to place the file info,
2757 * we have to allocate one.
2759 if (tmpdc == NULL) {
2760 if (!afs_freeDSList)
2761 afs_GetDownDSlot(4);
2762 if (!afs_freeDSList) {
2763 /* none free, making one is better than a panic */
2764 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2765 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2766 #ifdef KERNEL_HAVE_PIN
2767 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2770 tdc = afs_freeDSList;
2771 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2774 tdc->dflags = 0; /* up-to-date, not in free q */
2776 QAdd(&afs_DLRU, &tdc->lruq);
2777 if (tdc->lruq.prev == &tdc->lruq)
2778 osi_Panic("lruq 3");
2785 * Seek to the aslot'th entry and read it in.
2788 afs_osi_Read(afs_cacheInodep,
2789 sizeof(struct fcache) * aslot +
2790 sizeof(struct afs_fheader), (char *)(&tdc->f),
2791 sizeof(struct fcache));
2793 if (code != sizeof(struct fcache))
2795 if (!afs_CellNumValid(tdc->f.fid.Cell))
2799 tdc->f.fid.Cell = 0;
2800 tdc->f.fid.Fid.Volume = 0;
2802 hones(tdc->f.versionNo);
2803 tdc->dflags |= DFEntryMod;
2804 #if defined(KERNEL_HAVE_UERROR)
2805 last_error = getuerror();
2807 lasterrtime = osi_Time();
2808 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2809 tdc->f.states &= ~(DRO|DBackup|DRW);
2810 afs_DCMoveBucket(tdc, 0, 0);
2813 if (tdc->f.states & DRO) {
2814 afs_DCMoveBucket(tdc, 0, 2);
2815 } else if (tdc->f.states & DBackup) {
2816 afs_DCMoveBucket(tdc, 0, 1);
2818 afs_DCMoveBucket(tdc, 0, 1);
2824 if (tdc->f.chunk >= 0)
2825 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2830 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2831 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2832 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2835 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2836 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2837 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2838 ObtainReadLock(&tdc->tlock);
2841 * If we didn't read into a temporary dcache region, update the
2842 * slot pointer table.
2845 afs_indexTable[aslot] = tdc;
2848 } /*afs_UFSGetDSlot */
2853 * Write a particular dcache entry back to its home in the
2856 * \param adc Pointer to the dcache entry to write.
2857 * \param atime If true, set the modtime on the file to the current time.
2859 * \note Environment:
2860 * Must be called with the afs_xdcache lock at least read-locked,
2861 * and dcache entry at least read-locked.
2862 * The reference count is not changed.
2866 afs_WriteDCache(register struct dcache *adc, int atime)
2868 register afs_int32 code;
2870 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2872 AFS_STATCNT(afs_WriteDCache);
2873 osi_Assert(WriteLocked(&afs_xdcache));
2875 adc->f.modTime = osi_Time();
2877 * Seek to the right dcache slot and write the in-memory image out to disk.
2879 afs_cellname_write();
2881 afs_osi_Write(afs_cacheInodep,
2882 sizeof(struct fcache) * adc->index +
2883 sizeof(struct afs_fheader), (char *)(&adc->f),
2884 sizeof(struct fcache));
2885 if (code != sizeof(struct fcache))
2893 * Wake up users of a particular file waiting for stores to take
2896 * \param avc Ptr to related vcache entry.
2898 * \note Environment:
2899 * Nothing interesting.
2902 afs_wakeup(register struct vcache *avc)
2905 register struct brequest *tb;
2907 AFS_STATCNT(afs_wakeup);
2908 for (i = 0; i < NBRS; i++, tb++) {
2909 /* if request is valid and for this file, we've found it */
2910 if (tb->refCount > 0 && avc == tb->vc) {
2913 * If CSafeStore is on, then we don't awaken the guy
2914 * waiting for the store until the whole store has finished.
2915 * Otherwise, we do it now. Note that if CSafeStore is on,
2916 * the BStore routine actually wakes up the user, instead
2918 * I think this is redundant now because this sort of thing
2919 * is already being handled by the higher-level code.
2921 if ((avc->f.states & CSafeStore) == 0) {
2923 tb->flags |= BUVALID;
2924 if (tb->flags & BUWAIT) {
2925 tb->flags &= ~BUWAIT;
2937 * Given a file name and inode, set up that file to be an
2938 * active member in the AFS cache. This also involves checking
2939 * the usability of its data.
2941 * \param afile Name of the cache file to initialize.
2942 * \param ainode Inode of the file.
2944 * \note Environment:
2945 * This function is called only during initialization.
2948 afs_InitCacheFile(char *afile, ino_t ainode)
2950 register afs_int32 code;
2953 struct osi_file *tfile;
2954 struct osi_stat tstat;
2955 register struct dcache *tdc;
2957 AFS_STATCNT(afs_InitCacheFile);
2958 index = afs_stats_cmperf.cacheNumEntries;
2959 if (index >= afs_cacheFiles)
2962 MObtainWriteLock(&afs_xdcache, 282);
2963 tdc = afs_GetDSlot(index, NULL);
2964 ReleaseReadLock(&tdc->tlock);
2965 MReleaseWriteLock(&afs_xdcache);
2967 ObtainWriteLock(&tdc->lock, 621);
2968 MObtainWriteLock(&afs_xdcache, 622);
2970 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
2972 ReleaseWriteLock(&afs_xdcache);
2973 ReleaseWriteLock(&tdc->lock);
2978 /* Add any other 'complex' inode types here ... */
2979 #if defined(UKERNEL) || !defined(LINUX_USE_FH)
2980 tdc->f.inode.ufs = ainode;
2982 osi_Panic("Can't init cache with inode numbers when complex inodes are "
2987 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
2989 tfile = osi_UFSOpen(&tdc->f.inode);
2990 code = afs_osi_Stat(tfile, &tstat);
2992 osi_Panic("initcachefile stat");
2995 * If file size doesn't match the cache info file, it's probably bad.
2997 if (tdc->f.chunkBytes != tstat.size)
2999 tdc->f.chunkBytes = 0;
3002 * If file changed within T (120?) seconds of cache info file, it's
3003 * probably bad. In addition, if slot changed within last T seconds,
3004 * the cache info file may be incorrectly identified, and so slot
3007 if (cacheInfoModTime < tstat.mtime + 120)
3009 if (cacheInfoModTime < tdc->f.modTime + 120)
3011 /* In case write through is behind, make sure cache items entry is
3012 * at least as new as the chunk.
3014 if (tdc->f.modTime < tstat.mtime)
3017 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3018 if (tstat.size != 0)
3019 osi_UFSTruncate(tfile, 0);
3020 tdc->f.states &= ~(DRO|DBackup|DRW);
3021 afs_DCMoveBucket(tdc, 0, 0);
3022 /* put entry in free cache slot list */
3023 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3024 afs_freeDCList = index;
3026 afs_indexFlags[index] |= IFFree;
3027 afs_indexUnique[index] = 0;
3030 * We must put this entry in the appropriate hash tables.
3031 * Note that i is still set from the above DCHash call
3033 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3034 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3035 afs_dchashTbl[code] = tdc->index;
3036 code = DVHash(&tdc->f.fid);
3037 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3038 afs_dvhashTbl[code] = tdc->index;
3039 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3041 /* has nontrivial amt of data */
3042 afs_indexFlags[index] |= IFEverUsed;
3043 afs_stats_cmperf.cacheFilesReused++;
3045 * Initialize index times to file's mod times; init indexCounter
3048 hset32(afs_indexTimes[index], tstat.atime);
3049 if (hgetlo(afs_indexCounter) < tstat.atime) {
3050 hset32(afs_indexCounter, tstat.atime);
3052 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3053 } /*File is not bad */
3055 osi_UFSClose(tfile);
3056 tdc->f.states &= ~DWriting;
3057 tdc->dflags &= ~DFEntryMod;
3058 /* don't set f.modTime; we're just cleaning up */
3059 afs_WriteDCache(tdc, 0);
3060 ReleaseWriteLock(&afs_xdcache);
3061 ReleaseWriteLock(&tdc->lock);
3063 afs_stats_cmperf.cacheNumEntries++;
3068 /*Max # of struct dcache's resident at any time*/
3070 * If 'dchint' is enabled then in-memory dcache min is increased because of
3076 * Initialize dcache related variables.
3086 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3088 register struct dcache *tdp;
3092 afs_freeDCList = NULLIDX;
3093 afs_discardDCList = NULLIDX;
3094 afs_freeDCCount = 0;
3095 afs_freeDSList = NULL;
3096 hzero(afs_indexCounter);
3098 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3104 if (achunk < 0 || achunk > 30)
3105 achunk = 13; /* Use default */
3106 AFS_SETCHUNKSIZE(achunk);
3112 if (aflags & AFSCALL_INIT_MEMCACHE) {
3114 * Use a memory cache instead of a disk cache
3116 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3117 afs_cacheType = &afs_MemCacheOps;
3118 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3119 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3120 /* ablocks is reported in 1K blocks */
3121 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3123 printf("afsd: memory cache too large for available memory.\n");
3124 printf("afsd: AFS files cannot be accessed.\n\n");
3126 afiles = ablocks = 0;
3128 printf("Memory cache: Allocating %d dcache entries...",
3131 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3132 afs_cacheType = &afs_UfsCacheOps;
3135 if (aDentries > 512)
3136 afs_dhashsize = 2048;
3137 /* initialize hash tables */
3139 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3141 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3142 for (i = 0; i < afs_dhashsize; i++) {
3143 afs_dvhashTbl[i] = NULLIDX;
3144 afs_dchashTbl[i] = NULLIDX;
3146 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3147 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3148 for (i = 0; i < afiles; i++) {
3149 afs_dvnextTbl[i] = NULLIDX;
3150 afs_dcnextTbl[i] = NULLIDX;
3153 /* Allocate and zero the pointer array to the dcache entries */
3154 afs_indexTable = (struct dcache **)
3155 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3156 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3158 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3159 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3161 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3162 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3163 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3164 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3166 /* Allocate and thread the struct dcache entries themselves */
3167 tdp = afs_Initial_freeDSList =
3168 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3169 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3170 #ifdef KERNEL_HAVE_PIN
3171 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3172 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3173 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3174 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3175 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3176 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3177 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3178 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3179 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3182 afs_freeDSList = &tdp[0];
3183 for (i = 0; i < aDentries - 1; i++) {
3184 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3185 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3186 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3187 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3189 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3190 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3191 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3192 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3194 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3195 afs_cacheBlocks = ablocks;
3196 afs_ComputeCacheParms(); /* compute parms based on cache size */
3198 afs_dcentries = aDentries;
3200 afs_stats_cmperf.cacheBucket0_Discarded =
3201 afs_stats_cmperf.cacheBucket1_Discarded =
3202 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3208 * Shuts down the cache.
3212 shutdown_dcache(void)
3216 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3217 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3218 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3219 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3220 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3221 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3222 afs_osi_Free(afs_Initial_freeDSList,
3223 afs_dcentries * sizeof(struct dcache));
3224 #ifdef KERNEL_HAVE_PIN
3225 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3226 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3227 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3228 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3229 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3230 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3231 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3235 for (i = 0; i < afs_dhashsize; i++) {
3236 afs_dvhashTbl[i] = NULLIDX;
3237 afs_dchashTbl[i] = NULLIDX;
3240 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3241 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3243 afs_blocksUsed = afs_dcentries = 0;
3244 afs_stats_cmperf.cacheBucket0_Discarded =
3245 afs_stats_cmperf.cacheBucket1_Discarded =
3246 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3247 hzero(afs_indexCounter);
3249 afs_freeDCCount = 0;
3250 afs_freeDCList = NULLIDX;
3251 afs_discardDCList = NULLIDX;
3252 afs_freeDSList = afs_Initial_freeDSList = 0;
3254 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3260 * Get a dcache ready for writing, respecting the current cache size limits
3262 * len is required because afs_GetDCache with flag == 4 expects the length
3263 * field to be filled. It decides from this whether it's necessary to fetch
3264 * data into the chunk before writing or not (when the whole chunk is
3267 * \param avc The vcache to fetch a dcache for
3268 * \param filePos The start of the section to be written
3269 * \param len The length of the section to be written
3273 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3274 * must be released and afs_PutDCache() called to free dcache.
3277 * \note avc->lock must be held on entry. Function may release and reobtain
3278 * avc->lock and GLOCK.
3282 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3283 afs_size_t len, struct vrequest *areq,
3285 struct dcache *tdc = NULL;
3288 /* read the cached info */
3290 tdc = afs_FindDCache(avc, filePos);
3292 ObtainWriteLock(&tdc->lock, 657);
3293 } else if (afs_blocksUsed >
3294 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3295 tdc = afs_FindDCache(avc, filePos);
3297 ObtainWriteLock(&tdc->lock, 658);
3298 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3299 || (tdc->dflags & DFFetching)) {
3300 ReleaseWriteLock(&tdc->lock);
3306 afs_MaybeWakeupTruncateDaemon();
3307 while (afs_blocksUsed >
3308 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3309 ReleaseWriteLock(&avc->lock);
3310 if (afs_blocksUsed - afs_blocksDiscarded >
3311 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3312 afs_WaitForCacheDrain = 1;
3313 afs_osi_Sleep(&afs_WaitForCacheDrain);
3315 afs_MaybeFreeDiscardedDCache();
3316 afs_MaybeWakeupTruncateDaemon();
3317 ObtainWriteLock(&avc->lock, 509);
3319 avc->f.states |= CDirty;
3320 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3322 ObtainWriteLock(&tdc->lock, 659);
3325 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3327 ObtainWriteLock(&tdc->lock, 660);
3330 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3331 afs_stats_cmperf.cacheCurrDirtyChunks++;
3332 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3334 if (!(tdc->f.states & DWriting)) {
3335 /* don't mark entry as mod if we don't have to */
3336 tdc->f.states |= DWriting;
3337 tdc->dflags |= DFEntryMod;
3343 #if defined(AFS_DISCON_ENV)
3346 * Make a shadow copy of a dir's dcache. It's used for disconnected
3347 * operations like remove/create/rename to keep the original directory data.
3348 * On reconnection, we can diff the original data with the server and get the
3349 * server changes and with the local data to get the local changes.
3351 * \param avc The dir vnode.
3352 * \param adc The dir dcache.
3354 * \return 0 for success.
3356 * \note The vcache entry must be write locked.
3357 * \note The dcache entry must be read locked.
3359 int afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3361 int i, code, ret_code = 0, written, trans_size;
3362 struct dcache *new_dc = NULL;
3363 struct osi_file *tfile_src, *tfile_dst;
3364 struct VenusFid shadow_fid;
3367 /* Is this a dir? */
3368 if (vType(avc) != VDIR)
3371 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3374 /* Generate a fid for the shadow dir. */
3375 shadow_fid.Cell = avc->f.fid.Cell;
3376 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3377 afs_GenShadowFid(&shadow_fid);
3379 ObtainWriteLock(&afs_xdcache, 716);
3381 /* Get a fresh dcache. */
3382 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3384 ObtainReadLock(&adc->mflock);
3386 /* Set up the new fid. */
3387 /* Copy interesting data from original dir dcache. */
3388 new_dc->mflags = adc->mflags;
3389 new_dc->dflags = adc->dflags;
3390 new_dc->f.modTime = adc->f.modTime;
3391 new_dc->f.versionNo = adc->f.versionNo;
3392 new_dc->f.states = adc->f.states;
3393 new_dc->f.chunk= adc->f.chunk;
3394 new_dc->f.chunkBytes = adc->f.chunkBytes;
3396 ReleaseReadLock(&adc->mflock);
3398 /* Now add to the two hash chains */
3399 i = DCHash(&shadow_fid, 0);
3400 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3401 afs_dchashTbl[i] = new_dc->index;
3403 i = DVHash(&shadow_fid);
3404 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3405 afs_dvhashTbl[i] = new_dc->index;
3407 ReleaseWriteLock(&afs_xdcache);
3409 /* Alloc a 4k block. */
3410 data = (char *) afs_osi_Alloc(4096);
3412 printf("afs_MakeShadowDir: could not alloc data\n");
3417 /* Open the files. */
3418 tfile_src = afs_CFileOpen(&adc->f.inode);
3419 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3421 /* And now copy dir dcache data into this dcache,
3425 while (written < adc->f.chunkBytes) {
3426 trans_size = adc->f.chunkBytes - written;
3427 if (trans_size > 4096)
3430 /* Read a chunk from the dcache. */
3431 code = afs_CFileRead(tfile_src, written, data, trans_size);
3432 if (code < trans_size) {
3437 /* Write it to the new dcache. */
3438 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3439 if (code < trans_size) {
3444 written+=trans_size;
3447 afs_CFileClose(tfile_dst);
3448 afs_CFileClose(tfile_src);
3450 afs_osi_Free(data, 4096);
3452 ReleaseWriteLock(&new_dc->lock);
3453 afs_PutDCache(new_dc);
3456 ObtainWriteLock(&afs_xvcache, 763);
3457 ObtainWriteLock(&afs_disconDirtyLock, 765);
3458 QAdd(&afs_disconShadow, &avc->shadowq);
3460 ReleaseWriteLock(&afs_disconDirtyLock);
3461 ReleaseWriteLock(&afs_xvcache);
3463 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3464 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3472 * Delete the dcaches of a shadow dir.
3474 * \param avc The vcache containing the shadow fid.
3476 * \note avc must be write locked.
3478 void afs_DeleteShadowDir(struct vcache *avc)
3481 struct VenusFid shadow_fid;
3483 shadow_fid.Cell = avc->f.fid.Cell;
3484 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3485 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3486 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3488 tdc = afs_FindDCacheByFid(&shadow_fid);
3490 afs_HashOutDCache(tdc, 1);
3491 afs_DiscardDCache(tdc);
3494 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3495 ObtainWriteLock(&afs_disconDirtyLock, 708);
3496 QRemove(&avc->shadowq);
3497 ReleaseWriteLock(&afs_disconDirtyLock);
3498 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3502 * Populate a dcache with empty chunks up to a given file size,
3503 * used before extending a file in order to avoid 'holes' which
3504 * we can't access in disconnected mode.
3506 * \param avc The vcache which is being extended (locked)
3507 * \param alen The new length of the file
3510 void afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq) {
3512 afs_size_t len, offset;
3513 afs_int32 start, end;
3515 /* We're doing this to deal with the situation where we extend
3516 * by writing after lseek()ing past the end of the file . If that
3517 * extension skips chunks, then those chunks won't be created, and
3518 * GetDCache will assume that they have to be fetched from the server.
3519 * So, for each chunk between the current file position, and the new
3520 * length we GetDCache for that chunk.
3523 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3526 if (avc->f.m.Length == 0)
3529 start = AFS_CHUNK(avc->f.m.Length)+1;
3531 end = AFS_CHUNK(apos);
3534 len = AFS_CHUNKTOSIZE(start);
3535 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);