2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
13 #include <afsconfig.h>
14 #include "afs/param.h"
17 #include "afs/sysincludes.h" /*Standard vendor system headers */
18 #include "afsincludes.h" /*AFS-based standard headers */
19 #include "afs/afs_stats.h" /* statistics */
20 #include "afs/afs_cbqueue.h"
21 #include "afs/afs_osidnlc.h"
23 /* Forward declarations. */
24 static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
25 static void afs_FreeDiscardedDCache(void);
26 static void afs_DiscardDCache(struct dcache *);
27 static void afs_FreeDCache(struct dcache *);
29 static afs_int32 afs_DCGetBucket(struct vcache *);
30 static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
31 static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
32 static void afs_DCSizeInit(void);
33 static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);
36 * --------------------- Exported definitions ---------------------
39 afs_int32 afs_blocksUsed_0; /*!< 1K blocks in cache - in theory is zero */
40 afs_int32 afs_blocksUsed_1; /*!< 1K blocks in cache */
41 afs_int32 afs_blocksUsed_2; /*!< 1K blocks in cache */
42 afs_int32 afs_pct1 = -1;
43 afs_int32 afs_pct2 = -1;
44 afs_uint32 afs_tpct1 = 0;
45 afs_uint32 afs_tpct2 = 0;
46 afs_uint32 splitdcache = 0;
48 afs_lock_t afs_xdcache; /*!< Lock: alloc new disk cache entries */
49 afs_int32 afs_freeDCList; /*!< Free list for disk cache entries */
50 afs_int32 afs_freeDCCount; /*!< Count of elts in freeDCList */
51 afs_int32 afs_discardDCList; /*!< Discarded disk cache entries */
52 afs_int32 afs_discardDCCount; /*!< Count of elts in discardDCList */
53 struct dcache *afs_freeDSList; /*!< Free list for disk slots */
54 struct dcache *afs_Initial_freeDSList; /*!< Initial list for above */
55 afs_dcache_id_t cacheInode; /*!< Inode for CacheItems file */
56 struct osi_file *afs_cacheInodep = 0; /*!< file for CacheItems inode */
57 struct afs_q afs_DLRU; /*!< dcache LRU */
58 afs_int32 afs_dhashsize = 1024;
59 afs_int32 *afs_dvhashTbl; /*!< Data cache hash table: hashed by FID + chunk number. */
60 afs_int32 *afs_dchashTbl; /*!< Data cache hash table: hashed by FID. */
61 afs_int32 *afs_dvnextTbl; /*!< Dcache hash table links */
62 afs_int32 *afs_dcnextTbl; /*!< Dcache hash table links */
63 struct dcache **afs_indexTable; /*!< Pointers to dcache entries */
64 afs_hyper_t *afs_indexTimes; /*!< Dcache entry Access times */
65 afs_int32 *afs_indexUnique; /*!< dcache entry Fid.Unique */
66 unsigned char *afs_indexFlags; /*!< (only one) Is there data there? */
67 afs_hyper_t afs_indexCounter; /*!< Fake time for marking index
69 afs_int32 afs_cacheFiles = 0; /*!< Size of afs_indexTable */
70 afs_int32 afs_cacheBlocks; /*!< 1K blocks in cache */
71 afs_int32 afs_cacheStats; /*!< Stat entries in cache */
72 afs_int32 afs_blocksUsed; /*!< Number of blocks in use */
73 afs_int32 afs_blocksDiscarded; /*!<Blocks freed but not truncated */
74 afs_int32 afs_fsfragsize = AFS_MIN_FRAGSIZE; /*!< Underlying Filesystem minimum unit
75 *of disk allocation usually 1K
76 *this value is (truefrag -1 ) to
77 *save a bunch of subtracts... */
78 #ifdef AFS_64BIT_CLIENT
79 #ifdef AFS_VM_RDWR_ENV
80 afs_size_t afs_vmMappingEnd; /* !< For large files (>= 2GB) the VM
81 * mapping an 32bit addressing machines
82 * can only be used below the 2 GB
83 * line. From this point upwards we
84 * must do direct I/O into the cache
85 * files. The value should be on a
87 #endif /* AFS_VM_RDWR_ENV */
88 #endif /* AFS_64BIT_CLIENT */
90 /* The following is used to ensure that new dcache's aren't obtained when
91 * the cache is nearly full.
93 int afs_WaitForCacheDrain = 0;
94 int afs_TruncateDaemonRunning = 0;
95 int afs_CacheTooFull = 0;
97 afs_int32 afs_dcentries; /*!< In-memory dcache entries */
100 int dcacheDisabled = 0;
102 struct afs_cacheOps afs_UfsCacheOps = {
103 #if defined(AFS_SGI_ENV) && !defined(__c99)
116 .truncate = osi_UFSTruncate,
117 .fread = afs_osi_Read,
118 .fwrite = afs_osi_Write,
119 .close = osi_UFSClose,
120 .vread = afs_UFSRead,
121 .vwrite = afs_UFSWrite,
122 .GetDSlot = afs_UFSGetDSlot,
123 .GetVolSlot = afs_UFSGetVolSlot,
124 .HandleLink = afs_UFSHandleLink,
128 struct afs_cacheOps afs_MemCacheOps = {
129 #if (defined(AFS_SGI_ENV) && !defined(__c99))
131 afs_MemCacheTruncate,
141 .open = afs_MemCacheOpen,
142 .truncate = afs_MemCacheTruncate,
143 .fread = afs_MemReadBlk,
144 .fwrite = afs_MemWriteBlk,
145 .close = afs_MemCacheClose,
146 .vread = afs_MemRead,
147 .vwrite = afs_MemWrite,
148 .GetDSlot = afs_MemGetDSlot,
149 .GetVolSlot = afs_MemGetVolSlot,
150 .HandleLink = afs_MemHandleLink,
154 int cacheDiskType; /*Type of backing disk for cache */
155 struct afs_cacheOps *afs_cacheType;
158 * Where is this vcache's entry associated dcache located/
159 * \param avc The vcache entry.
160 * \return Bucket index:
165 afs_DCGetBucket(struct vcache *avc)
170 /* This should be replaced with some sort of user configurable function */
171 if (avc->f.states & CRO) {
173 } else if (avc->f.states & CBackup) {
183 * Readjust a dcache's size.
185 * \param adc The dcache to be adjusted.
186 * \param oldSize Old size for the dcache.
187 * \param newSize The new size to be adjusted to.
191 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
193 afs_int32 adjustSize = newSize - oldSize;
201 afs_blocksUsed_0 += adjustSize;
202 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
205 afs_blocksUsed_1 += adjustSize;
206 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
209 afs_blocksUsed_2 += adjustSize;
210 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
218 * Move a dcache from one bucket to another.
220 * \param adc Operate on this dcache.
221 * \param size Size in bucket (?).
222 * \param newBucket Destination bucket.
226 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
231 /* Substract size from old bucket. */
235 afs_blocksUsed_0 -= size;
238 afs_blocksUsed_1 -= size;
241 afs_blocksUsed_2 -= size;
245 /* Set new bucket and increase destination bucket size. */
246 adc->bucket = newBucket;
251 afs_blocksUsed_0 += size;
254 afs_blocksUsed_1 += size;
257 afs_blocksUsed_2 += size;
265 * Init split caches size.
270 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
279 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
284 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
285 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
287 /* Short cut: if we don't know about it, try to kill it */
288 if (phase < 2 && afs_blocksUsed_0)
291 if (afs_pct1 > afs_tpct1)
293 if (afs_pct2 > afs_tpct2)
295 return 0; /* unlikely */
300 * Warn about failing to store a file.
302 * \param acode Associated error code.
303 * \param avolume Volume involved.
304 * \param aflags How to handle the output:
305 * aflags & 1: Print out on console
306 * aflags & 2: Print out on controlling tty
308 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
312 afs_StoreWarn(register afs_int32 acode, afs_int32 avolume,
313 register afs_int32 aflags)
315 static char problem_fmt[] =
316 "afs: failed to store file in volume %d (%s)\n";
317 static char problem_fmt_w_error[] =
318 "afs: failed to store file in volume %d (error %d)\n";
319 static char netproblems[] = "network problems";
320 static char partfull[] = "partition full";
321 static char overquota[] = "over quota";
323 AFS_STATCNT(afs_StoreWarn);
329 afs_warn(problem_fmt, avolume, netproblems);
331 afs_warnuser(problem_fmt, avolume, netproblems);
332 } else if (acode == ENOSPC) {
337 afs_warn(problem_fmt, avolume, partfull);
339 afs_warnuser(problem_fmt, avolume, partfull);
342 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
343 * Instead ENOSPC will be sent...
345 if (acode == EDQUOT) {
350 afs_warn(problem_fmt, avolume, overquota);
352 afs_warnuser(problem_fmt, avolume, overquota);
360 afs_warn(problem_fmt_w_error, avolume, acode);
362 afs_warnuser(problem_fmt_w_error, avolume, acode);
367 * Try waking up truncation daemon, if it's worth it.
370 afs_MaybeWakeupTruncateDaemon(void)
372 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
373 afs_CacheTooFull = 1;
374 if (!afs_TruncateDaemonRunning)
375 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
376 } else if (!afs_TruncateDaemonRunning
377 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
378 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
385 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
386 * struct so we need only export one symbol for AIX.
388 static struct CTD_stats {
389 osi_timeval_t CTD_beforeSleep;
390 osi_timeval_t CTD_afterSleep;
391 osi_timeval_t CTD_sleepTime;
392 osi_timeval_t CTD_runTime;
396 u_int afs_min_cache = 0;
399 * Keeps the cache clean and free by truncating uneeded files, when used.
404 afs_CacheTruncateDaemon(void)
406 osi_timeval_t CTD_tmpTime;
410 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
412 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
414 osi_GetuTime(&CTD_stats.CTD_afterSleep);
415 afs_TruncateDaemonRunning = 1;
417 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
418 ObtainWriteLock(&afs_xdcache, 266);
419 if (afs_CacheTooFull) {
420 int space_needed, slots_needed;
421 /* if we get woken up, we should try to clean something out */
422 for (counter = 0; counter < 10; counter++) {
424 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
426 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
427 afs_GetDownD(slots_needed, &space_needed, 0);
428 if ((space_needed <= 0) && (slots_needed <= 0)) {
431 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
434 if (!afs_CacheIsTooFull())
435 afs_CacheTooFull = 0;
436 } /* end of cache cleanup */
437 ReleaseWriteLock(&afs_xdcache);
440 * This is a defensive check to try to avoid starving threads
441 * that may need the global lock so thay can help free some
442 * cache space. If this thread won't be sleeping or truncating
443 * any cache files then give up the global lock so other
444 * threads get a chance to run.
446 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
447 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
448 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
452 * This is where we free the discarded cache elements.
454 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
455 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
456 afs_FreeDiscardedDCache();
459 /* See if we need to continue to run. Someone may have
460 * signalled us while we were executing.
462 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
463 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
464 /* Collect statistics on truncate daemon. */
465 CTD_stats.CTD_nSleeps++;
466 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
467 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
468 CTD_stats.CTD_beforeSleep);
469 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
471 afs_TruncateDaemonRunning = 0;
472 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
473 afs_TruncateDaemonRunning = 1;
475 osi_GetuTime(&CTD_stats.CTD_afterSleep);
476 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
477 CTD_stats.CTD_afterSleep);
478 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
480 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
481 afs_termState = AFSOP_STOP_AFSDB;
482 afs_osi_Wakeup(&afs_termState);
490 * Make adjustment for the new size in the disk cache entry
492 * \note Major Assumptions Here:
493 * Assumes that frag size is an integral power of two, less one,
494 * and that this is a two's complement machine. I don't
495 * know of any filesystems which violate this assumption...
497 * \param adc Ptr to dcache entry.
498 * \param anewsize New size desired.
503 afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
505 register afs_int32 oldSize;
507 AFS_STATCNT(afs_AdjustSize);
509 adc->dflags |= DFEntryMod;
510 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
511 adc->f.chunkBytes = newSize;
514 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
515 afs_DCAdjustSize(adc, oldSize, newSize);
516 if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {
518 /* We're growing the file, wakeup the daemon */
519 afs_MaybeWakeupTruncateDaemon();
521 afs_blocksUsed += (newSize - oldSize);
522 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
527 * This routine is responsible for moving at least one entry (but up
528 * to some number of them) from the LRU queue to the free queue.
530 * \param anumber Number of entries that should ideally be moved.
531 * \param aneedSpace How much space we need (1K blocks);
534 * The anumber parameter is just a hint; at least one entry MUST be
535 * moved, or we'll panic. We must be called with afs_xdcache
536 * write-locked. We should try to satisfy both anumber and aneedspace,
537 * whichever is more demanding - need to do several things:
538 * 1. only grab up to anumber victims if aneedSpace <= 0, not
539 * the whole set of MAXATONCE.
540 * 2. dynamically choose MAXATONCE to reflect severity of
541 * demand: something like (*aneedSpace >> (logChunk - 9))
543 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
544 * indicates that the cache is not properly configured/tuned or
545 * something. We should be able to automatically correct that problem.
548 #define MAXATONCE 16 /* max we can obtain at once */
550 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
554 struct VenusFid *afid;
558 register struct vcache *tvc;
559 afs_uint32 victims[MAXATONCE];
560 struct dcache *victimDCs[MAXATONCE];
561 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
562 afs_uint32 victimPtr; /* next free item in victim arrays */
563 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
564 afs_uint32 maxVictimPtr; /* where it is */
567 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
571 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
572 vfslocked = VFS_LOCK_GIANT(afs_globalVFS);
575 AFS_STATCNT(afs_GetDownD);
577 if (CheckLock(&afs_xdcache) != -1)
578 osi_Panic("getdownd nolock");
579 /* decrement anumber first for all dudes in free list */
580 /* SHOULD always decrement anumber first, even if aneedSpace >0,
581 * because we should try to free space even if anumber <=0 */
582 if (!aneedSpace || *aneedSpace <= 0) {
583 anumber -= afs_freeDCCount;
585 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
586 VFS_UNLOCK_GIANT(vfslocked);
588 return; /* enough already free */
592 /* bounds check parameter */
593 if (anumber > MAXATONCE)
594 anumber = MAXATONCE; /* all we can do */
596 /* rewrite so phases include a better eligiblity for gc test*/
598 * The phase variable manages reclaims. Set to 0, the first pass,
599 * we don't reclaim active entries, or other than target bucket.
600 * Set to 1, we reclaim even active ones in target bucket.
601 * Set to 2, we reclaim any inactive one.
602 * Set to 3, we reclaim even active ones.
610 for (i = 0; i < afs_cacheFiles; i++)
611 /* turn off all flags */
612 afs_indexFlags[i] &= ~IFFlag;
614 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
615 /* find oldest entries for reclamation */
616 maxVictimPtr = victimPtr = 0;
617 hzero(maxVictimTime);
618 curbucket = afs_DCWhichBucket(phase, buckethint);
619 /* select victims from access time array */
620 for (i = 0; i < afs_cacheFiles; i++) {
621 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
622 /* skip if dirty or already free */
625 tdc = afs_indexTable[i];
626 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
628 /* Wrong bucket; can't use it! */
631 if (tdc && (tdc->refCount != 0)) {
632 /* Referenced; can't use it! */
635 hset(vtime, afs_indexTimes[i]);
637 /* if we've already looked at this one, skip it */
638 if (afs_indexFlags[i] & IFFlag)
641 if (victimPtr < MAXATONCE) {
642 /* if there's at least one free victim slot left */
643 victims[victimPtr] = i;
644 hset(victimTimes[victimPtr], vtime);
645 if (hcmp(vtime, maxVictimTime) > 0) {
646 hset(maxVictimTime, vtime);
647 maxVictimPtr = victimPtr;
650 } else if (hcmp(vtime, maxVictimTime) < 0) {
652 * We're older than youngest victim, so we replace at
655 /* find youngest (largest LRU) victim */
658 osi_Panic("getdownd local");
660 hset(victimTimes[j], vtime);
661 /* recompute maxVictimTime */
662 hset(maxVictimTime, vtime);
663 for (j = 0; j < victimPtr; j++)
664 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
665 hset(maxVictimTime, victimTimes[j]);
671 /* now really reclaim the victims */
672 j = 0; /* flag to track if we actually got any of the victims */
673 /* first, hold all the victims, since we're going to release the lock
674 * during the truncate operation.
676 for (i = 0; i < victimPtr; i++) {
677 tdc = afs_GetDSlot(victims[i], 0);
678 /* We got tdc->tlock(R) here */
679 if (tdc->refCount == 1)
683 ReleaseReadLock(&tdc->tlock);
687 for (i = 0; i < victimPtr; i++) {
688 /* q is first elt in dcache entry */
690 /* now, since we're dropping the afs_xdcache lock below, we
691 * have to verify, before proceeding, that there are no other
692 * references to this dcache entry, even now. Note that we
693 * compare with 1, since we bumped it above when we called
694 * afs_GetDSlot to preserve the entry's identity.
696 if (tdc && tdc->refCount == 1) {
697 unsigned char chunkFlags;
698 afs_size_t tchunkoffset = 0;
700 /* xdcache is lower than the xvcache lock */
701 ReleaseWriteLock(&afs_xdcache);
702 ObtainReadLock(&afs_xvcache);
703 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
704 ReleaseReadLock(&afs_xvcache);
705 ObtainWriteLock(&afs_xdcache, 527);
707 if (tdc->refCount > 1)
710 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
711 chunkFlags = afs_indexFlags[tdc->index];
712 if (((phase & 1) == 0) && osi_Active(tvc))
714 if (((phase & 1) == 1) && osi_Active(tvc)
715 && (tvc->f.states & CDCLock)
716 && (chunkFlags & IFAnyPages))
718 if (chunkFlags & IFDataMod)
720 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
721 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
722 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
723 ICL_HANDLE_OFFSET(tchunkoffset));
725 #if defined(AFS_SUN5_ENV)
727 * Now we try to invalidate pages. We do this only for
728 * Solaris. For other platforms, it's OK to recycle a
729 * dcache entry out from under a page, because the strategy
730 * function can call afs_GetDCache().
732 if (!skip && (chunkFlags & IFAnyPages)) {
735 ReleaseWriteLock(&afs_xdcache);
736 ObtainWriteLock(&tvc->vlock, 543);
737 if (tvc->multiPage) {
741 /* block locking pages */
742 tvc->vstates |= VPageCleaning;
743 /* block getting new pages */
745 ReleaseWriteLock(&tvc->vlock);
746 /* One last recheck */
747 ObtainWriteLock(&afs_xdcache, 333);
748 chunkFlags = afs_indexFlags[tdc->index];
749 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
750 || (osi_Active(tvc) && (tvc->f.states & CDCLock)
751 && (chunkFlags & IFAnyPages))) {
753 ReleaseWriteLock(&afs_xdcache);
756 ReleaseWriteLock(&afs_xdcache);
758 code = osi_VM_GetDownD(tvc, tdc);
760 ObtainWriteLock(&afs_xdcache, 269);
761 /* we actually removed all pages, clean and dirty */
763 afs_indexFlags[tdc->index] &=
764 ~(IFDirtyPages | IFAnyPages);
767 ReleaseWriteLock(&afs_xdcache);
769 ObtainWriteLock(&tvc->vlock, 544);
770 if (--tvc->activeV == 0
771 && (tvc->vstates & VRevokeWait)) {
772 tvc->vstates &= ~VRevokeWait;
773 afs_osi_Wakeup((char *)&tvc->vstates);
776 if (tvc->vstates & VPageCleaning) {
777 tvc->vstates &= ~VPageCleaning;
778 afs_osi_Wakeup((char *)&tvc->vstates);
781 ReleaseWriteLock(&tvc->vlock);
783 #endif /* AFS_SUN5_ENV */
785 ReleaseWriteLock(&afs_xdcache);
788 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
789 ObtainWriteLock(&afs_xdcache, 528);
790 if (afs_indexFlags[tdc->index] &
791 (IFDataMod | IFDirtyPages | IFAnyPages))
793 if (tdc->refCount > 1)
796 #if defined(AFS_SUN5_ENV)
798 /* no vnode, so IFDirtyPages is spurious (we don't
799 * sweep dcaches on vnode recycling, so we can have
800 * DIRTYPAGES set even when all pages are gone). Just
802 * Hold vcache lock to prevent vnode from being
803 * created while we're clearing IFDirtyPages.
805 afs_indexFlags[tdc->index] &=
806 ~(IFDirtyPages | IFAnyPages);
810 /* skip this guy and mark him as recently used */
811 afs_indexFlags[tdc->index] |= IFFlag;
812 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
813 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
814 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
815 ICL_HANDLE_OFFSET(tchunkoffset));
817 /* flush this dude from the data cache and reclaim;
818 * first, make sure no one will care that we damage
819 * it, by removing it from all hash tables. Then,
820 * melt it down for parts. Note that any concurrent
821 * (new possibility!) calls to GetDownD won't touch
822 * this guy because his reference count is > 0. */
823 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
824 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
825 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
826 ICL_HANDLE_OFFSET(tchunkoffset));
827 AFS_STATCNT(afs_gget);
828 afs_HashOutDCache(tdc, 1);
829 if (tdc->f.chunkBytes != 0) {
833 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
838 afs_DiscardDCache(tdc);
843 j = 1; /* we reclaimed at least one victim */
847 } /* end of for victims loop */
850 /* Phase is 0 and no one was found, so try phase 1 (ignore
851 * osi_Active flag) */
854 for (i = 0; i < afs_cacheFiles; i++)
855 /* turn off all flags */
856 afs_indexFlags[i] &= ~IFFlag;
859 /* found no one in phases 0-5, we're hosed */
863 } /* big while loop */
865 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
866 VFS_UNLOCK_GIANT(vfslocked);
875 * Remove adc from any hash tables that would allow it to be located
876 * again by afs_FindDCache or afs_GetDCache.
878 * \param adc Pointer to dcache entry to remove from hash tables.
880 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
884 afs_HashOutDCache(struct dcache *adc, int zap)
888 AFS_STATCNT(afs_glink);
890 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
892 /* if this guy is in the hash table, pull him out */
893 if (adc->f.fid.Fid.Volume != 0) {
894 /* remove entry from first hash chains */
895 i = DCHash(&adc->f.fid, adc->f.chunk);
896 us = afs_dchashTbl[i];
897 if (us == adc->index) {
898 /* first dude in the list */
899 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
901 /* somewhere on the chain */
902 while (us != NULLIDX) {
903 if (afs_dcnextTbl[us] == adc->index) {
904 /* found item pointing at the one to delete */
905 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
908 us = afs_dcnextTbl[us];
911 osi_Panic("dcache hc");
913 /* remove entry from *other* hash chain */
914 i = DVHash(&adc->f.fid);
915 us = afs_dvhashTbl[i];
916 if (us == adc->index) {
917 /* first dude in the list */
918 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
920 /* somewhere on the chain */
921 while (us != NULLIDX) {
922 if (afs_dvnextTbl[us] == adc->index) {
923 /* found item pointing at the one to delete */
924 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
927 us = afs_dvnextTbl[us];
930 osi_Panic("dcache hv");
935 /* prevent entry from being found on a reboot (it is already out of
936 * the hash table, but after a crash, we just look at fid fields of
937 * stable (old) entries).
939 adc->f.fid.Fid.Volume = 0; /* invalid */
941 /* mark entry as modified */
942 adc->dflags |= DFEntryMod;
947 } /*afs_HashOutDCache */
950 * Flush the given dcache entry, pulling it from hash chains
951 * and truncating the associated cache file.
953 * \param adc Ptr to dcache entry to flush.
956 * This routine must be called with the afs_xdcache lock held
960 afs_FlushDCache(register struct dcache *adc)
962 AFS_STATCNT(afs_FlushDCache);
964 * Bump the number of cache files flushed.
966 afs_stats_cmperf.cacheFlushes++;
968 /* remove from all hash tables */
969 afs_HashOutDCache(adc, 1);
971 /* Free its space; special case null operation, since truncate operation
972 * in UFS is slow even in this case, and this allows us to pre-truncate
973 * these files at more convenient times with fewer locks set
974 * (see afs_GetDownD).
976 if (adc->f.chunkBytes != 0) {
977 afs_DiscardDCache(adc);
978 afs_MaybeWakeupTruncateDaemon();
983 if (afs_WaitForCacheDrain) {
984 if (afs_blocksUsed <=
985 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
986 afs_WaitForCacheDrain = 0;
987 afs_osi_Wakeup(&afs_WaitForCacheDrain);
990 } /*afs_FlushDCache */
994 * Put a dcache entry on the free dcache entry list.
996 * \param adc dcache entry to free.
998 * \note Environment: called with afs_xdcache lock write-locked.
1001 afs_FreeDCache(register struct dcache *adc)
1003 /* Thread on free list, update free list count and mark entry as
1004 * freed in its indexFlags element. Also, ensure DCache entry gets
1005 * written out (set DFEntryMod).
1008 afs_dvnextTbl[adc->index] = afs_freeDCList;
1009 afs_freeDCList = adc->index;
1011 afs_indexFlags[adc->index] |= IFFree;
1012 adc->dflags |= DFEntryMod;
1014 if (afs_WaitForCacheDrain) {
1015 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1016 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1017 afs_WaitForCacheDrain = 0;
1018 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1021 } /* afs_FreeDCache */
1024 * Discard the cache element by moving it to the discardDCList.
1025 * This puts the cache element into a quasi-freed state, where
1026 * the space may be reused, but the file has not been truncated.
1028 * \note Major Assumptions Here:
1029 * Assumes that frag size is an integral power of two, less one,
1030 * and that this is a two's complement machine. I don't
1031 * know of any filesystems which violate this assumption...
1033 * \param adr Ptr to dcache entry.
1035 * \note Environment:
1036 * Must be called with afs_xdcache write-locked.
1040 afs_DiscardDCache(register struct dcache *adc)
1042 register afs_int32 size;
1044 AFS_STATCNT(afs_DiscardDCache);
1046 osi_Assert(adc->refCount == 1);
1048 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1049 afs_blocksDiscarded += size;
1050 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1052 afs_dvnextTbl[adc->index] = afs_discardDCList;
1053 afs_discardDCList = adc->index;
1054 afs_discardDCCount++;
1056 adc->f.fid.Fid.Volume = 0;
1057 adc->dflags |= DFEntryMod;
1058 afs_indexFlags[adc->index] |= IFDiscarded;
1060 if (afs_WaitForCacheDrain) {
1061 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1062 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1063 afs_WaitForCacheDrain = 0;
1064 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1068 } /*afs_DiscardDCache */
1071 * Free the next element on the list of discarded cache elements.
1074 afs_FreeDiscardedDCache(void)
1076 register struct dcache *tdc;
1077 register struct osi_file *tfile;
1078 register afs_int32 size;
1080 AFS_STATCNT(afs_FreeDiscardedDCache);
1082 ObtainWriteLock(&afs_xdcache, 510);
1083 if (!afs_blocksDiscarded) {
1084 ReleaseWriteLock(&afs_xdcache);
1089 * Get an entry from the list of discarded cache elements
1091 tdc = afs_GetDSlot(afs_discardDCList, 0);
1092 osi_Assert(tdc->refCount == 1);
1093 ReleaseReadLock(&tdc->tlock);
1095 afs_discardDCList = afs_dvnextTbl[tdc->index];
1096 afs_dvnextTbl[tdc->index] = NULLIDX;
1097 afs_discardDCCount--;
1098 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1099 afs_blocksDiscarded -= size;
1100 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1101 /* We can lock because we just took it off the free list */
1102 ObtainWriteLock(&tdc->lock, 626);
1103 ReleaseWriteLock(&afs_xdcache);
1106 * Truncate the element to reclaim its space
1108 tfile = afs_CFileOpen(&tdc->f.inode);
1109 afs_CFileTruncate(tfile, 0);
1110 afs_CFileClose(tfile);
1111 afs_AdjustSize(tdc, 0);
1112 afs_DCMoveBucket(tdc, 0, 0);
1115 * Free the element we just truncated
1117 ObtainWriteLock(&afs_xdcache, 511);
1118 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1119 afs_FreeDCache(tdc);
1120 tdc->f.states &= ~(DRO|DBackup|DRW);
1121 ReleaseWriteLock(&tdc->lock);
1123 ReleaseWriteLock(&afs_xdcache);
1127 * Free as many entries from the list of discarded cache elements
1128 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1133 afs_MaybeFreeDiscardedDCache(void)
1136 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1138 while (afs_blocksDiscarded
1139 && (afs_blocksUsed >
1140 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1141 afs_FreeDiscardedDCache();
1147 * Try to free up a certain number of disk slots.
1149 * \param anumber Targeted number of disk slots to free up.
1151 * \note Environment:
1152 * Must be called with afs_xdcache write-locked.
1156 afs_GetDownDSlot(int anumber)
1158 struct afs_q *tq, *nq;
1163 AFS_STATCNT(afs_GetDownDSlot);
1164 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1165 osi_Panic("diskless getdowndslot");
1167 if (CheckLock(&afs_xdcache) != -1)
1168 osi_Panic("getdowndslot nolock");
1170 /* decrement anumber first for all dudes in free list */
1171 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1174 return; /* enough already free */
1176 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1178 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1179 nq = QPrev(tq); /* in case we remove it */
1180 if (tdc->refCount == 0) {
1181 if ((ix = tdc->index) == NULLIDX)
1182 osi_Panic("getdowndslot");
1183 /* pull the entry out of the lruq and put it on the free list */
1184 QRemove(&tdc->lruq);
1186 /* write-through if modified */
1187 if (tdc->dflags & DFEntryMod) {
1188 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1190 * ask proxy to do this for us - we don't have the stack space
1192 while (tdc->dflags & DFEntryMod) {
1195 s = SPLOCK(afs_sgibklock);
1196 if (afs_sgibklist == NULL) {
1197 /* if slot is free, grab it. */
1198 afs_sgibklist = tdc;
1199 SV_SIGNAL(&afs_sgibksync);
1201 /* wait for daemon to (start, then) finish. */
1202 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1206 tdc->dflags &= ~DFEntryMod;
1207 afs_WriteDCache(tdc, 1);
1211 /* finally put the entry in the free list */
1212 afs_indexTable[ix] = NULL;
1213 afs_indexFlags[ix] &= ~IFEverUsed;
1214 tdc->index = NULLIDX;
1215 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1216 afs_freeDSList = tdc;
1220 } /*afs_GetDownDSlot */
1227 * Increment the reference count on a disk cache entry,
1228 * which already has a non-zero refcount. In order to
1229 * increment the refcount of a zero-reference entry, you
1230 * have to hold afs_xdcache.
1233 * adc : Pointer to the dcache entry to increment.
1236 * Nothing interesting.
1239 afs_RefDCache(struct dcache *adc)
1241 ObtainWriteLock(&adc->tlock, 627);
1242 if (adc->refCount < 0)
1243 osi_Panic("RefDCache: negative refcount");
1245 ReleaseWriteLock(&adc->tlock);
1254 * Decrement the reference count on a disk cache entry.
1257 * ad : Ptr to the dcache entry to decrement.
1260 * Nothing interesting.
1263 afs_PutDCache(register struct dcache *adc)
1265 AFS_STATCNT(afs_PutDCache);
1266 ObtainWriteLock(&adc->tlock, 276);
1267 if (adc->refCount <= 0)
1268 osi_Panic("putdcache");
1270 ReleaseWriteLock(&adc->tlock);
1279 * Try to discard all data associated with this file from the
1283 * avc : Pointer to the cache info for the file.
1286 * Both pvnLock and lock are write held.
1289 afs_TryToSmush(register struct vcache *avc, afs_ucred_t *acred, int sync)
1291 register struct dcache *tdc;
1294 AFS_STATCNT(afs_TryToSmush);
1295 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1296 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
1297 sync = 1; /* XX Temp testing XX */
1299 #if defined(AFS_SUN5_ENV)
1300 ObtainWriteLock(&avc->vlock, 573);
1301 avc->activeV++; /* block new getpages */
1302 ReleaseWriteLock(&avc->vlock);
1305 /* Flush VM pages */
1306 osi_VM_TryToSmush(avc, acred, sync);
1309 * Get the hash chain containing all dce's for this fid
1311 i = DVHash(&avc->f.fid);
1312 ObtainWriteLock(&afs_xdcache, 277);
1313 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1314 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1315 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1316 int releaseTlock = 1;
1317 tdc = afs_GetDSlot(index, NULL);
1318 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1320 if ((afs_indexFlags[index] & IFDataMod) == 0
1321 && tdc->refCount == 1) {
1322 ReleaseReadLock(&tdc->tlock);
1324 afs_FlushDCache(tdc);
1327 afs_indexTable[index] = 0;
1330 ReleaseReadLock(&tdc->tlock);
1334 #if defined(AFS_SUN5_ENV)
1335 ObtainWriteLock(&avc->vlock, 545);
1336 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1337 avc->vstates &= ~VRevokeWait;
1338 afs_osi_Wakeup((char *)&avc->vstates);
1340 ReleaseWriteLock(&avc->vlock);
1342 ReleaseWriteLock(&afs_xdcache);
1344 * It's treated like a callback so that when we do lookups we'll
1345 * invalidate the unique bit if any
1346 * trytoSmush occured during the lookup call
1352 * afs_DCacheMissingChunks
1355 * Given the cached info for a file, return the number of chunks that
1356 * are not available from the dcache.
1359 * avc: Pointer to the (held) vcache entry to look in.
1362 * The number of chunks which are not currently cached.
1365 * The vcache entry is held upon entry.
1369 afs_DCacheMissingChunks(struct vcache *avc)
1372 afs_size_t totalLength = 0;
1373 afs_uint32 totalChunks = 0;
1376 totalLength = avc->f.m.Length;
1377 if (avc->f.truncPos < totalLength)
1378 totalLength = avc->f.truncPos;
1380 /* Length is 0, no chunk missing. */
1381 if (totalLength == 0)
1384 /* If totalLength is a multiple of chunksize, the last byte appears
1385 * as being part of the next chunk, which does not exist.
1386 * Decrementing totalLength by one fixes that.
1389 totalChunks = (AFS_CHUNK(totalLength) + 1);
1391 /* If we're a directory, we only ever have one chunk, regardless of
1392 * the size of the dir.
1394 if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
1398 printf("Should have %d chunks for %u bytes\n",
1399 totalChunks, (totalLength + 1));
1401 i = DVHash(&avc->f.fid);
1402 ObtainWriteLock(&afs_xdcache, 1001);
1403 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1404 i = afs_dvnextTbl[index];
1405 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1406 tdc = afs_GetDSlot(index, NULL);
1407 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1410 ReleaseReadLock(&tdc->tlock);
1414 ReleaseWriteLock(&afs_xdcache);
1416 /*printf("Missing %d chunks\n", totalChunks);*/
1418 return (totalChunks);
1425 * Given the cached info for a file and a byte offset into the
1426 * file, make sure the dcache entry for that file and containing
1427 * the given byte is available, returning it to our caller.
1430 * avc : Pointer to the (held) vcache entry to look in.
1431 * abyte : Which byte we want to get to.
1434 * Pointer to the dcache entry covering the file & desired byte,
1435 * or NULL if not found.
1438 * The vcache entry is held upon entry.
1442 afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1445 register afs_int32 i, index;
1446 register struct dcache *tdc = NULL;
1448 AFS_STATCNT(afs_FindDCache);
1449 chunk = AFS_CHUNK(abyte);
1452 * Hash on the [fid, chunk] and get the corresponding dcache index
1453 * after write-locking the dcache.
1455 i = DCHash(&avc->f.fid, chunk);
1456 ObtainWriteLock(&afs_xdcache, 278);
1457 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1458 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1459 tdc = afs_GetDSlot(index, NULL);
1460 ReleaseReadLock(&tdc->tlock);
1461 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1462 break; /* leaving refCount high for caller */
1466 index = afs_dcnextTbl[index];
1468 if (index != NULLIDX) {
1469 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1470 hadd32(afs_indexCounter, 1);
1471 ReleaseWriteLock(&afs_xdcache);
1474 ReleaseWriteLock(&afs_xdcache);
1476 } /*afs_FindDCache */
1480 * Get a fresh dcache from the free or discarded list.
1482 * \param avc Who's dcache is this going to be?
1483 * \param chunk The position where it will be placed in.
1484 * \param lock How are locks held.
1485 * \param ashFid If this dcache going to be used for a shadow dir,
1488 * \note Required locks:
1490 * - avc (R if (lock & 1) set and W otherwise)
1491 * \note It write locks the new dcache. The caller must unlock it.
1493 * \return The new dcache.
1495 struct dcache *afs_AllocDCache(struct vcache *avc,
1498 struct VenusFid *ashFid)
1500 struct dcache *tdc = NULL;
1501 afs_uint32 size = 0;
1502 struct osi_file *file;
1504 if (afs_discardDCList == NULLIDX
1505 || ((lock & 2) && afs_freeDCList != NULLIDX)) {
1507 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1508 tdc = afs_GetDSlot(afs_freeDCList, 0);
1509 osi_Assert(tdc->refCount == 1);
1510 ReleaseReadLock(&tdc->tlock);
1511 ObtainWriteLock(&tdc->lock, 604);
1512 afs_freeDCList = afs_dvnextTbl[tdc->index];
1515 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1516 tdc = afs_GetDSlot(afs_discardDCList, 0);
1517 osi_Assert(tdc->refCount == 1);
1518 ReleaseReadLock(&tdc->tlock);
1519 ObtainWriteLock(&tdc->lock, 605);
1520 afs_discardDCList = afs_dvnextTbl[tdc->index];
1521 afs_discardDCCount--;
1523 ((tdc->f.chunkBytes +
1524 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1525 tdc->f.states &= ~(DRO|DBackup|DRW);
1526 afs_DCMoveBucket(tdc, size, 0);
1527 afs_blocksDiscarded -= size;
1528 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1530 /* Truncate the chunk so zeroes get filled properly */
1531 file = afs_CFileOpen(&tdc->f.inode);
1532 afs_CFileTruncate(file, 0);
1533 afs_CFileClose(file);
1534 afs_AdjustSize(tdc, 0);
1540 * avc->lock(R) if setLocks
1541 * avc->lock(W) if !setLocks
1547 * Fill in the newly-allocated dcache record.
1549 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1551 /* Use shadow fid if provided. */
1552 tdc->f.fid = *ashFid;
1554 /* Use normal vcache's fid otherwise. */
1555 tdc->f.fid = avc->f.fid;
1556 if (avc->f.states & CRO)
1557 tdc->f.states = DRO;
1558 else if (avc->f.states & CBackup)
1559 tdc->f.states = DBackup;
1561 tdc->f.states = DRW;
1562 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1563 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1565 hones(tdc->f.versionNo); /* invalid value */
1566 tdc->f.chunk = chunk;
1567 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1569 if (tdc->lruq.prev == &tdc->lruq)
1570 osi_Panic("lruq 1");
1579 * This function is called to obtain a reference to data stored in
1580 * the disk cache, locating a chunk of data containing the desired
1581 * byte and returning a reference to the disk cache entry, with its
1582 * reference count incremented.
1586 * avc : Ptr to a vcache entry (unlocked)
1587 * abyte : Byte position in the file desired
1588 * areq : Request structure identifying the requesting user.
1589 * aflags : Settings as follows:
1591 * 2 : Return after creating entry.
1592 * 4 : called from afs_vnop_write.c
1593 * *alen contains length of data to be written.
1595 * aoffset : Set to the offset within the chunk where the resident
1597 * alen : Set to the number of bytes of data after the desired
1598 * byte (including the byte itself) which can be read
1602 * The vcache entry pointed to by avc is unlocked upon entry.
1606 * Update the vnode-to-dcache hint if we can get the vnode lock
1607 * right away. Assumes dcache entry is at least read-locked.
1610 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1612 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1613 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1616 ReleaseWriteLock(&v->lock);
1620 /* avc - Write-locked unless aflags & 1 */
1622 afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1623 register struct vrequest *areq, afs_size_t * aoffset,
1624 afs_size_t * alen, int aflags)
1626 register afs_int32 i, code, shortcut;
1627 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1628 register afs_int32 adjustsize = 0;
1634 afs_size_t maxGoodLength; /* amount of good data at server */
1635 afs_size_t Position = 0;
1636 afs_int32 size, tlen; /* size of segment to transfer */
1637 struct afs_FetchOutput *tsmall = 0;
1638 register struct dcache *tdc;
1639 register struct osi_file *file;
1640 register struct afs_conn *tc;
1642 struct server *newCallback = NULL;
1643 char setNewCallback;
1644 char setVcacheStatus;
1645 char doVcacheUpdate;
1647 int doAdjustSize = 0;
1648 int doReallyAdjustSize = 0;
1649 int overWriteWholeChunk = 0;
1652 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1653 int fromReplica; /*Are we reading from a replica? */
1654 int numFetchLoops; /*# times around the fetch/analyze loop */
1655 #endif /* AFS_NOSTATS */
1657 AFS_STATCNT(afs_GetDCache);
1661 setLocks = aflags & 1;
1664 * Determine the chunk number and offset within the chunk corresponding
1665 * to the desired byte.
1667 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1670 chunk = AFS_CHUNK(abyte);
1673 /* come back to here if we waited for the cache to drain. */
1676 setNewCallback = setVcacheStatus = 0;
1680 ObtainWriteLock(&avc->lock, 616);
1682 ObtainReadLock(&avc->lock);
1687 * avc->lock(R) if setLocks && !slowPass
1688 * avc->lock(W) if !setLocks || slowPass
1693 /* check hints first! (might could use bcmp or some such...) */
1694 if ((tdc = avc->dchint)) {
1698 * The locking order between afs_xdcache and dcache lock matters.
1699 * The hint dcache entry could be anywhere, even on the free list.
1700 * Locking afs_xdcache ensures that noone is trying to pull dcache
1701 * entries from the free list, and thereby assuming them to be not
1702 * referenced and not locked.
1704 ObtainReadLock(&afs_xdcache);
1705 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1707 if (dcLocked && (tdc->index != NULLIDX)
1708 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1709 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1710 /* got the right one. It might not be the right version, and it
1711 * might be fetching, but it's the right dcache entry.
1713 /* All this code should be integrated better with what follows:
1714 * I can save a good bit more time under a write lock if I do..
1716 ObtainWriteLock(&tdc->tlock, 603);
1718 ReleaseWriteLock(&tdc->tlock);
1720 ReleaseReadLock(&afs_xdcache);
1723 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1724 && !(tdc->dflags & DFFetching)) {
1726 afs_stats_cmperf.dcacheHits++;
1727 ObtainWriteLock(&afs_xdcache, 559);
1728 QRemove(&tdc->lruq);
1729 QAdd(&afs_DLRU, &tdc->lruq);
1730 ReleaseWriteLock(&afs_xdcache);
1733 * avc->lock(R) if setLocks && !slowPass
1734 * avc->lock(W) if !setLocks || slowPass
1741 ReleaseSharedLock(&tdc->lock);
1742 ReleaseReadLock(&afs_xdcache);
1750 * avc->lock(R) if setLocks && !slowPass
1751 * avc->lock(W) if !setLocks || slowPass
1752 * tdc->lock(S) if tdc
1755 if (!tdc) { /* If the hint wasn't the right dcache entry */
1757 * Hash on the [fid, chunk] and get the corresponding dcache index
1758 * after write-locking the dcache.
1763 * avc->lock(R) if setLocks && !slowPass
1764 * avc->lock(W) if !setLocks || slowPass
1767 i = DCHash(&avc->f.fid, chunk);
1768 /* check to make sure our space is fine */
1769 afs_MaybeWakeupTruncateDaemon();
1771 ObtainWriteLock(&afs_xdcache, 280);
1773 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1774 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1775 tdc = afs_GetDSlot(index, NULL);
1776 ReleaseReadLock(&tdc->tlock);
1779 * avc->lock(R) if setLocks && !slowPass
1780 * avc->lock(W) if !setLocks || slowPass
1783 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1784 /* Move it up in the beginning of the list */
1785 if (afs_dchashTbl[i] != index) {
1786 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1787 afs_dcnextTbl[index] = afs_dchashTbl[i];
1788 afs_dchashTbl[i] = index;
1790 ReleaseWriteLock(&afs_xdcache);
1791 ObtainSharedLock(&tdc->lock, 606);
1792 break; /* leaving refCount high for caller */
1798 index = afs_dcnextTbl[index];
1802 * If we didn't find the entry, we'll create one.
1804 if (index == NULLIDX) {
1807 * avc->lock(R) if setLocks
1808 * avc->lock(W) if !setLocks
1811 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1812 avc, ICL_TYPE_INT32, chunk);
1814 /* Make sure there is a free dcache entry for us to use */
1815 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1818 avc->f.states |= CDCLock;
1819 /* just need slots */
1820 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1822 avc->f.states &= ~CDCLock;
1823 if (afs_discardDCList != NULLIDX
1824 || afs_freeDCList != NULLIDX)
1826 /* If we can't get space for 5 mins we give up and panic */
1827 if (++downDCount > 300) {
1828 #if defined(AFS_CACHE_BYPASS)
1829 afs_warn("GetDCache calling osi_Panic: No space in five minutes.\n downDCount: %d\n aoffset: %d alen: %d\n", downDCount, aoffset, alen);
1831 osi_Panic("getdcache");
1833 ReleaseWriteLock(&afs_xdcache);
1836 * avc->lock(R) if setLocks
1837 * avc->lock(W) if !setLocks
1839 afs_osi_Wait(1000, 0, 0);
1844 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1847 * Now add to the two hash chains - note that i is still set
1848 * from the above DCHash call.
1850 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1851 afs_dchashTbl[i] = tdc->index;
1852 i = DVHash(&avc->f.fid);
1853 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1854 afs_dvhashTbl[i] = tdc->index;
1855 tdc->dflags = DFEntryMod;
1857 afs_MaybeWakeupTruncateDaemon();
1858 ReleaseWriteLock(&afs_xdcache);
1859 ConvertWToSLock(&tdc->lock);
1864 /* vcache->dcache hint failed */
1867 * avc->lock(R) if setLocks && !slowPass
1868 * avc->lock(W) if !setLocks || slowPass
1871 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1872 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1873 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1874 hgetlo(avc->f.m.DataVersion));
1876 * Here we have the entry in tdc, with its refCount incremented.
1877 * Note: we don't use the S-lock on avc; it costs concurrency when
1878 * storing a file back to the server.
1882 * Not a newly created file so we need to check the file's length and
1883 * compare data versions since someone could have changed the data or we're
1884 * reading a file written elsewhere. We only want to bypass doing no-op
1885 * read rpcs on newly created files (dv of 0) since only then we guarantee
1886 * that this chunk's data hasn't been filled by another client.
1888 size = AFS_CHUNKSIZE(abyte);
1889 if (aflags & 4) /* called from write */
1891 else /* called from read */
1892 tlen = tdc->validPos - abyte;
1893 Position = AFS_CHUNKTOBASE(chunk);
1894 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1895 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1896 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1897 ICL_HANDLE_OFFSET(Position));
1898 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
1900 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
1901 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
1902 overWriteWholeChunk = 1;
1903 if (doAdjustSize || overWriteWholeChunk) {
1904 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1906 #ifdef AFS_SGI64_ENV
1909 #else /* AFS_SGI64_ENV */
1912 #endif /* AFS_SGI64_ENV */
1913 #else /* AFS_SGI_ENV */
1916 #endif /* AFS_SGI_ENV */
1917 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
1918 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1919 #if defined(AFS_SUN5_ENV)
1920 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
1922 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
1924 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1925 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
1926 doReallyAdjustSize = 1;
1928 if (doReallyAdjustSize || overWriteWholeChunk) {
1929 /* no data in file to read at this position */
1930 UpgradeSToWLock(&tdc->lock, 607);
1931 file = afs_CFileOpen(&tdc->f.inode);
1932 afs_CFileTruncate(file, 0);
1933 afs_CFileClose(file);
1934 afs_AdjustSize(tdc, 0);
1935 hset(tdc->f.versionNo, avc->f.m.DataVersion);
1936 tdc->dflags |= DFEntryMod;
1938 ConvertWToSLock(&tdc->lock);
1943 * We must read in the whole chunk if the version number doesn't
1947 /* don't need data, just a unique dcache entry */
1948 ObtainWriteLock(&afs_xdcache, 608);
1949 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1950 hadd32(afs_indexCounter, 1);
1951 ReleaseWriteLock(&afs_xdcache);
1953 updateV2DC(setLocks, avc, tdc, 553);
1954 if (vType(avc) == VDIR)
1957 *aoffset = AFS_CHUNKOFFSET(abyte);
1958 if (tdc->validPos < abyte)
1959 *alen = (afs_size_t) 0;
1961 *alen = tdc->validPos - abyte;
1962 ReleaseSharedLock(&tdc->lock);
1965 ReleaseWriteLock(&avc->lock);
1967 ReleaseReadLock(&avc->lock);
1969 return tdc; /* check if we're done */
1974 * avc->lock(R) if setLocks && !slowPass
1975 * avc->lock(W) if !setLocks || slowPass
1978 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
1980 setNewCallback = setVcacheStatus = 0;
1984 * avc->lock(R) if setLocks && !slowPass
1985 * avc->lock(W) if !setLocks || slowPass
1988 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
1990 * Version number mismatch.
1993 * If we are disconnected, then we can't do much of anything
1994 * because the data doesn't match the file.
1996 if (AFS_IS_DISCONNECTED) {
1997 ReleaseSharedLock(&tdc->lock);
2000 ReleaseWriteLock(&avc->lock);
2002 ReleaseReadLock(&avc->lock);
2004 /* Flush the Dcache */
2009 UpgradeSToWLock(&tdc->lock, 609);
2012 * If data ever existed for this vnode, and this is a text object,
2013 * do some clearing. Now, you'd think you need only do the flush
2014 * when VTEXT is on, but VTEXT is turned off when the text object
2015 * is freed, while pages are left lying around in memory marked
2016 * with this vnode. If we would reactivate (create a new text
2017 * object from) this vnode, we could easily stumble upon some of
2018 * these old pages in pagein. So, we always flush these guys.
2019 * Sun has a wonderful lack of useful invariants in this system.
2021 * avc->flushDV is the data version # of the file at the last text
2022 * flush. Clearly, at least, we don't have to flush the file more
2023 * often than it changes
2025 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2027 * By here, the cache entry is always write-locked. We can
2028 * deadlock if we call osi_Flush with the cache entry locked...
2029 * Unlock the dcache too.
2031 ReleaseWriteLock(&tdc->lock);
2032 if (setLocks && !slowPass)
2033 ReleaseReadLock(&avc->lock);
2035 ReleaseWriteLock(&avc->lock);
2039 * Call osi_FlushPages in open, read/write, and map, since it
2040 * is too hard here to figure out if we should lock the
2043 if (setLocks && !slowPass)
2044 ObtainReadLock(&avc->lock);
2046 ObtainWriteLock(&avc->lock, 66);
2047 ObtainWriteLock(&tdc->lock, 610);
2052 * avc->lock(R) if setLocks && !slowPass
2053 * avc->lock(W) if !setLocks || slowPass
2057 /* Watch for standard race condition around osi_FlushText */
2058 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2059 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2060 afs_stats_cmperf.dcacheHits++;
2061 ConvertWToSLock(&tdc->lock);
2065 /* Sleep here when cache needs to be drained. */
2066 if (setLocks && !slowPass
2067 && (afs_blocksUsed >
2068 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2069 /* Make sure truncate daemon is running */
2070 afs_MaybeWakeupTruncateDaemon();
2071 ObtainWriteLock(&tdc->tlock, 614);
2072 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2073 ReleaseWriteLock(&tdc->tlock);
2074 ReleaseWriteLock(&tdc->lock);
2075 ReleaseReadLock(&avc->lock);
2076 while ((afs_blocksUsed - afs_blocksDiscarded) >
2077 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2078 afs_WaitForCacheDrain = 1;
2079 afs_osi_Sleep(&afs_WaitForCacheDrain);
2081 afs_MaybeFreeDiscardedDCache();
2082 /* need to check if someone else got the chunk first. */
2083 goto RetryGetDCache;
2086 /* Do not fetch data beyond truncPos. */
2087 maxGoodLength = avc->f.m.Length;
2088 if (avc->f.truncPos < maxGoodLength)
2089 maxGoodLength = avc->f.truncPos;
2090 Position = AFS_CHUNKBASE(abyte);
2091 if (vType(avc) == VDIR) {
2092 size = avc->f.m.Length;
2093 if (size > tdc->f.chunkBytes) {
2094 /* pre-reserve space for file */
2095 afs_AdjustSize(tdc, size);
2097 size = 999999999; /* max size for transfer */
2099 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2100 /* don't read past end of good data on server */
2101 if (Position + size > maxGoodLength)
2102 size = maxGoodLength - Position;
2104 size = 0; /* Handle random races */
2105 if (size > tdc->f.chunkBytes) {
2106 /* pre-reserve space for file */
2107 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2108 /* max size for transfer still in size */
2111 if (afs_mariner && !tdc->f.chunk)
2112 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2114 * Right now, we only have one tool, and it's a hammer. So, we
2115 * fetch the whole file.
2117 DZap(tdc); /* pages in cache may be old */
2118 file = afs_CFileOpen(&tdc->f.inode);
2119 afs_RemoveVCB(&avc->f.fid);
2120 tdc->f.states |= DWriting;
2121 tdc->dflags |= DFFetching;
2122 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2123 if (tdc->mflags & DFFetchReq) {
2124 tdc->mflags &= ~DFFetchReq;
2125 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2126 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2127 __FILE__, ICL_TYPE_INT32, __LINE__,
2128 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2132 (struct afs_FetchOutput *)osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2133 setVcacheStatus = 0;
2136 * Remember if we are doing the reading from a replicated volume,
2137 * and how many times we've zipped around the fetch/analyze loop.
2139 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2141 accP = &(afs_stats_cmfullperf.accessinf);
2143 (accP->replicatedRefs)++;
2145 (accP->unreplicatedRefs)++;
2146 #endif /* AFS_NOSTATS */
2147 /* this is a cache miss */
2148 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2149 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2150 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2153 afs_stats_cmperf.dcacheMisses++;
2156 * Dynamic root support: fetch data from local memory.
2158 if (afs_IsDynroot(avc)) {
2162 afs_GetDynroot(&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 */
2180 } else if (afs_IsDynrootMount(avc)) {
2184 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2186 dynrootDir += Position;
2187 dynrootLen -= Position;
2188 if (size > dynrootLen)
2192 code = afs_CFileWrite(file, 0, dynrootDir, size);
2200 tdc->validPos = Position + size;
2201 afs_CFileTruncate(file, size); /* prune it */
2204 * Not a dynamic vnode: do the real fetch.
2209 * avc->lock(R) if setLocks && !slowPass
2210 * avc->lock(W) if !setLocks || slowPass
2214 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK);
2219 (accP->numReplicasAccessed)++;
2221 #endif /* AFS_NOSTATS */
2222 if (!setLocks || slowPass) {
2223 avc->callback = tc->srvr->server;
2225 newCallback = tc->srvr->server;
2229 code = afs_CacheFetchProc(tc, file, Position, tdc,
2235 /* callback could have been broken (or expired) in a race here,
2236 * but we return the data anyway. It's as good as we knew about
2237 * when we started. */
2239 * validPos is updated by CacheFetchProc, and can only be
2240 * modifed under a dcache write lock, which we've blocked out
2242 size = tdc->validPos - Position; /* actual segment size */
2245 afs_CFileTruncate(file, size); /* prune it */
2247 if (!setLocks || slowPass) {
2248 ObtainWriteLock(&afs_xcbhash, 453);
2249 afs_DequeueCallback(avc);
2250 avc->f.states &= ~(CStatd | CUnique);
2251 avc->callback = NULL;
2252 ReleaseWriteLock(&afs_xcbhash);
2253 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2254 osi_dnlc_purgedp(avc);
2256 /* Something lost. Forget about performance, and go
2257 * back with a vcache write lock.
2259 afs_CFileTruncate(file, 0);
2260 afs_AdjustSize(tdc, 0);
2261 afs_CFileClose(file);
2262 osi_FreeLargeSpace(tsmall);
2264 ReleaseWriteLock(&tdc->lock);
2267 ReleaseReadLock(&avc->lock);
2269 goto RetryGetDCache;
2273 } while (afs_Analyze
2274 (tc, code, &avc->f.fid, areq,
2275 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2279 * avc->lock(R) if setLocks && !slowPass
2280 * avc->lock(W) if !setLocks || slowPass
2286 * In the case of replicated access, jot down info on the number of
2287 * attempts it took before we got through or gave up.
2290 if (numFetchLoops <= 1)
2291 (accP->refFirstReplicaOK)++;
2292 if (numFetchLoops > accP->maxReplicasPerRef)
2293 accP->maxReplicasPerRef = numFetchLoops;
2295 #endif /* AFS_NOSTATS */
2297 tdc->dflags &= ~DFFetching;
2298 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2299 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2300 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2301 tdc, ICL_TYPE_INT32, tdc->dflags);
2302 if (avc->execsOrWriters == 0)
2303 tdc->f.states &= ~DWriting;
2305 /* now, if code != 0, we have an error and should punt.
2306 * note that we have the vcache write lock, either because
2307 * !setLocks or slowPass.
2310 afs_CFileTruncate(file, 0);
2311 afs_AdjustSize(tdc, 0);
2312 afs_CFileClose(file);
2313 ZapDCE(tdc); /* sets DFEntryMod */
2314 if (vType(avc) == VDIR) {
2317 tdc->f.states &= ~(DRO|DBackup|DRW);
2318 afs_DCMoveBucket(tdc, 0, 0);
2319 ReleaseWriteLock(&tdc->lock);
2321 if (!afs_IsDynroot(avc)) {
2322 ObtainWriteLock(&afs_xcbhash, 454);
2323 afs_DequeueCallback(avc);
2324 avc->f.states &= ~(CStatd | CUnique);
2325 ReleaseWriteLock(&afs_xcbhash);
2326 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2327 osi_dnlc_purgedp(avc);
2330 * avc->lock(W); assert(!setLocks || slowPass)
2332 osi_Assert(!setLocks || slowPass);
2338 /* otherwise we copy in the just-fetched info */
2339 afs_CFileClose(file);
2340 afs_AdjustSize(tdc, size); /* new size */
2342 * Copy appropriate fields into vcache. Status is
2343 * copied later where we selectively acquire the
2344 * vcache write lock.
2347 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2349 setVcacheStatus = 1;
2350 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2351 tsmall->OutStatus.DataVersion);
2352 tdc->dflags |= DFEntryMod;
2353 afs_indexFlags[tdc->index] |= IFEverUsed;
2354 ConvertWToSLock(&tdc->lock);
2355 } /*Data version numbers don't match */
2358 * Data version numbers match.
2360 afs_stats_cmperf.dcacheHits++;
2361 } /*Data version numbers match */
2363 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2367 * avc->lock(R) if setLocks && !slowPass
2368 * avc->lock(W) if !setLocks || slowPass
2369 * tdc->lock(S) if tdc
2373 * See if this was a reference to a file in the local cell.
2375 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2376 afs_stats_cmperf.dlocalAccesses++;
2378 afs_stats_cmperf.dremoteAccesses++;
2380 /* Fix up LRU info */
2383 ObtainWriteLock(&afs_xdcache, 602);
2384 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2385 hadd32(afs_indexCounter, 1);
2386 ReleaseWriteLock(&afs_xdcache);
2388 /* return the data */
2389 if (vType(avc) == VDIR)
2392 *aoffset = AFS_CHUNKOFFSET(abyte);
2393 *alen = (tdc->f.chunkBytes - *aoffset);
2394 ReleaseSharedLock(&tdc->lock);
2399 * avc->lock(R) if setLocks && !slowPass
2400 * avc->lock(W) if !setLocks || slowPass
2403 /* Fix up the callback and status values in the vcache */
2405 if (setLocks && !slowPass) {
2408 * This is our dirty little secret to parallel fetches.
2409 * We don't write-lock the vcache while doing the fetch,
2410 * but potentially we'll need to update the vcache after
2411 * the fetch is done.
2413 * Drop the read lock and try to re-obtain the write
2414 * lock. If the vcache still has the same DV, it's
2415 * ok to go ahead and install the new data.
2417 afs_hyper_t currentDV, statusDV;
2419 hset(currentDV, avc->f.m.DataVersion);
2421 if (setNewCallback && avc->callback != newCallback)
2425 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2426 tsmall->OutStatus.DataVersion);
2428 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2430 if (setVcacheStatus && !hsame(currentDV, statusDV))
2434 ReleaseReadLock(&avc->lock);
2436 if (doVcacheUpdate) {
2437 ObtainWriteLock(&avc->lock, 615);
2438 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2439 /* We lose. Someone will beat us to it. */
2441 ReleaseWriteLock(&avc->lock);
2446 /* With slow pass, we've already done all the updates */
2448 ReleaseWriteLock(&avc->lock);
2451 /* Check if we need to perform any last-minute fixes with a write-lock */
2452 if (!setLocks || doVcacheUpdate) {
2454 avc->callback = newCallback;
2455 if (tsmall && setVcacheStatus)
2456 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2458 ReleaseWriteLock(&avc->lock);
2462 osi_FreeLargeSpace(tsmall);
2465 } /*afs_GetDCache */
2469 * afs_WriteThroughDSlots
2472 * Sweep through the dcache slots and write out any modified
2473 * in-memory data back on to our caching store.
2479 * The afs_xdcache is write-locked through this whole affair.
2482 afs_WriteThroughDSlots(void)
2484 register struct dcache *tdc;
2485 register afs_int32 i, touchedit = 0;
2487 struct afs_q DirtyQ, *tq;
2489 AFS_STATCNT(afs_WriteThroughDSlots);
2492 * Because of lock ordering, we can't grab dcache locks while
2493 * holding afs_xdcache. So we enter xdcache, get a reference
2494 * for every dcache entry, and exit xdcache.
2496 ObtainWriteLock(&afs_xdcache, 283);
2498 for (i = 0; i < afs_cacheFiles; i++) {
2499 tdc = afs_indexTable[i];
2501 /* Grab tlock in case the existing refcount isn't zero */
2502 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2503 ObtainWriteLock(&tdc->tlock, 623);
2505 ReleaseWriteLock(&tdc->tlock);
2507 QAdd(&DirtyQ, &tdc->dirty);
2510 ReleaseWriteLock(&afs_xdcache);
2513 * Now, for each dcache entry we found, check if it's dirty.
2514 * If so, get write-lock, get afs_xdcache, which protects
2515 * afs_cacheInodep, and flush it. Don't forget to put back
2519 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2521 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2523 if (tdc->dflags & DFEntryMod) {
2526 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2528 /* Now that we have the write lock, double-check */
2529 if (wrLock && (tdc->dflags & DFEntryMod)) {
2530 tdc->dflags &= ~DFEntryMod;
2531 ObtainWriteLock(&afs_xdcache, 620);
2532 afs_WriteDCache(tdc, 1);
2533 ReleaseWriteLock(&afs_xdcache);
2537 ReleaseWriteLock(&tdc->lock);
2543 ObtainWriteLock(&afs_xdcache, 617);
2544 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2545 /* Touch the file to make sure that the mtime on the file is kept
2546 * up-to-date to avoid losing cached files on cold starts because
2547 * their mtime seems old...
2549 struct afs_fheader theader;
2551 theader.magic = AFS_FHMAGIC;
2552 theader.firstCSize = AFS_FIRSTCSIZE;
2553 theader.otherCSize = AFS_OTHERCSIZE;
2554 theader.version = AFS_CI_VERSION;
2555 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2557 ReleaseWriteLock(&afs_xdcache);
2564 * Return a pointer to an freshly initialized dcache entry using
2565 * a memory-based cache. The tlock will be read-locked.
2568 * aslot : Dcache slot to look at.
2569 * tmpdc : Ptr to dcache entry.
2572 * Must be called with afs_xdcache write-locked.
2576 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2578 register struct dcache *tdc;
2581 AFS_STATCNT(afs_MemGetDSlot);
2582 if (CheckLock(&afs_xdcache) != -1)
2583 osi_Panic("getdslot nolock");
2584 if (aslot < 0 || aslot >= afs_cacheFiles)
2585 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2586 tdc = afs_indexTable[aslot];
2588 QRemove(&tdc->lruq); /* move to queue head */
2589 QAdd(&afs_DLRU, &tdc->lruq);
2590 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2591 ObtainWriteLock(&tdc->tlock, 624);
2593 ConvertWToRLock(&tdc->tlock);
2596 if (tmpdc == NULL) {
2597 if (!afs_freeDSList)
2598 afs_GetDownDSlot(4);
2599 if (!afs_freeDSList) {
2600 /* none free, making one is better than a panic */
2601 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2602 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2603 #ifdef KERNEL_HAVE_PIN
2604 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2607 tdc = afs_freeDSList;
2608 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2611 tdc->dflags = 0; /* up-to-date, not in free q */
2613 QAdd(&afs_DLRU, &tdc->lruq);
2614 if (tdc->lruq.prev == &tdc->lruq)
2615 osi_Panic("lruq 3");
2621 /* initialize entry */
2622 tdc->f.fid.Cell = 0;
2623 tdc->f.fid.Fid.Volume = 0;
2625 hones(tdc->f.versionNo);
2626 tdc->f.inode.mem = aslot;
2627 tdc->dflags |= DFEntryMod;
2630 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2633 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2634 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2635 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2638 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2639 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2640 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2641 ObtainReadLock(&tdc->tlock);
2644 afs_indexTable[aslot] = tdc;
2647 } /*afs_MemGetDSlot */
2649 unsigned int last_error = 0, lasterrtime = 0;
2655 * Return a pointer to an freshly initialized dcache entry using
2656 * a UFS-based disk cache. The dcache tlock will be read-locked.
2659 * aslot : Dcache slot to look at.
2660 * tmpdc : Ptr to dcache entry.
2663 * afs_xdcache lock write-locked.
2666 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2668 register afs_int32 code;
2669 register struct dcache *tdc;
2673 AFS_STATCNT(afs_UFSGetDSlot);
2674 if (CheckLock(&afs_xdcache) != -1)
2675 osi_Panic("getdslot nolock");
2676 if (aslot < 0 || aslot >= afs_cacheFiles)
2677 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2678 tdc = afs_indexTable[aslot];
2680 QRemove(&tdc->lruq); /* move to queue head */
2681 QAdd(&afs_DLRU, &tdc->lruq);
2682 /* Grab tlock in case refCount != 0 */
2683 ObtainWriteLock(&tdc->tlock, 625);
2685 ConvertWToRLock(&tdc->tlock);
2688 /* otherwise we should read it in from the cache file */
2690 * If we weren't passed an in-memory region to place the file info,
2691 * we have to allocate one.
2693 if (tmpdc == NULL) {
2694 if (!afs_freeDSList)
2695 afs_GetDownDSlot(4);
2696 if (!afs_freeDSList) {
2697 /* none free, making one is better than a panic */
2698 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2699 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2700 #ifdef KERNEL_HAVE_PIN
2701 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2704 tdc = afs_freeDSList;
2705 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2708 tdc->dflags = 0; /* up-to-date, not in free q */
2710 QAdd(&afs_DLRU, &tdc->lruq);
2711 if (tdc->lruq.prev == &tdc->lruq)
2712 osi_Panic("lruq 3");
2719 * Seek to the aslot'th entry and read it in.
2722 afs_osi_Read(afs_cacheInodep,
2723 sizeof(struct fcache) * aslot +
2724 sizeof(struct afs_fheader), (char *)(&tdc->f),
2725 sizeof(struct fcache));
2727 if (code != sizeof(struct fcache))
2729 if (!afs_CellNumValid(tdc->f.fid.Cell))
2733 tdc->f.fid.Cell = 0;
2734 tdc->f.fid.Fid.Volume = 0;
2736 hones(tdc->f.versionNo);
2737 tdc->dflags |= DFEntryMod;
2738 #if defined(KERNEL_HAVE_UERROR)
2739 last_error = getuerror();
2741 lasterrtime = osi_Time();
2742 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2743 tdc->f.states &= ~(DRO|DBackup|DRW);
2744 afs_DCMoveBucket(tdc, 0, 0);
2747 if (tdc->f.states & DRO) {
2748 afs_DCMoveBucket(tdc, 0, 2);
2749 } else if (tdc->f.states & DBackup) {
2750 afs_DCMoveBucket(tdc, 0, 1);
2752 afs_DCMoveBucket(tdc, 0, 1);
2758 if (tdc->f.chunk >= 0)
2759 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2764 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2765 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2766 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2769 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2770 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2771 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2772 ObtainReadLock(&tdc->tlock);
2775 * If we didn't read into a temporary dcache region, update the
2776 * slot pointer table.
2779 afs_indexTable[aslot] = tdc;
2782 } /*afs_UFSGetDSlot */
2787 * Write a particular dcache entry back to its home in the
2790 * \param adc Pointer to the dcache entry to write.
2791 * \param atime If true, set the modtime on the file to the current time.
2793 * \note Environment:
2794 * Must be called with the afs_xdcache lock at least read-locked,
2795 * and dcache entry at least read-locked.
2796 * The reference count is not changed.
2800 afs_WriteDCache(register struct dcache *adc, int atime)
2802 register afs_int32 code;
2804 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2806 AFS_STATCNT(afs_WriteDCache);
2807 osi_Assert(WriteLocked(&afs_xdcache));
2809 adc->f.modTime = osi_Time();
2811 * Seek to the right dcache slot and write the in-memory image out to disk.
2813 afs_cellname_write();
2815 afs_osi_Write(afs_cacheInodep,
2816 sizeof(struct fcache) * adc->index +
2817 sizeof(struct afs_fheader), (char *)(&adc->f),
2818 sizeof(struct fcache));
2819 if (code != sizeof(struct fcache))
2827 * Wake up users of a particular file waiting for stores to take
2830 * \param avc Ptr to related vcache entry.
2832 * \note Environment:
2833 * Nothing interesting.
2836 afs_wakeup(register struct vcache *avc)
2839 register struct brequest *tb;
2841 AFS_STATCNT(afs_wakeup);
2842 for (i = 0; i < NBRS; i++, tb++) {
2843 /* if request is valid and for this file, we've found it */
2844 if (tb->refCount > 0 && avc == tb->vc) {
2847 * If CSafeStore is on, then we don't awaken the guy
2848 * waiting for the store until the whole store has finished.
2849 * Otherwise, we do it now. Note that if CSafeStore is on,
2850 * the BStore routine actually wakes up the user, instead
2852 * I think this is redundant now because this sort of thing
2853 * is already being handled by the higher-level code.
2855 if ((avc->f.states & CSafeStore) == 0) {
2857 tb->flags |= BUVALID;
2858 if (tb->flags & BUWAIT) {
2859 tb->flags &= ~BUWAIT;
2871 * Given a file name and inode, set up that file to be an
2872 * active member in the AFS cache. This also involves checking
2873 * the usability of its data.
2875 * \param afile Name of the cache file to initialize.
2876 * \param ainode Inode of the file.
2878 * \note Environment:
2879 * This function is called only during initialization.
2882 afs_InitCacheFile(char *afile, ino_t ainode)
2884 register afs_int32 code;
2887 struct osi_file *tfile;
2888 struct osi_stat tstat;
2889 register struct dcache *tdc;
2891 AFS_STATCNT(afs_InitCacheFile);
2892 index = afs_stats_cmperf.cacheNumEntries;
2893 if (index >= afs_cacheFiles)
2896 ObtainWriteLock(&afs_xdcache, 282);
2897 tdc = afs_GetDSlot(index, NULL);
2898 ReleaseReadLock(&tdc->tlock);
2899 ReleaseWriteLock(&afs_xdcache);
2901 ObtainWriteLock(&tdc->lock, 621);
2902 ObtainWriteLock(&afs_xdcache, 622);
2904 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
2906 ReleaseWriteLock(&afs_xdcache);
2907 ReleaseWriteLock(&tdc->lock);
2912 /* Add any other 'complex' inode types here ... */
2913 #if defined(UKERNEL) || !(defined(LINUX_USE_FH) || defined(AFS_CACHE_VNODE_PATH))
2914 tdc->f.inode.ufs = ainode;
2916 osi_Panic("Can't init cache with inode numbers when complex inodes are "
2921 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
2923 tfile = osi_UFSOpen(&tdc->f.inode);
2924 code = afs_osi_Stat(tfile, &tstat);
2926 osi_Panic("initcachefile stat");
2929 * If file size doesn't match the cache info file, it's probably bad.
2931 if (tdc->f.chunkBytes != tstat.size)
2933 tdc->f.chunkBytes = 0;
2936 * If file changed within T (120?) seconds of cache info file, it's
2937 * probably bad. In addition, if slot changed within last T seconds,
2938 * the cache info file may be incorrectly identified, and so slot
2941 if (cacheInfoModTime < tstat.mtime + 120)
2943 if (cacheInfoModTime < tdc->f.modTime + 120)
2945 /* In case write through is behind, make sure cache items entry is
2946 * at least as new as the chunk.
2948 if (tdc->f.modTime < tstat.mtime)
2951 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
2952 if (tstat.size != 0)
2953 osi_UFSTruncate(tfile, 0);
2954 tdc->f.states &= ~(DRO|DBackup|DRW);
2955 afs_DCMoveBucket(tdc, 0, 0);
2956 /* put entry in free cache slot list */
2957 afs_dvnextTbl[tdc->index] = afs_freeDCList;
2958 afs_freeDCList = index;
2960 afs_indexFlags[index] |= IFFree;
2961 afs_indexUnique[index] = 0;
2964 * We must put this entry in the appropriate hash tables.
2965 * Note that i is still set from the above DCHash call
2967 code = DCHash(&tdc->f.fid, tdc->f.chunk);
2968 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
2969 afs_dchashTbl[code] = tdc->index;
2970 code = DVHash(&tdc->f.fid);
2971 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
2972 afs_dvhashTbl[code] = tdc->index;
2973 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
2975 /* has nontrivial amt of data */
2976 afs_indexFlags[index] |= IFEverUsed;
2977 afs_stats_cmperf.cacheFilesReused++;
2979 * Initialize index times to file's mod times; init indexCounter
2982 hset32(afs_indexTimes[index], tstat.atime);
2983 if (hgetlo(afs_indexCounter) < tstat.atime) {
2984 hset32(afs_indexCounter, tstat.atime);
2986 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
2987 } /*File is not bad */
2989 osi_UFSClose(tfile);
2990 tdc->f.states &= ~DWriting;
2991 tdc->dflags &= ~DFEntryMod;
2992 /* don't set f.modTime; we're just cleaning up */
2993 afs_WriteDCache(tdc, 0);
2994 ReleaseWriteLock(&afs_xdcache);
2995 ReleaseWriteLock(&tdc->lock);
2997 afs_stats_cmperf.cacheNumEntries++;
3002 /*Max # of struct dcache's resident at any time*/
3004 * If 'dchint' is enabled then in-memory dcache min is increased because of
3010 * Initialize dcache related variables.
3020 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3022 register struct dcache *tdp;
3026 afs_freeDCList = NULLIDX;
3027 afs_discardDCList = NULLIDX;
3028 afs_freeDCCount = 0;
3029 afs_freeDSList = NULL;
3030 hzero(afs_indexCounter);
3032 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3038 if (achunk < 0 || achunk > 30)
3039 achunk = 13; /* Use default */
3040 AFS_SETCHUNKSIZE(achunk);
3046 if (aflags & AFSCALL_INIT_MEMCACHE) {
3048 * Use a memory cache instead of a disk cache
3050 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3051 afs_cacheType = &afs_MemCacheOps;
3052 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3053 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3054 /* ablocks is reported in 1K blocks */
3055 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3057 afs_warn("afsd: memory cache too large for available memory.\n");
3058 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3060 afiles = ablocks = 0;
3062 afs_warn("Memory cache: Allocating %d dcache entries...",
3065 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3066 afs_cacheType = &afs_UfsCacheOps;
3069 if (aDentries > 512)
3070 afs_dhashsize = 2048;
3071 /* initialize hash tables */
3073 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3075 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3076 for (i = 0; i < afs_dhashsize; i++) {
3077 afs_dvhashTbl[i] = NULLIDX;
3078 afs_dchashTbl[i] = NULLIDX;
3080 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3081 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3082 for (i = 0; i < afiles; i++) {
3083 afs_dvnextTbl[i] = NULLIDX;
3084 afs_dcnextTbl[i] = NULLIDX;
3087 /* Allocate and zero the pointer array to the dcache entries */
3088 afs_indexTable = (struct dcache **)
3089 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3090 memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3092 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3093 memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3095 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3096 memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3097 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3098 memset(afs_indexFlags, 0, afiles * sizeof(char));
3100 /* Allocate and thread the struct dcache entries themselves */
3101 tdp = afs_Initial_freeDSList =
3102 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3103 memset(tdp, 0, aDentries * sizeof(struct dcache));
3104 #ifdef KERNEL_HAVE_PIN
3105 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3106 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3107 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3108 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3109 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3110 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3111 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3112 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3113 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3116 afs_freeDSList = &tdp[0];
3117 for (i = 0; i < aDentries - 1; i++) {
3118 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3119 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3120 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3121 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3123 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3124 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3125 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3126 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3128 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3129 afs_cacheBlocks = ablocks;
3130 afs_ComputeCacheParms(); /* compute parms based on cache size */
3132 afs_dcentries = aDentries;
3134 afs_stats_cmperf.cacheBucket0_Discarded =
3135 afs_stats_cmperf.cacheBucket1_Discarded =
3136 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3142 * Shuts down the cache.
3146 shutdown_dcache(void)
3150 #ifdef AFS_CACHE_VNODE_PATH
3151 if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
3153 for (i = 0; i < afs_cacheFiles; i++) {
3154 tdc = afs_indexTable[i];
3156 afs_osi_FreeStr(tdc->f.inode.ufs);
3162 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3163 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3164 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3165 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3166 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3167 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3168 afs_osi_Free(afs_Initial_freeDSList,
3169 afs_dcentries * sizeof(struct dcache));
3170 #ifdef KERNEL_HAVE_PIN
3171 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3172 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3173 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3174 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3175 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3176 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3177 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3181 for (i = 0; i < afs_dhashsize; i++) {
3182 afs_dvhashTbl[i] = NULLIDX;
3183 afs_dchashTbl[i] = NULLIDX;
3186 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3187 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3189 afs_blocksUsed = afs_dcentries = 0;
3190 afs_stats_cmperf.cacheBucket0_Discarded =
3191 afs_stats_cmperf.cacheBucket1_Discarded =
3192 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3193 hzero(afs_indexCounter);
3195 afs_freeDCCount = 0;
3196 afs_freeDCList = NULLIDX;
3197 afs_discardDCList = NULLIDX;
3198 afs_freeDSList = afs_Initial_freeDSList = 0;
3200 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3206 * Get a dcache ready for writing, respecting the current cache size limits
3208 * len is required because afs_GetDCache with flag == 4 expects the length
3209 * field to be filled. It decides from this whether it's necessary to fetch
3210 * data into the chunk before writing or not (when the whole chunk is
3213 * \param avc The vcache to fetch a dcache for
3214 * \param filePos The start of the section to be written
3215 * \param len The length of the section to be written
3219 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3220 * must be released and afs_PutDCache() called to free dcache.
3223 * \note avc->lock must be held on entry. Function may release and reobtain
3224 * avc->lock and GLOCK.
3228 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3229 afs_size_t len, struct vrequest *areq,
3231 struct dcache *tdc = NULL;
3234 /* read the cached info */
3236 tdc = afs_FindDCache(avc, filePos);
3238 ObtainWriteLock(&tdc->lock, 657);
3239 } else if (afs_blocksUsed >
3240 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3241 tdc = afs_FindDCache(avc, filePos);
3243 ObtainWriteLock(&tdc->lock, 658);
3244 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3245 || (tdc->dflags & DFFetching)) {
3246 ReleaseWriteLock(&tdc->lock);
3252 afs_MaybeWakeupTruncateDaemon();
3253 while (afs_blocksUsed >
3254 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3255 ReleaseWriteLock(&avc->lock);
3256 if (afs_blocksUsed - afs_blocksDiscarded >
3257 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3258 afs_WaitForCacheDrain = 1;
3259 afs_osi_Sleep(&afs_WaitForCacheDrain);
3261 afs_MaybeFreeDiscardedDCache();
3262 afs_MaybeWakeupTruncateDaemon();
3263 ObtainWriteLock(&avc->lock, 509);
3265 avc->f.states |= CDirty;
3266 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3268 ObtainWriteLock(&tdc->lock, 659);
3271 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3273 ObtainWriteLock(&tdc->lock, 660);
3276 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3277 afs_stats_cmperf.cacheCurrDirtyChunks++;
3278 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3280 if (!(tdc->f.states & DWriting)) {
3281 /* don't mark entry as mod if we don't have to */
3282 tdc->f.states |= DWriting;
3283 tdc->dflags |= DFEntryMod;
3289 #if defined(AFS_DISCON_ENV)
3292 * Make a shadow copy of a dir's dcache. It's used for disconnected
3293 * operations like remove/create/rename to keep the original directory data.
3294 * On reconnection, we can diff the original data with the server and get the
3295 * server changes and with the local data to get the local changes.
3297 * \param avc The dir vnode.
3298 * \param adc The dir dcache.
3300 * \return 0 for success.
3302 * \note The vcache entry must be write locked.
3303 * \note The dcache entry must be read locked.
3306 afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3308 int i, code, ret_code = 0, written, trans_size;
3309 struct dcache *new_dc = NULL;
3310 struct osi_file *tfile_src, *tfile_dst;
3311 struct VenusFid shadow_fid;
3314 /* Is this a dir? */
3315 if (vType(avc) != VDIR)
3318 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3321 /* Generate a fid for the shadow dir. */
3322 shadow_fid.Cell = avc->f.fid.Cell;
3323 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3324 afs_GenShadowFid(&shadow_fid);
3326 ObtainWriteLock(&afs_xdcache, 716);
3328 /* Get a fresh dcache. */
3329 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3331 ObtainReadLock(&adc->mflock);
3333 /* Set up the new fid. */
3334 /* Copy interesting data from original dir dcache. */
3335 new_dc->mflags = adc->mflags;
3336 new_dc->dflags = adc->dflags;
3337 new_dc->f.modTime = adc->f.modTime;
3338 new_dc->f.versionNo = adc->f.versionNo;
3339 new_dc->f.states = adc->f.states;
3340 new_dc->f.chunk= adc->f.chunk;
3341 new_dc->f.chunkBytes = adc->f.chunkBytes;
3343 ReleaseReadLock(&adc->mflock);
3345 /* Now add to the two hash chains */
3346 i = DCHash(&shadow_fid, 0);
3347 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3348 afs_dchashTbl[i] = new_dc->index;
3350 i = DVHash(&shadow_fid);
3351 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3352 afs_dvhashTbl[i] = new_dc->index;
3354 ReleaseWriteLock(&afs_xdcache);
3356 /* Alloc a 4k block. */
3357 data = (char *) afs_osi_Alloc(4096);
3359 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3364 /* Open the files. */
3365 tfile_src = afs_CFileOpen(&adc->f.inode);
3366 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3368 /* And now copy dir dcache data into this dcache,
3372 while (written < adc->f.chunkBytes) {
3373 trans_size = adc->f.chunkBytes - written;
3374 if (trans_size > 4096)
3377 /* Read a chunk from the dcache. */
3378 code = afs_CFileRead(tfile_src, written, data, trans_size);
3379 if (code < trans_size) {
3384 /* Write it to the new dcache. */
3385 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3386 if (code < trans_size) {
3391 written+=trans_size;
3394 afs_CFileClose(tfile_dst);
3395 afs_CFileClose(tfile_src);
3397 afs_osi_Free(data, 4096);
3399 ReleaseWriteLock(&new_dc->lock);
3400 afs_PutDCache(new_dc);
3403 ObtainWriteLock(&afs_xvcache, 763);
3404 ObtainWriteLock(&afs_disconDirtyLock, 765);
3405 QAdd(&afs_disconShadow, &avc->shadowq);
3407 ReleaseWriteLock(&afs_disconDirtyLock);
3408 ReleaseWriteLock(&afs_xvcache);
3410 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3411 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3419 * Delete the dcaches of a shadow dir.
3421 * \param avc The vcache containing the shadow fid.
3423 * \note avc must be write locked.
3426 afs_DeleteShadowDir(struct vcache *avc)
3429 struct VenusFid shadow_fid;
3431 shadow_fid.Cell = avc->f.fid.Cell;
3432 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3433 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3434 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3436 tdc = afs_FindDCacheByFid(&shadow_fid);
3438 afs_HashOutDCache(tdc, 1);
3439 afs_DiscardDCache(tdc);
3442 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3443 ObtainWriteLock(&afs_disconDirtyLock, 708);
3444 QRemove(&avc->shadowq);
3445 ReleaseWriteLock(&afs_disconDirtyLock);
3446 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3450 * Populate a dcache with empty chunks up to a given file size,
3451 * used before extending a file in order to avoid 'holes' which
3452 * we can't access in disconnected mode.
3454 * \param avc The vcache which is being extended (locked)
3455 * \param alen The new length of the file
3459 afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
3462 afs_size_t len, offset;
3463 afs_int32 start, end;
3465 /* We're doing this to deal with the situation where we extend
3466 * by writing after lseek()ing past the end of the file . If that
3467 * extension skips chunks, then those chunks won't be created, and
3468 * GetDCache will assume that they have to be fetched from the server.
3469 * So, for each chunk between the current file position, and the new
3470 * length we GetDCache for that chunk.
3473 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3476 if (avc->f.m.Length == 0)
3479 start = AFS_CHUNK(avc->f.m.Length)+1;
3481 end = AFS_CHUNK(apos);
3484 len = AFS_CHUNKTOSIZE(start);
3485 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);