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"
19 #include "afs/sysincludes.h" /*Standard vendor system headers */
20 #include "afsincludes.h" /*AFS-based standard headers */
21 #include "afs/afs_stats.h" /* statistics */
22 #include "afs/afs_cbqueue.h"
23 #include "afs/afs_osidnlc.h"
25 /* Forward declarations. */
26 static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
27 static void afs_FreeDiscardedDCache(void);
28 static void afs_DiscardDCache(struct dcache *);
29 static void afs_FreeDCache(struct dcache *);
31 static afs_int32 afs_DCGetBucket(struct vcache *);
32 static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
33 static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
34 static void afs_DCSizeInit(void);
35 static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);
39 * --------------------- Exported definitions ---------------------
42 afs_int32 afs_blocksUsed_0; /*1K blocks in cache - in theory is zero */
43 afs_int32 afs_blocksUsed_1; /*1K blocks in cache */
44 afs_int32 afs_blocksUsed_2; /*1K blocks in cache */
45 afs_int32 afs_pct1 = -1;
46 afs_int32 afs_pct2 = -1;
47 afs_uint32 afs_tpct1 = 0;
48 afs_uint32 afs_tpct2 = 0;
49 afs_uint32 splitdcache = 0;
51 afs_lock_t afs_xdcache; /*Lock: alloc new disk cache entries */
52 afs_int32 afs_freeDCList; /*Free list for disk cache entries */
53 afs_int32 afs_freeDCCount; /*Count of elts in freeDCList */
54 afs_int32 afs_discardDCList; /*Discarded disk cache entries */
55 afs_int32 afs_discardDCCount; /*Count of elts in discardDCList */
56 struct dcache *afs_freeDSList; /*Free list for disk slots */
57 struct dcache *afs_Initial_freeDSList; /*Initial list for above */
58 ino_t cacheInode; /*Inode for CacheItems file */
59 struct osi_file *afs_cacheInodep = 0; /* file for CacheItems inode */
60 struct afs_q afs_DLRU; /*dcache LRU */
61 afs_int32 afs_dhashsize = 1024;
62 afs_int32 *afs_dvhashTbl; /*Data cache hash table */
63 afs_int32 *afs_dchashTbl; /*Data cache hash table */
64 afs_int32 *afs_dvnextTbl; /*Dcache hash table links */
65 afs_int32 *afs_dcnextTbl; /*Dcache hash table links */
66 struct dcache **afs_indexTable; /*Pointers to dcache entries */
67 afs_hyper_t *afs_indexTimes; /*Dcache entry Access times */
68 afs_int32 *afs_indexUnique; /*dcache entry Fid.Unique */
69 unsigned char *afs_indexFlags; /*(only one) Is there data there? */
70 afs_hyper_t afs_indexCounter; /*Fake time for marking index
72 afs_int32 afs_cacheFiles = 0; /*Size of afs_indexTable */
73 afs_int32 afs_cacheBlocks; /*1K blocks in cache */
74 afs_int32 afs_cacheStats; /*Stat entries in cache */
75 afs_int32 afs_blocksUsed; /*Number of blocks in use */
76 afs_int32 afs_blocksDiscarded; /*Blocks freed but not truncated */
77 afs_int32 afs_fsfragsize = 1023; /*Underlying Filesystem minimum unit
78 *of disk allocation usually 1K
79 *this value is (truefrag -1 ) to
80 *save a bunch of subtracts... */
81 #ifdef AFS_64BIT_CLIENT
82 #ifdef AFS_VM_RDWR_ENV
83 afs_size_t afs_vmMappingEnd; /* for large files (>= 2GB) the VM
84 * mapping an 32bit addressing machines
85 * can only be used below the 2 GB
86 * line. From this point upwards we
87 * must do direct I/O into the cache
88 * files. The value should be on a
90 #endif /* AFS_VM_RDWR_ENV */
91 #endif /* AFS_64BIT_CLIENT */
93 /* The following is used to ensure that new dcache's aren't obtained when
94 * the cache is nearly full.
96 int afs_WaitForCacheDrain = 0;
97 int afs_TruncateDaemonRunning = 0;
98 int afs_CacheTooFull = 0;
100 afs_int32 afs_dcentries; /* In-memory dcache entries */
103 int dcacheDisabled = 0;
105 static int afs_UFSCacheFetchProc(), afs_UFSCacheStoreProc();
106 struct afs_cacheOps afs_UfsCacheOps = {
114 afs_UFSCacheFetchProc,
115 afs_UFSCacheStoreProc,
121 struct afs_cacheOps afs_MemCacheOps = {
123 afs_MemCacheTruncate,
129 afs_MemCacheFetchProc,
130 afs_MemCacheStoreProc,
136 int cacheDiskType; /*Type of backing disk for cache */
137 struct afs_cacheOps *afs_cacheType;
140 afs_DCGetBucket(struct vcache *avc)
142 /* This should be replaced with some sort of user configurable function */
143 if (avc->states & CRO) {
145 } else if (avc->states & CBackup) {
155 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
157 afs_int32 adjustSize = newSize - oldSize;
165 afs_blocksUsed_0 += adjustSize;
166 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
169 afs_blocksUsed_1 += adjustSize;
170 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
173 afs_blocksUsed_2 += adjustSize;
174 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
182 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
190 afs_blocksUsed_0 -= size;
193 afs_blocksUsed_1 -= size;
196 afs_blocksUsed_2 -= size;
200 adc->bucket = newBucket;
205 afs_blocksUsed_0 += size;
208 afs_blocksUsed_1 += size;
211 afs_blocksUsed_2 += size;
221 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
225 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
230 afs_pct1 = afs_blocksUsed_1*100/afs_cacheBlocks;
231 afs_pct2 = afs_blocksUsed_2*100/afs_cacheBlocks;
233 /* Short cut: if we don't know about it, try to kill it */
234 if (phase < 2 && afs_blocksUsed_0)
237 if (afs_pct1 > afs_tpct1)
239 if (afs_pct2 > afs_tpct2)
241 return 0; /* unlikely */
249 * Warn about failing to store a file.
252 * acode : Associated error code.
253 * avolume : Volume involved.
254 * aflags : How to handle the output:
255 * aflags & 1: Print out on console
256 * aflags & 2: Print out on controlling tty
259 * Call this from close call when vnodeops is RCS unlocked.
263 afs_StoreWarn(register afs_int32 acode, afs_int32 avolume,
264 register afs_int32 aflags)
266 static char problem_fmt[] =
267 "afs: failed to store file in volume %d (%s)\n";
268 static char problem_fmt_w_error[] =
269 "afs: failed to store file in volume %d (error %d)\n";
270 static char netproblems[] = "network problems";
271 static char partfull[] = "partition full";
272 static char overquota[] = "over quota";
274 AFS_STATCNT(afs_StoreWarn);
280 afs_warn(problem_fmt, avolume, netproblems);
282 afs_warnuser(problem_fmt, avolume, netproblems);
283 } else if (acode == ENOSPC) {
288 afs_warn(problem_fmt, avolume, partfull);
290 afs_warnuser(problem_fmt, avolume, partfull);
293 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
294 * Instead ENOSPC will be sent...
296 if (acode == EDQUOT) {
301 afs_warn(problem_fmt, avolume, overquota);
303 afs_warnuser(problem_fmt, avolume, overquota);
311 afs_warn(problem_fmt_w_error, avolume, acode);
313 afs_warnuser(problem_fmt_w_error, avolume, acode);
318 afs_MaybeWakeupTruncateDaemon(void)
320 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
321 afs_CacheTooFull = 1;
322 if (!afs_TruncateDaemonRunning)
323 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
324 } else if (!afs_TruncateDaemonRunning
325 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
326 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
330 /* Keep statistics on run time for afs_CacheTruncateDaemon. This is a
331 * struct so we need only export one symbol for AIX.
333 static struct CTD_stats {
334 osi_timeval_t CTD_beforeSleep;
335 osi_timeval_t CTD_afterSleep;
336 osi_timeval_t CTD_sleepTime;
337 osi_timeval_t CTD_runTime;
341 u_int afs_min_cache = 0;
343 afs_CacheTruncateDaemon(void)
345 osi_timeval_t CTD_tmpTime;
349 (100 - CM_DCACHECOUNTFREEPCT +
350 CM_DCACHEEXTRAPCT) * afs_cacheFiles / 100;
352 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
354 osi_GetuTime(&CTD_stats.CTD_afterSleep);
355 afs_TruncateDaemonRunning = 1;
357 cb_lowat = ((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT)
358 * afs_cacheBlocks) / 100;
359 MObtainWriteLock(&afs_xdcache, 266);
360 if (afs_CacheTooFull) {
361 int space_needed, slots_needed;
362 /* if we get woken up, we should try to clean something out */
363 for (counter = 0; counter < 10; counter++) {
365 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
367 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
368 afs_GetDownD(slots_needed, &space_needed, 0);
369 if ((space_needed <= 0) && (slots_needed <= 0)) {
372 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
375 if (!afs_CacheIsTooFull())
376 afs_CacheTooFull = 0;
378 MReleaseWriteLock(&afs_xdcache);
381 * This is a defensive check to try to avoid starving threads
382 * that may need the global lock so thay can help free some
383 * cache space. If this thread won't be sleeping or truncating
384 * any cache files then give up the global lock so other
385 * threads get a chance to run.
387 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
388 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
389 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
393 * This is where we free the discarded cache elements.
395 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
396 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
397 afs_FreeDiscardedDCache();
400 /* See if we need to continue to run. Someone may have
401 * signalled us while we were executing.
403 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
404 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
405 /* Collect statistics on truncate daemon. */
406 CTD_stats.CTD_nSleeps++;
407 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
408 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
409 CTD_stats.CTD_beforeSleep);
410 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
412 afs_TruncateDaemonRunning = 0;
413 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
414 afs_TruncateDaemonRunning = 1;
416 osi_GetuTime(&CTD_stats.CTD_afterSleep);
417 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
418 CTD_stats.CTD_afterSleep);
419 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
421 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
423 afs_termState = AFSOP_STOP_AFSDB;
425 afs_termState = AFSOP_STOP_RXEVENT;
427 afs_osi_Wakeup(&afs_termState);
438 * Make adjustment for the new size in the disk cache entry
440 * Major Assumptions Here:
441 * Assumes that frag size is an integral power of two, less one,
442 * and that this is a two's complement machine. I don't
443 * know of any filesystems which violate this assumption...
446 * adc : Ptr to dcache entry.
447 * anewsize : New size desired.
451 afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
453 register afs_int32 oldSize;
455 AFS_STATCNT(afs_AdjustSize);
457 adc->dflags |= DFEntryMod;
458 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
459 adc->f.chunkBytes = newSize;
462 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
463 afs_DCAdjustSize(adc, oldSize, newSize);
464 if (newSize > oldSize) {
465 /* We're growing the file, wakeup the daemon */
466 afs_MaybeWakeupTruncateDaemon();
468 afs_blocksUsed += (newSize - oldSize);
469 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
477 * This routine is responsible for moving at least one entry (but up
478 * to some number of them) from the LRU queue to the free queue.
481 * anumber : Number of entries that should ideally be moved.
482 * aneedSpace : How much space we need (1K blocks);
485 * The anumber parameter is just a hint; at least one entry MUST be
486 * moved, or we'll panic. We must be called with afs_xdcache
487 * write-locked. We should try to satisfy both anumber and aneedspace,
488 * whichever is more demanding - need to do several things:
489 * 1. only grab up to anumber victims if aneedSpace <= 0, not
490 * the whole set of MAXATONCE.
491 * 2. dynamically choose MAXATONCE to reflect severity of
492 * demand: something like (*aneedSpace >> (logChunk - 9))
493 * N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
494 * indicates that the cache is not properly configured/tuned or
495 * something. We should be able to automatically correct that problem.
498 #define MAXATONCE 16 /* max we can obtain at once */
500 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
504 struct VenusFid *afid;
508 register struct vcache *tvc;
509 afs_uint32 victims[MAXATONCE];
510 struct dcache *victimDCs[MAXATONCE];
511 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
512 afs_uint32 victimPtr; /* next free item in victim arrays */
513 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
514 afs_uint32 maxVictimPtr; /* where it is */
518 AFS_STATCNT(afs_GetDownD);
519 if (CheckLock(&afs_xdcache) != -1)
520 osi_Panic("getdownd nolock");
521 /* decrement anumber first for all dudes in free list */
522 /* SHOULD always decrement anumber first, even if aneedSpace >0,
523 * because we should try to free space even if anumber <=0 */
524 if (!aneedSpace || *aneedSpace <= 0) {
525 anumber -= afs_freeDCCount;
527 return; /* enough already free */
529 /* bounds check parameter */
530 if (anumber > MAXATONCE)
531 anumber = MAXATONCE; /* all we can do */
533 /* rewrite so phases include a better eligiblity for gc test*/
535 * The phase variable manages reclaims. Set to 0, the first pass,
536 * we don't reclaim active entries, or other than target bucket.
537 * Set to 1, we reclaim even active ones in target bucket.
538 * Set to 2, we reclaim any inactive one.
539 * Set to 3, we reclaim even active ones.
547 for (i = 0; i < afs_cacheFiles; i++)
548 /* turn off all flags */
549 afs_indexFlags[i] &= ~IFFlag;
551 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
552 /* find oldest entries for reclamation */
553 maxVictimPtr = victimPtr = 0;
554 hzero(maxVictimTime);
555 curbucket = afs_DCWhichBucket(phase, buckethint);
556 /* select victims from access time array */
557 for (i = 0; i < afs_cacheFiles; i++) {
558 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
559 /* skip if dirty or already free */
562 tdc = afs_indexTable[i];
563 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
565 /* Wrong bucket; can't use it! */
568 if (tdc && (tdc->refCount != 0)) {
569 /* Referenced; can't use it! */
572 hset(vtime, afs_indexTimes[i]);
574 /* if we've already looked at this one, skip it */
575 if (afs_indexFlags[i] & IFFlag)
578 if (victimPtr < MAXATONCE) {
579 /* if there's at least one free victim slot left */
580 victims[victimPtr] = i;
581 hset(victimTimes[victimPtr], vtime);
582 if (hcmp(vtime, maxVictimTime) > 0) {
583 hset(maxVictimTime, vtime);
584 maxVictimPtr = victimPtr;
587 } else if (hcmp(vtime, maxVictimTime) < 0) {
589 * We're older than youngest victim, so we replace at
592 /* find youngest (largest LRU) victim */
595 osi_Panic("getdownd local");
597 hset(victimTimes[j], vtime);
598 /* recompute maxVictimTime */
599 hset(maxVictimTime, vtime);
600 for (j = 0; j < victimPtr; j++)
601 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
602 hset(maxVictimTime, victimTimes[j]);
608 /* now really reclaim the victims */
609 j = 0; /* flag to track if we actually got any of the victims */
610 /* first, hold all the victims, since we're going to release the lock
611 * during the truncate operation.
613 for (i = 0; i < victimPtr; i++) {
614 tdc = afs_GetDSlot(victims[i], 0);
615 /* We got tdc->tlock(R) here */
616 if (tdc->refCount == 1)
620 ReleaseReadLock(&tdc->tlock);
624 for (i = 0; i < victimPtr; i++) {
625 /* q is first elt in dcache entry */
627 /* now, since we're dropping the afs_xdcache lock below, we
628 * have to verify, before proceeding, that there are no other
629 * references to this dcache entry, even now. Note that we
630 * compare with 1, since we bumped it above when we called
631 * afs_GetDSlot to preserve the entry's identity.
633 if (tdc && tdc->refCount == 1) {
634 unsigned char chunkFlags;
635 afs_size_t tchunkoffset = 0;
637 /* xdcache is lower than the xvcache lock */
638 MReleaseWriteLock(&afs_xdcache);
639 MObtainReadLock(&afs_xvcache);
640 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
641 MReleaseReadLock(&afs_xvcache);
642 MObtainWriteLock(&afs_xdcache, 527);
644 if (tdc->refCount > 1)
647 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
648 chunkFlags = afs_indexFlags[tdc->index];
649 if ((((phase / 2) & 1) == 0) && osi_Active(tvc))
651 if ((((phase / 2) & 1) == 1) && osi_Active(tvc)
652 && (tvc->states & CDCLock)
653 && (chunkFlags & IFAnyPages))
655 if (chunkFlags & IFDataMod)
657 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
658 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
659 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
660 ICL_HANDLE_OFFSET(tchunkoffset));
662 #if defined(AFS_SUN5_ENV)
664 * Now we try to invalidate pages. We do this only for
665 * Solaris. For other platforms, it's OK to recycle a
666 * dcache entry out from under a page, because the strategy
667 * function can call afs_GetDCache().
669 if (!skip && (chunkFlags & IFAnyPages)) {
672 MReleaseWriteLock(&afs_xdcache);
673 MObtainWriteLock(&tvc->vlock, 543);
674 if (tvc->multiPage) {
678 /* block locking pages */
679 tvc->vstates |= VPageCleaning;
680 /* block getting new pages */
682 MReleaseWriteLock(&tvc->vlock);
683 /* One last recheck */
684 MObtainWriteLock(&afs_xdcache, 333);
685 chunkFlags = afs_indexFlags[tdc->index];
686 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
687 || (osi_Active(tvc) && (tvc->states & CDCLock)
688 && (chunkFlags & IFAnyPages))) {
690 MReleaseWriteLock(&afs_xdcache);
693 MReleaseWriteLock(&afs_xdcache);
695 code = osi_VM_GetDownD(tvc, tdc);
697 MObtainWriteLock(&afs_xdcache, 269);
698 /* we actually removed all pages, clean and dirty */
700 afs_indexFlags[tdc->index] &=
701 ~(IFDirtyPages | IFAnyPages);
704 MReleaseWriteLock(&afs_xdcache);
706 MObtainWriteLock(&tvc->vlock, 544);
707 if (--tvc->activeV == 0
708 && (tvc->vstates & VRevokeWait)) {
709 tvc->vstates &= ~VRevokeWait;
710 afs_osi_Wakeup((char *)&tvc->vstates);
713 if (tvc->vstates & VPageCleaning) {
714 tvc->vstates &= ~VPageCleaning;
715 afs_osi_Wakeup((char *)&tvc->vstates);
718 MReleaseWriteLock(&tvc->vlock);
720 #endif /* AFS_SUN5_ENV */
722 MReleaseWriteLock(&afs_xdcache);
726 MObtainWriteLock(&afs_xdcache, 528);
727 if (afs_indexFlags[tdc->index] &
728 (IFDataMod | IFDirtyPages | IFAnyPages))
730 if (tdc->refCount > 1)
733 #if defined(AFS_SUN5_ENV)
735 /* no vnode, so IFDirtyPages is spurious (we don't
736 * sweep dcaches on vnode recycling, so we can have
737 * DIRTYPAGES set even when all pages are gone). Just
739 * Hold vcache lock to prevent vnode from being
740 * created while we're clearing IFDirtyPages.
742 afs_indexFlags[tdc->index] &=
743 ~(IFDirtyPages | IFAnyPages);
747 /* skip this guy and mark him as recently used */
748 afs_indexFlags[tdc->index] |= IFFlag;
749 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
750 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
751 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
752 ICL_HANDLE_OFFSET(tchunkoffset));
754 /* flush this dude from the data cache and reclaim;
755 * first, make sure no one will care that we damage
756 * it, by removing it from all hash tables. Then,
757 * melt it down for parts. Note that any concurrent
758 * (new possibility!) calls to GetDownD won't touch
759 * this guy because his reference count is > 0. */
760 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
761 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
762 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
763 ICL_HANDLE_OFFSET(tchunkoffset));
764 AFS_STATCNT(afs_gget);
765 afs_HashOutDCache(tdc);
766 if (tdc->f.chunkBytes != 0) {
770 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
775 afs_DiscardDCache(tdc);
780 j = 1; /* we reclaimed at least one victim */
787 /* Phase is 0 and no one was found, so try phase 1 (ignore
788 * osi_Active flag) */
791 for (i = 0; i < afs_cacheFiles; i++)
792 /* turn off all flags */
793 afs_indexFlags[i] &= ~IFFlag;
796 /* found no one in phases 0-5, we're hosed */
800 } /* big while loop */
807 * Description: remove adc from any hash tables that would allow it to be located
808 * again by afs_FindDCache or afs_GetDCache.
810 * Parameters: adc -- pointer to dcache entry to remove from hash tables.
812 * Locks: Must have the afs_xdcache lock write-locked to call this function.
815 afs_HashOutDCache(struct dcache *adc)
819 AFS_STATCNT(afs_glink);
820 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
822 /* if this guy is in the hash table, pull him out */
823 if (adc->f.fid.Fid.Volume != 0) {
824 /* remove entry from first hash chains */
825 i = DCHash(&adc->f.fid, adc->f.chunk);
826 us = afs_dchashTbl[i];
827 if (us == adc->index) {
828 /* first dude in the list */
829 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
831 /* somewhere on the chain */
832 while (us != NULLIDX) {
833 if (afs_dcnextTbl[us] == adc->index) {
834 /* found item pointing at the one to delete */
835 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
838 us = afs_dcnextTbl[us];
841 osi_Panic("dcache hc");
843 /* remove entry from *other* hash chain */
844 i = DVHash(&adc->f.fid);
845 us = afs_dvhashTbl[i];
846 if (us == adc->index) {
847 /* first dude in the list */
848 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
850 /* somewhere on the chain */
851 while (us != NULLIDX) {
852 if (afs_dvnextTbl[us] == adc->index) {
853 /* found item pointing at the one to delete */
854 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
857 us = afs_dvnextTbl[us];
860 osi_Panic("dcache hv");
864 /* prevent entry from being found on a reboot (it is already out of
865 * the hash table, but after a crash, we just look at fid fields of
866 * stable (old) entries).
868 adc->f.fid.Fid.Volume = 0; /* invalid */
870 /* mark entry as modified */
871 adc->dflags |= DFEntryMod;
875 } /*afs_HashOutDCache */
882 * Flush the given dcache entry, pulling it from hash chains
883 * and truncating the associated cache file.
886 * adc: Ptr to dcache entry to flush.
889 * This routine must be called with the afs_xdcache lock held
894 afs_FlushDCache(register struct dcache *adc)
896 AFS_STATCNT(afs_FlushDCache);
898 * Bump the number of cache files flushed.
900 afs_stats_cmperf.cacheFlushes++;
902 /* remove from all hash tables */
903 afs_HashOutDCache(adc);
905 /* Free its space; special case null operation, since truncate operation
906 * in UFS is slow even in this case, and this allows us to pre-truncate
907 * these files at more convenient times with fewer locks set
908 * (see afs_GetDownD).
910 if (adc->f.chunkBytes != 0) {
911 afs_DiscardDCache(adc);
912 afs_MaybeWakeupTruncateDaemon();
917 if (afs_WaitForCacheDrain) {
918 if (afs_blocksUsed <=
919 (CM_CACHESIZEDRAINEDPCT * afs_cacheBlocks) / 100) {
920 afs_WaitForCacheDrain = 0;
921 afs_osi_Wakeup(&afs_WaitForCacheDrain);
924 } /*afs_FlushDCache */
930 * Description: put a dcache entry on the free dcache entry list.
932 * Parameters: adc -- dcache entry to free
934 * Environment: called with afs_xdcache lock write-locked.
937 afs_FreeDCache(register struct dcache *adc)
939 /* Thread on free list, update free list count and mark entry as
940 * freed in its indexFlags element. Also, ensure DCache entry gets
941 * written out (set DFEntryMod).
944 afs_dvnextTbl[adc->index] = afs_freeDCList;
945 afs_freeDCList = adc->index;
947 afs_indexFlags[adc->index] |= IFFree;
948 adc->dflags |= DFEntryMod;
950 if (afs_WaitForCacheDrain) {
951 if ((afs_blocksUsed - afs_blocksDiscarded) <=
952 (CM_CACHESIZEDRAINEDPCT * afs_cacheBlocks) / 100) {
953 afs_WaitForCacheDrain = 0;
954 afs_osi_Wakeup(&afs_WaitForCacheDrain);
963 * Discard the cache element by moving it to the discardDCList.
964 * This puts the cache element into a quasi-freed state, where
965 * the space may be reused, but the file has not been truncated.
967 * Major Assumptions Here:
968 * Assumes that frag size is an integral power of two, less one,
969 * and that this is a two's complement machine. I don't
970 * know of any filesystems which violate this assumption...
973 * adc : Ptr to dcache entry.
976 * Must be called with afs_xdcache write-locked.
980 afs_DiscardDCache(register struct dcache *adc)
982 register afs_int32 size;
984 AFS_STATCNT(afs_DiscardDCache);
986 osi_Assert(adc->refCount == 1);
988 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
989 afs_blocksDiscarded += size;
990 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
992 afs_dvnextTbl[adc->index] = afs_discardDCList;
993 afs_discardDCList = adc->index;
994 afs_discardDCCount++;
996 adc->f.fid.Fid.Volume = 0;
997 adc->dflags |= DFEntryMod;
998 afs_indexFlags[adc->index] |= IFDiscarded;
1000 if (afs_WaitForCacheDrain) {
1001 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1002 (CM_CACHESIZEDRAINEDPCT * afs_cacheBlocks) / 100) {
1003 afs_WaitForCacheDrain = 0;
1004 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1008 } /*afs_DiscardDCache */
1011 * afs_FreeDiscardedDCache
1014 * Free the next element on the list of discarded cache elements.
1017 afs_FreeDiscardedDCache(void)
1019 register struct dcache *tdc;
1020 register struct osi_file *tfile;
1021 register afs_int32 size;
1023 AFS_STATCNT(afs_FreeDiscardedDCache);
1025 MObtainWriteLock(&afs_xdcache, 510);
1026 if (!afs_blocksDiscarded) {
1027 MReleaseWriteLock(&afs_xdcache);
1032 * Get an entry from the list of discarded cache elements
1034 tdc = afs_GetDSlot(afs_discardDCList, 0);
1035 osi_Assert(tdc->refCount == 1);
1036 ReleaseReadLock(&tdc->tlock);
1038 afs_discardDCList = afs_dvnextTbl[tdc->index];
1039 afs_dvnextTbl[tdc->index] = NULLIDX;
1040 afs_discardDCCount--;
1041 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1042 afs_blocksDiscarded -= size;
1043 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1044 /* We can lock because we just took it off the free list */
1045 ObtainWriteLock(&tdc->lock, 626);
1046 MReleaseWriteLock(&afs_xdcache);
1049 * Truncate the element to reclaim its space
1051 tfile = afs_CFileOpen(tdc->f.inode);
1052 afs_CFileTruncate(tfile, 0);
1053 afs_CFileClose(tfile);
1054 afs_AdjustSize(tdc, 0);
1055 tdc->f.states &= ~(DRO|DBackup|DRW);
1056 afs_DCMoveBucket(tdc, 0, 0);
1059 * Free the element we just truncated
1061 MObtainWriteLock(&afs_xdcache, 511);
1062 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1063 afs_FreeDCache(tdc);
1064 ReleaseWriteLock(&tdc->lock);
1066 MReleaseWriteLock(&afs_xdcache);
1070 * afs_MaybeFreeDiscardedDCache
1073 * Free as many entries from the list of discarded cache elements
1074 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1080 afs_MaybeFreeDiscardedDCache(void)
1083 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1085 while (afs_blocksDiscarded
1086 && (afs_blocksUsed >
1087 (CM_WAITFORDRAINPCT * afs_cacheBlocks) / 100)) {
1088 afs_FreeDiscardedDCache();
1097 * Try to free up a certain number of disk slots.
1100 * anumber : Targeted number of disk slots to free up.
1103 * Must be called with afs_xdcache write-locked.
1106 afs_GetDownDSlot(int anumber)
1108 struct afs_q *tq, *nq;
1113 AFS_STATCNT(afs_GetDownDSlot);
1114 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1115 osi_Panic("diskless getdowndslot");
1117 if (CheckLock(&afs_xdcache) != -1)
1118 osi_Panic("getdowndslot nolock");
1120 /* decrement anumber first for all dudes in free list */
1121 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1124 return; /* enough already free */
1126 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1128 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1129 nq = QPrev(tq); /* in case we remove it */
1130 if (tdc->refCount == 0) {
1131 if ((ix = tdc->index) == NULLIDX)
1132 osi_Panic("getdowndslot");
1133 /* pull the entry out of the lruq and put it on the free list */
1134 QRemove(&tdc->lruq);
1136 /* write-through if modified */
1137 if (tdc->dflags & DFEntryMod) {
1138 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1140 * ask proxy to do this for us - we don't have the stack space
1142 while (tdc->dflags & DFEntryMod) {
1145 s = SPLOCK(afs_sgibklock);
1146 if (afs_sgibklist == NULL) {
1147 /* if slot is free, grab it. */
1148 afs_sgibklist = tdc;
1149 SV_SIGNAL(&afs_sgibksync);
1151 /* wait for daemon to (start, then) finish. */
1152 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1156 tdc->dflags &= ~DFEntryMod;
1157 afs_WriteDCache(tdc, 1);
1161 /* finally put the entry in the free list */
1162 afs_indexTable[ix] = NULL;
1163 afs_indexFlags[ix] &= ~IFEverUsed;
1164 tdc->index = NULLIDX;
1165 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1166 afs_freeDSList = tdc;
1170 } /*afs_GetDownDSlot */
1177 * Increment the reference count on a disk cache entry,
1178 * which already has a non-zero refcount. In order to
1179 * increment the refcount of a zero-reference entry, you
1180 * have to hold afs_xdcache.
1183 * adc : Pointer to the dcache entry to increment.
1186 * Nothing interesting.
1189 afs_RefDCache(struct dcache *adc)
1191 ObtainWriteLock(&adc->tlock, 627);
1192 if (adc->refCount < 0)
1193 osi_Panic("RefDCache: negative refcount");
1195 ReleaseWriteLock(&adc->tlock);
1204 * Decrement the reference count on a disk cache entry.
1207 * ad : Ptr to the dcache entry to decrement.
1210 * Nothing interesting.
1213 afs_PutDCache(register struct dcache *adc)
1215 AFS_STATCNT(afs_PutDCache);
1216 ObtainWriteLock(&adc->tlock, 276);
1217 if (adc->refCount <= 0)
1218 osi_Panic("putdcache");
1220 ReleaseWriteLock(&adc->tlock);
1229 * Try to discard all data associated with this file from the
1233 * avc : Pointer to the cache info for the file.
1236 * Both pvnLock and lock are write held.
1239 afs_TryToSmush(register struct vcache *avc, struct AFS_UCRED *acred, int sync)
1241 register struct dcache *tdc;
1244 AFS_STATCNT(afs_TryToSmush);
1245 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1246 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length));
1247 sync = 1; /* XX Temp testing XX */
1249 #if defined(AFS_SUN5_ENV)
1250 ObtainWriteLock(&avc->vlock, 573);
1251 avc->activeV++; /* block new getpages */
1252 ReleaseWriteLock(&avc->vlock);
1255 /* Flush VM pages */
1256 osi_VM_TryToSmush(avc, acred, sync);
1259 * Get the hash chain containing all dce's for this fid
1261 i = DVHash(&avc->fid);
1262 MObtainWriteLock(&afs_xdcache, 277);
1263 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1264 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1265 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1266 int releaseTlock = 1;
1267 tdc = afs_GetDSlot(index, NULL);
1268 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1270 if ((afs_indexFlags[index] & IFDataMod) == 0
1271 && tdc->refCount == 1) {
1272 ReleaseReadLock(&tdc->tlock);
1274 afs_FlushDCache(tdc);
1277 afs_indexTable[index] = 0;
1280 ReleaseReadLock(&tdc->tlock);
1284 #if defined(AFS_SUN5_ENV)
1285 ObtainWriteLock(&avc->vlock, 545);
1286 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1287 avc->vstates &= ~VRevokeWait;
1288 afs_osi_Wakeup((char *)&avc->vstates);
1290 ReleaseWriteLock(&avc->vlock);
1292 MReleaseWriteLock(&afs_xdcache);
1294 * It's treated like a callback so that when we do lookups we'll invalidate the unique bit if any
1295 * trytoSmush occured during the lookup call
1304 * Given the cached info for a file and a byte offset into the
1305 * file, make sure the dcache entry for that file and containing
1306 * the given byte is available, returning it to our caller.
1309 * avc : Pointer to the (held) vcache entry to look in.
1310 * abyte : Which byte we want to get to.
1313 * Pointer to the dcache entry covering the file & desired byte,
1314 * or NULL if not found.
1317 * The vcache entry is held upon entry.
1321 afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1324 register afs_int32 i, index;
1325 register struct dcache *tdc = NULL;
1327 AFS_STATCNT(afs_FindDCache);
1328 chunk = AFS_CHUNK(abyte);
1331 * Hash on the [fid, chunk] and get the corresponding dcache index
1332 * after write-locking the dcache.
1334 i = DCHash(&avc->fid, chunk);
1335 MObtainWriteLock(&afs_xdcache, 278);
1336 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1337 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1338 tdc = afs_GetDSlot(index, NULL);
1339 ReleaseReadLock(&tdc->tlock);
1340 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1341 break; /* leaving refCount high for caller */
1346 index = afs_dcnextTbl[index];
1348 MReleaseWriteLock(&afs_xdcache);
1349 if (index != NULLIDX) {
1350 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1351 hadd32(afs_indexCounter, 1);
1356 } /*afs_FindDCache */
1360 * afs_UFSCacheStoreProc
1363 * Called upon store.
1366 * acall : Ptr to the Rx call structure involved.
1367 * afile : Ptr to the related file descriptor.
1368 * alen : Size of the file in bytes.
1369 * avc : Ptr to the vcache entry.
1370 * shouldWake : is it "safe" to return early from close() ?
1371 * abytesToXferP : Set to the number of bytes to xfer.
1372 * NOTE: This parameter is only used if AFS_NOSTATS
1374 * abytesXferredP : Set to the number of bytes actually xferred.
1375 * NOTE: This parameter is only used if AFS_NOSTATS
1379 * Nothing interesting.
1382 afs_UFSCacheStoreProc(register struct rx_call *acall, struct osi_file *afile,
1383 register afs_int32 alen, struct vcache *avc,
1384 int *shouldWake, afs_size_t * abytesToXferP,
1385 afs_size_t * abytesXferredP)
1387 afs_int32 code, got;
1388 register char *tbuffer;
1391 AFS_STATCNT(UFS_CacheStoreProc);
1395 * In this case, alen is *always* the amount of data we'll be trying
1398 (*abytesToXferP) = alen;
1399 (*abytesXferredP) = 0;
1400 #endif /* AFS_NOSTATS */
1402 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1403 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1404 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1405 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1407 tlen = (alen > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : alen);
1408 got = afs_osi_Read(afile, -1, tbuffer, tlen);
1410 #if defined(KERNEL_HAVE_UERROR)
1411 || (got != tlen && getuerror())
1414 osi_FreeLargeSpace(tbuffer);
1417 afs_Trace2(afs_iclSetp, CM_TRACE_STOREPROC2, ICL_TYPE_OFFSET,
1418 ICL_HANDLE_OFFSET(*tbuffer), ICL_TYPE_INT32, got);
1420 code = rx_Write(acall, tbuffer, got); /* writing 0 bytes will
1421 * push a short packet. Is that really what we want, just because the
1422 * data didn't come back from the disk yet? Let's try it and see. */
1425 (*abytesXferredP) += code;
1426 #endif /* AFS_NOSTATS */
1428 code = rx_Error(acall);
1429 osi_FreeLargeSpace(tbuffer);
1430 return code ? code : -33;
1434 * If file has been locked on server, we can allow the store
1437 if (shouldWake && *shouldWake && (rx_GetRemoteStatus(acall) & 1)) {
1438 *shouldWake = 0; /* only do this once */
1442 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1443 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1444 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1445 osi_FreeLargeSpace(tbuffer);
1448 } /* afs_UFSCacheStoreProc */
1452 * afs_UFSCacheFetchProc
1455 * Routine called on fetch; also tells people waiting for data
1456 * that more has arrived.
1459 * acall : Ptr to the Rx call structure.
1460 * afile : File descriptor for the cache file.
1461 * abase : Base offset to fetch.
1462 * adc : Ptr to the dcache entry for the file, write-locked.
1463 * avc : Ptr to the vcache entry for the file.
1464 * abytesToXferP : Set to the number of bytes to xfer.
1465 * NOTE: This parameter is only used if AFS_NOSTATS
1467 * abytesXferredP : Set to the number of bytes actually xferred.
1468 * NOTE: This parameter is only used if AFS_NOSTATS
1472 * Nothing interesting.
1476 afs_UFSCacheFetchProc(register struct rx_call *acall, struct osi_file *afile,
1477 afs_size_t abase, struct dcache *adc,
1478 struct vcache *avc, afs_size_t * abytesToXferP,
1479 afs_size_t * abytesXferredP, afs_int32 lengthFound)
1482 register afs_int32 code;
1483 register char *tbuffer;
1487 AFS_STATCNT(UFS_CacheFetchProc);
1488 osi_Assert(WriteLocked(&adc->lock));
1489 afile->offset = 0; /* Each time start from the beginning */
1490 length = lengthFound;
1492 (*abytesToXferP) = 0;
1493 (*abytesXferredP) = 0;
1494 #endif /* AFS_NOSTATS */
1495 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1496 adc->validPos = abase;
1500 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
1502 length = ntohl(length);
1503 if (code != sizeof(afs_int32)) {
1504 osi_FreeLargeSpace(tbuffer);
1505 code = rx_Error(acall);
1506 return (code ? code : -1); /* try to return code, not -1 */
1510 * The fetch protocol is extended for the AFS/DFS translator
1511 * to allow multiple blocks of data, each with its own length,
1512 * to be returned. As long as the top bit is set, there are more
1515 * We do not do this for AFS file servers because they sometimes
1516 * return large negative numbers as the transfer size.
1518 if (avc->states & CForeign) {
1519 moredata = length & 0x80000000;
1520 length &= ~0x80000000;
1525 (*abytesToXferP) += length;
1526 #endif /* AFS_NOSTATS */
1527 while (length > 0) {
1528 tlen = (length > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : length);
1529 #ifdef RX_KERNEL_TRACE
1530 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1534 code = rx_Read(acall, tbuffer, tlen);
1536 #ifdef RX_KERNEL_TRACE
1537 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1541 (*abytesXferredP) += code;
1542 #endif /* AFS_NOSTATS */
1544 osi_FreeLargeSpace(tbuffer);
1545 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64READ,
1546 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
1547 ICL_TYPE_INT32, length);
1550 code = afs_osi_Write(afile, -1, tbuffer, tlen);
1552 osi_FreeLargeSpace(tbuffer);
1557 adc->validPos = abase;
1558 if (afs_osi_Wakeup(&adc->validPos) == 0)
1559 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
1560 __FILE__, ICL_TYPE_INT32, __LINE__,
1561 ICL_TYPE_POINTER, adc, ICL_TYPE_INT32,
1565 osi_FreeLargeSpace(tbuffer);
1568 } /* afs_UFSCacheFetchProc */
1574 * This function is called to obtain a reference to data stored in
1575 * the disk cache, locating a chunk of data containing the desired
1576 * byte and returning a reference to the disk cache entry, with its
1577 * reference count incremented.
1581 * avc : Ptr to a vcache entry (unlocked)
1582 * abyte : Byte position in the file desired
1583 * areq : Request structure identifying the requesting user.
1584 * aflags : Settings as follows:
1586 * 2 : Return after creating entry.
1587 * 4 : called from afs_vnop_write.c
1588 * *alen contains length of data to be written.
1590 * aoffset : Set to the offset within the chunk where the resident
1592 * alen : Set to the number of bytes of data after the desired
1593 * byte (including the byte itself) which can be read
1597 * The vcache entry pointed to by avc is unlocked upon entry.
1601 struct AFSVolSync tsync;
1602 struct AFSFetchStatus OutStatus;
1603 struct AFSCallBack CallBack;
1607 * Update the vnode-to-dcache hint if we can get the vnode lock
1608 * right away. Assumes dcache entry is at least read-locked.
1611 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1613 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1614 if (hsame(v->m.DataVersion, d->f.versionNo) && v->callback)
1617 ReleaseWriteLock(&v->lock);
1621 /* avc - Write-locked unless aflags & 1 */
1623 afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1624 register struct vrequest *areq, afs_size_t * aoffset,
1625 afs_size_t * alen, int aflags)
1627 register afs_int32 i, code, code1 = 0, shortcut;
1628 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1629 register afs_int32 adjustsize = 0;
1635 afs_size_t maxGoodLength; /* amount of good data at server */
1636 struct rx_call *tcall;
1637 afs_size_t Position = 0;
1638 #ifdef AFS_64BIT_CLIENT
1640 afs_size_t lengthFound; /* as returned from server */
1641 #endif /* AFS_64BIT_CLIENT */
1642 afs_int32 size, tlen; /* size of segment to transfer */
1643 struct tlocal1 *tsmall = 0;
1644 register struct dcache *tdc;
1645 register struct osi_file *file;
1646 register struct conn *tc;
1648 struct server *newCallback = NULL;
1649 char setNewCallback;
1650 char setVcacheStatus;
1651 char doVcacheUpdate;
1653 int doAdjustSize = 0;
1654 int doReallyAdjustSize = 0;
1655 int overWriteWholeChunk = 0;
1659 struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
1660 osi_timeval_t xferStartTime, /*FS xfer start time */
1661 xferStopTime; /*FS xfer stop time */
1662 afs_size_t bytesToXfer; /* # bytes to xfer */
1663 afs_size_t bytesXferred; /* # bytes actually xferred */
1664 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1665 int fromReplica; /*Are we reading from a replica? */
1666 int numFetchLoops; /*# times around the fetch/analyze loop */
1667 #endif /* AFS_NOSTATS */
1669 AFS_STATCNT(afs_GetDCache);
1674 setLocks = aflags & 1;
1677 * Determine the chunk number and offset within the chunk corresponding
1678 * to the desired byte.
1680 if (avc->fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1683 chunk = AFS_CHUNK(abyte);
1686 /* come back to here if we waited for the cache to drain. */
1689 setNewCallback = setVcacheStatus = 0;
1693 ObtainWriteLock(&avc->lock, 616);
1695 ObtainReadLock(&avc->lock);
1700 * avc->lock(R) if setLocks && !slowPass
1701 * avc->lock(W) if !setLocks || slowPass
1706 /* check hints first! (might could use bcmp or some such...) */
1707 if ((tdc = avc->dchint)) {
1711 * The locking order between afs_xdcache and dcache lock matters.
1712 * The hint dcache entry could be anywhere, even on the free list.
1713 * Locking afs_xdcache ensures that noone is trying to pull dcache
1714 * entries from the free list, and thereby assuming them to be not
1715 * referenced and not locked.
1717 MObtainReadLock(&afs_xdcache);
1718 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1720 if (dcLocked && (tdc->index != NULLIDX)
1721 && !FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk
1722 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1723 /* got the right one. It might not be the right version, and it
1724 * might be fetching, but it's the right dcache entry.
1726 /* All this code should be integrated better with what follows:
1727 * I can save a good bit more time under a write lock if I do..
1729 ObtainWriteLock(&tdc->tlock, 603);
1731 ReleaseWriteLock(&tdc->tlock);
1733 MReleaseReadLock(&afs_xdcache);
1736 if (hsame(tdc->f.versionNo, avc->m.DataVersion)
1737 && !(tdc->dflags & DFFetching)) {
1739 afs_stats_cmperf.dcacheHits++;
1740 MObtainWriteLock(&afs_xdcache, 559);
1741 QRemove(&tdc->lruq);
1742 QAdd(&afs_DLRU, &tdc->lruq);
1743 MReleaseWriteLock(&afs_xdcache);
1746 * avc->lock(R) if setLocks && !slowPass
1747 * avc->lock(W) if !setLocks || slowPass
1754 ReleaseSharedLock(&tdc->lock);
1755 MReleaseReadLock(&afs_xdcache);
1763 * avc->lock(R) if setLocks && !slowPass
1764 * avc->lock(W) if !setLocks || slowPass
1765 * tdc->lock(S) if tdc
1768 if (!tdc) { /* If the hint wasn't the right dcache entry */
1770 * Hash on the [fid, chunk] and get the corresponding dcache index
1771 * after write-locking the dcache.
1776 * avc->lock(R) if setLocks && !slowPass
1777 * avc->lock(W) if !setLocks || slowPass
1780 i = DCHash(&avc->fid, chunk);
1781 /* check to make sure our space is fine */
1782 afs_MaybeWakeupTruncateDaemon();
1784 MObtainWriteLock(&afs_xdcache, 280);
1786 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1787 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1788 tdc = afs_GetDSlot(index, NULL);
1789 ReleaseReadLock(&tdc->tlock);
1792 * avc->lock(R) if setLocks && !slowPass
1793 * avc->lock(W) if !setLocks || slowPass
1796 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1797 /* Move it up in the beginning of the list */
1798 if (afs_dchashTbl[i] != index) {
1799 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1800 afs_dcnextTbl[index] = afs_dchashTbl[i];
1801 afs_dchashTbl[i] = index;
1803 MReleaseWriteLock(&afs_xdcache);
1804 ObtainSharedLock(&tdc->lock, 606);
1805 break; /* leaving refCount high for caller */
1811 index = afs_dcnextTbl[index];
1815 * If we didn't find the entry, we'll create one.
1817 if (index == NULLIDX) {
1820 * avc->lock(R) if setLocks
1821 * avc->lock(W) if !setLocks
1824 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1825 avc, ICL_TYPE_INT32, chunk);
1827 /* Make sure there is a free dcache entry for us to use */
1828 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1831 avc->states |= CDCLock;
1832 /* just need slots */
1833 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1835 avc->states &= ~CDCLock;
1836 if (afs_discardDCList != NULLIDX
1837 || afs_freeDCList != NULLIDX)
1839 /* If we can't get space for 5 mins we give up and panic */
1840 if (++downDCount > 300)
1841 osi_Panic("getdcache");
1842 MReleaseWriteLock(&afs_xdcache);
1845 * avc->lock(R) if setLocks
1846 * avc->lock(W) if !setLocks
1848 afs_osi_Wait(1000, 0, 0);
1853 if (afs_discardDCList == NULLIDX
1854 || ((aflags & 2) && afs_freeDCList != NULLIDX)) {
1856 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1857 tdc = afs_GetDSlot(afs_freeDCList, 0);
1858 osi_Assert(tdc->refCount == 1);
1859 ReleaseReadLock(&tdc->tlock);
1860 ObtainWriteLock(&tdc->lock, 604);
1861 afs_freeDCList = afs_dvnextTbl[tdc->index];
1864 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1865 tdc = afs_GetDSlot(afs_discardDCList, 0);
1866 osi_Assert(tdc->refCount == 1);
1867 ReleaseReadLock(&tdc->tlock);
1868 ObtainWriteLock(&tdc->lock, 605);
1869 afs_discardDCList = afs_dvnextTbl[tdc->index];
1870 afs_discardDCCount--;
1872 ((tdc->f.chunkBytes +
1873 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1874 tdc->f.states &= ~(DRO|DBackup|DRW);
1875 afs_DCMoveBucket(tdc, size, 0);
1876 afs_blocksDiscarded -= size;
1877 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1879 /* Truncate the chunk so zeroes get filled properly */
1880 file = afs_CFileOpen(tdc->f.inode);
1881 afs_CFileTruncate(file, 0);
1882 afs_CFileClose(file);
1883 afs_AdjustSize(tdc, 0);
1889 * avc->lock(R) if setLocks
1890 * avc->lock(W) if !setLocks
1896 * Fill in the newly-allocated dcache record.
1898 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1899 tdc->f.fid = avc->fid;
1900 if (avc->states & CRO)
1901 tdc->f.states = DRO;
1902 else if (avc->states & CBackup)
1903 tdc->f.states = DBackup;
1905 tdc->f.states = DRW;
1906 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1907 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1908 hones(tdc->f.versionNo); /* invalid value */
1909 tdc->f.chunk = chunk;
1910 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1912 if (tdc->lruq.prev == &tdc->lruq)
1913 osi_Panic("lruq 1");
1916 * Now add to the two hash chains - note that i is still set
1917 * from the above DCHash call.
1919 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1920 afs_dchashTbl[i] = tdc->index;
1921 i = DVHash(&avc->fid);
1922 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1923 afs_dvhashTbl[i] = tdc->index;
1924 tdc->dflags = DFEntryMod;
1926 afs_MaybeWakeupTruncateDaemon();
1927 MReleaseWriteLock(&afs_xdcache);
1928 ConvertWToSLock(&tdc->lock);
1933 /* vcache->dcache hint failed */
1936 * avc->lock(R) if setLocks && !slowPass
1937 * avc->lock(W) if !setLocks || slowPass
1940 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1941 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1942 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1943 hgetlo(avc->m.DataVersion));
1945 * Here we have the entry in tdc, with its refCount incremented.
1946 * Note: we don't use the S-lock on avc; it costs concurrency when
1947 * storing a file back to the server.
1951 * Not a newly created file so we need to check the file's length and
1952 * compare data versions since someone could have changed the data or we're
1953 * reading a file written elsewhere. We only want to bypass doing no-op
1954 * read rpcs on newly created files (dv of 0) since only then we guarantee
1955 * that this chunk's data hasn't been filled by another client.
1957 size = AFS_CHUNKSIZE(abyte);
1958 if (aflags & 4) /* called from write */
1960 else /* called from read */
1961 tlen = tdc->validPos - abyte;
1962 Position = AFS_CHUNKTOBASE(chunk);
1963 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1964 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1965 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1966 ICL_HANDLE_OFFSET(Position));
1967 if ((aflags & 4) && (hiszero(avc->m.DataVersion)))
1969 if ((aflags & 4) && (abyte == Position) && (tlen >= size))
1970 overWriteWholeChunk = 1;
1971 if (doAdjustSize || overWriteWholeChunk) {
1972 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1974 #ifdef AFS_SGI64_ENV
1977 #else /* AFS_SGI64_ENV */
1980 #endif /* AFS_SGI64_ENV */
1981 #else /* AFS_SGI_ENV */
1984 #endif /* AFS_SGI_ENV */
1985 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->m.Length &&
1986 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1987 #if defined(AFS_SUN5_ENV) || defined(AFS_OSF_ENV)
1988 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
1990 if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
1992 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1993 !hsame(avc->m.DataVersion, tdc->f.versionNo))
1994 doReallyAdjustSize = 1;
1996 if (doReallyAdjustSize || overWriteWholeChunk) {
1997 /* no data in file to read at this position */
1998 UpgradeSToWLock(&tdc->lock, 607);
2000 file = afs_CFileOpen(tdc->f.inode);
2001 afs_CFileTruncate(file, 0);
2002 afs_CFileClose(file);
2003 afs_AdjustSize(tdc, 0);
2004 hset(tdc->f.versionNo, avc->m.DataVersion);
2005 tdc->dflags |= DFEntryMod;
2007 ConvertWToSLock(&tdc->lock);
2012 * We must read in the whole chunk if the version number doesn't
2016 /* don't need data, just a unique dcache entry */
2017 ObtainWriteLock(&afs_xdcache, 608);
2018 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2019 hadd32(afs_indexCounter, 1);
2020 ReleaseWriteLock(&afs_xdcache);
2022 updateV2DC(setLocks, avc, tdc, 553);
2023 if (vType(avc) == VDIR)
2026 *aoffset = AFS_CHUNKOFFSET(abyte);
2027 if (tdc->validPos < abyte)
2028 *alen = (afs_size_t) 0;
2030 *alen = tdc->validPos - abyte;
2031 ReleaseSharedLock(&tdc->lock);
2034 ReleaseWriteLock(&avc->lock);
2036 ReleaseReadLock(&avc->lock);
2038 return tdc; /* check if we're done */
2043 * avc->lock(R) if setLocks && !slowPass
2044 * avc->lock(W) if !setLocks || slowPass
2047 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2049 setNewCallback = setVcacheStatus = 0;
2053 * avc->lock(R) if setLocks && !slowPass
2054 * avc->lock(W) if !setLocks || slowPass
2057 if (!hsame(avc->m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2059 * Version number mismatch.
2061 UpgradeSToWLock(&tdc->lock, 609);
2064 * If data ever existed for this vnode, and this is a text object,
2065 * do some clearing. Now, you'd think you need only do the flush
2066 * when VTEXT is on, but VTEXT is turned off when the text object
2067 * is freed, while pages are left lying around in memory marked
2068 * with this vnode. If we would reactivate (create a new text
2069 * object from) this vnode, we could easily stumble upon some of
2070 * these old pages in pagein. So, we always flush these guys.
2071 * Sun has a wonderful lack of useful invariants in this system.
2073 * avc->flushDV is the data version # of the file at the last text
2074 * flush. Clearly, at least, we don't have to flush the file more
2075 * often than it changes
2077 if (hcmp(avc->flushDV, avc->m.DataVersion) < 0) {
2079 * By here, the cache entry is always write-locked. We can
2080 * deadlock if we call osi_Flush with the cache entry locked...
2081 * Unlock the dcache too.
2083 ReleaseWriteLock(&tdc->lock);
2084 if (setLocks && !slowPass)
2085 ReleaseReadLock(&avc->lock);
2087 ReleaseWriteLock(&avc->lock);
2091 * Call osi_FlushPages in open, read/write, and map, since it
2092 * is too hard here to figure out if we should lock the
2095 if (setLocks && !slowPass)
2096 ObtainReadLock(&avc->lock);
2098 ObtainWriteLock(&avc->lock, 66);
2099 ObtainWriteLock(&tdc->lock, 610);
2104 * avc->lock(R) if setLocks && !slowPass
2105 * avc->lock(W) if !setLocks || slowPass
2109 /* Watch for standard race condition around osi_FlushText */
2110 if (hsame(avc->m.DataVersion, tdc->f.versionNo)) {
2111 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2112 afs_stats_cmperf.dcacheHits++;
2113 ConvertWToSLock(&tdc->lock);
2117 /* Sleep here when cache needs to be drained. */
2118 if (setLocks && !slowPass
2119 && (afs_blocksUsed >
2120 (CM_WAITFORDRAINPCT * afs_cacheBlocks) / 100)) {
2121 /* Make sure truncate daemon is running */
2122 afs_MaybeWakeupTruncateDaemon();
2123 ObtainWriteLock(&tdc->tlock, 614);
2124 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2125 ReleaseWriteLock(&tdc->tlock);
2126 ReleaseWriteLock(&tdc->lock);
2127 ReleaseReadLock(&avc->lock);
2128 while ((afs_blocksUsed - afs_blocksDiscarded) >
2129 (CM_WAITFORDRAINPCT * afs_cacheBlocks) / 100) {
2130 afs_WaitForCacheDrain = 1;
2131 afs_osi_Sleep(&afs_WaitForCacheDrain);
2133 afs_MaybeFreeDiscardedDCache();
2134 /* need to check if someone else got the chunk first. */
2135 goto RetryGetDCache;
2138 /* Do not fetch data beyond truncPos. */
2139 maxGoodLength = avc->m.Length;
2140 if (avc->truncPos < maxGoodLength)
2141 maxGoodLength = avc->truncPos;
2142 Position = AFS_CHUNKBASE(abyte);
2143 if (vType(avc) == VDIR) {
2144 size = avc->m.Length;
2145 if (size > tdc->f.chunkBytes) {
2146 /* pre-reserve space for file */
2147 afs_AdjustSize(tdc, size);
2149 size = 999999999; /* max size for transfer */
2151 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2152 /* don't read past end of good data on server */
2153 if (Position + size > maxGoodLength)
2154 size = maxGoodLength - Position;
2156 size = 0; /* Handle random races */
2157 if (size > tdc->f.chunkBytes) {
2158 /* pre-reserve space for file */
2159 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2160 /* max size for transfer still in size */
2163 if (afs_mariner && !tdc->f.chunk)
2164 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2166 * Right now, we only have one tool, and it's a hammer. So, we
2167 * fetch the whole file.
2169 DZap(tdc); /* pages in cache may be old */
2170 file = afs_CFileOpen(tdc->f.inode);
2171 afs_RemoveVCB(&avc->fid);
2172 tdc->f.states |= DWriting;
2173 tdc->dflags |= DFFetching;
2174 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2175 if (tdc->mflags & DFFetchReq) {
2176 tdc->mflags &= ~DFFetchReq;
2177 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2178 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2179 __FILE__, ICL_TYPE_INT32, __LINE__,
2180 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2184 (struct tlocal1 *)osi_AllocLargeSpace(sizeof(struct tlocal1));
2185 setVcacheStatus = 0;
2188 * Remember if we are doing the reading from a replicated volume,
2189 * and how many times we've zipped around the fetch/analyze loop.
2191 fromReplica = (avc->states & CRO) ? 1 : 0;
2193 accP = &(afs_stats_cmfullperf.accessinf);
2195 (accP->replicatedRefs)++;
2197 (accP->unreplicatedRefs)++;
2198 #endif /* AFS_NOSTATS */
2199 /* this is a cache miss */
2200 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2201 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
2202 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2205 afs_stats_cmperf.dcacheMisses++;
2208 * Dynamic root support: fetch data from local memory.
2210 if (afs_IsDynroot(avc)) {
2214 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2216 dynrootDir += Position;
2217 dynrootLen -= Position;
2218 if (size > dynrootLen)
2222 code = afs_CFileWrite(file, 0, dynrootDir, size);
2230 tdc->validPos = Position + size;
2231 afs_CFileTruncate(file, size); /* prune it */
2234 * Not a dynamic vnode: do the real fetch.
2239 * avc->lock(R) if setLocks && !slowPass
2240 * avc->lock(W) if !setLocks || slowPass
2244 tc = afs_Conn(&avc->fid, areq, SHARED_LOCK);
2246 afs_int32 length_hi, length, bytes;
2250 (accP->numReplicasAccessed)++;
2252 #endif /* AFS_NOSTATS */
2253 if (!setLocks || slowPass) {
2254 avc->callback = tc->srvr->server;
2256 newCallback = tc->srvr->server;
2261 tcall = rx_NewCall(tc->id);
2264 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
2265 #ifdef AFS_64BIT_CLIENT
2266 length_hi = code = 0;
2267 if (!afs_serverHasNo64Bit(tc)) {
2271 StartRXAFS_FetchData64(tcall,
2272 (struct AFSFid *)&avc->fid.
2273 Fid, Position, tsize);
2276 afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
2277 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32,
2281 rx_Read(tcall, (char *)&length_hi,
2284 if (bytes == sizeof(afs_int32)) {
2285 length_hi = ntohl(length_hi);
2288 code = rx_Error(tcall);
2290 code1 = rx_EndCall(tcall, code);
2292 tcall = (struct rx_call *)0;
2296 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
2297 if (Position > 0x7FFFFFFF) {
2304 tcall = rx_NewCall(tc->id);
2306 StartRXAFS_FetchData(tcall, (struct AFSFid *)
2311 afs_serverSetNo64Bit(tc);
2316 rx_Read(tcall, (char *)&length,
2319 if (bytes == sizeof(afs_int32)) {
2320 length = ntohl(length);
2322 code = rx_Error(tcall);
2325 FillInt64(lengthFound, length_hi, length);
2326 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64LENG,
2327 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
2329 ICL_HANDLE_OFFSET(lengthFound));
2330 #else /* AFS_64BIT_CLIENT */
2333 StartRXAFS_FetchData(tcall,
2334 (struct AFSFid *)&avc->fid.Fid,
2340 rx_Read(tcall, (char *)&length,
2343 if (bytes == sizeof(afs_int32)) {
2344 length = ntohl(length);
2346 code = rx_Error(tcall);
2349 #endif /* AFS_64BIT_CLIENT */
2354 &(afs_stats_cmfullperf.rpc.
2355 fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
2356 osi_GetuTime(&xferStartTime);
2359 afs_CacheFetchProc(tcall, file,
2360 (afs_size_t) Position, tdc,
2362 &bytesXferred, length);
2364 osi_GetuTime(&xferStopTime);
2365 (xferP->numXfers)++;
2367 (xferP->numSuccesses)++;
2368 afs_stats_XferSumBytes
2369 [AFS_STATS_FS_XFERIDX_FETCHDATA] +=
2371 (xferP->sumBytes) +=
2372 (afs_stats_XferSumBytes
2373 [AFS_STATS_FS_XFERIDX_FETCHDATA] >> 10);
2374 afs_stats_XferSumBytes
2375 [AFS_STATS_FS_XFERIDX_FETCHDATA] &= 0x3FF;
2376 if (bytesXferred < xferP->minBytes)
2377 xferP->minBytes = bytesXferred;
2378 if (bytesXferred > xferP->maxBytes)
2379 xferP->maxBytes = bytesXferred;
2382 * Tally the size of the object. Note: we tally the actual size,
2383 * NOT the number of bytes that made it out over the wire.
2385 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
2386 (xferP->count[0])++;
2387 else if (bytesToXfer <=
2388 AFS_STATS_MAXBYTES_BUCKET1)
2389 (xferP->count[1])++;
2390 else if (bytesToXfer <=
2391 AFS_STATS_MAXBYTES_BUCKET2)
2392 (xferP->count[2])++;
2393 else if (bytesToXfer <=
2394 AFS_STATS_MAXBYTES_BUCKET3)
2395 (xferP->count[3])++;
2396 else if (bytesToXfer <=
2397 AFS_STATS_MAXBYTES_BUCKET4)
2398 (xferP->count[4])++;
2399 else if (bytesToXfer <=
2400 AFS_STATS_MAXBYTES_BUCKET5)
2401 (xferP->count[5])++;
2402 else if (bytesToXfer <=
2403 AFS_STATS_MAXBYTES_BUCKET6)
2404 (xferP->count[6])++;
2405 else if (bytesToXfer <=
2406 AFS_STATS_MAXBYTES_BUCKET7)
2407 (xferP->count[7])++;
2409 (xferP->count[8])++;
2411 afs_stats_GetDiff(elapsedTime, xferStartTime,
2413 afs_stats_AddTo((xferP->sumTime), elapsedTime);
2414 afs_stats_SquareAddTo((xferP->sqrTime),
2416 if (afs_stats_TimeLessThan
2417 (elapsedTime, (xferP->minTime))) {
2418 afs_stats_TimeAssign((xferP->minTime),
2421 if (afs_stats_TimeGreaterThan
2422 (elapsedTime, (xferP->maxTime))) {
2423 afs_stats_TimeAssign((xferP->maxTime),
2429 afs_CacheFetchProc(tcall, file, Position, tdc,
2431 #endif /* AFS_NOSTATS */
2436 EndRXAFS_FetchData(tcall, &tsmall->OutStatus,
2444 code1 = rx_EndCall(tcall, code);
2453 /* callback could have been broken (or expired) in a race here,
2454 * but we return the data anyway. It's as good as we knew about
2455 * when we started. */
2457 * validPos is updated by CacheFetchProc, and can only be
2458 * modifed under a dcache write lock, which we've blocked out
2460 size = tdc->validPos - Position; /* actual segment size */
2463 afs_CFileTruncate(file, size); /* prune it */
2465 if (!setLocks || slowPass) {
2466 ObtainWriteLock(&afs_xcbhash, 453);
2467 afs_DequeueCallback(avc);
2468 avc->states &= ~(CStatd | CUnique);
2469 avc->callback = NULL;
2470 ReleaseWriteLock(&afs_xcbhash);
2471 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2472 osi_dnlc_purgedp(avc);
2474 /* Something lost. Forget about performance, and go
2475 * back with a vcache write lock.
2477 afs_CFileTruncate(file, 0);
2478 afs_AdjustSize(tdc, 0);
2479 afs_CFileClose(file);
2480 osi_FreeLargeSpace(tsmall);
2482 ReleaseWriteLock(&tdc->lock);
2485 ReleaseReadLock(&avc->lock);
2487 goto RetryGetDCache;
2491 } while (afs_Analyze
2492 (tc, code, &avc->fid, areq,
2493 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2497 * avc->lock(R) if setLocks && !slowPass
2498 * avc->lock(W) if !setLocks || slowPass
2504 * In the case of replicated access, jot down info on the number of
2505 * attempts it took before we got through or gave up.
2508 if (numFetchLoops <= 1)
2509 (accP->refFirstReplicaOK)++;
2510 if (numFetchLoops > accP->maxReplicasPerRef)
2511 accP->maxReplicasPerRef = numFetchLoops;
2513 #endif /* AFS_NOSTATS */
2515 tdc->dflags &= ~DFFetching;
2516 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2517 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2518 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2519 tdc, ICL_TYPE_INT32, tdc->dflags);
2520 if (avc->execsOrWriters == 0)
2521 tdc->f.states &= ~DWriting;
2523 /* now, if code != 0, we have an error and should punt.
2524 * note that we have the vcache write lock, either because
2525 * !setLocks or slowPass.
2528 afs_CFileTruncate(file, 0);
2529 afs_AdjustSize(tdc, 0);
2530 afs_CFileClose(file);
2531 ZapDCE(tdc); /* sets DFEntryMod */
2532 if (vType(avc) == VDIR) {
2535 tdc->f.states &= ~(DRO|DBackup|DRW);
2536 afs_DCMoveBucket(tdc, 0, 0);
2537 ReleaseWriteLock(&tdc->lock);
2539 ObtainWriteLock(&afs_xcbhash, 454);
2540 afs_DequeueCallback(avc);
2541 avc->states &= ~(CStatd | CUnique);
2542 ReleaseWriteLock(&afs_xcbhash);
2543 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2544 osi_dnlc_purgedp(avc);
2547 * avc->lock(W); assert(!setLocks || slowPass)
2549 osi_Assert(!setLocks || slowPass);
2550 tdc->f.states &= ~(DRO|DBackup|DRW);
2551 afs_DCMoveBucket(tdc, 0, 0);
2556 /* otherwise we copy in the just-fetched info */
2557 afs_CFileClose(file);
2558 afs_AdjustSize(tdc, size); /* new size */
2560 * Copy appropriate fields into vcache. Status is
2561 * copied later where we selectively acquire the
2562 * vcache write lock.
2565 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2567 setVcacheStatus = 1;
2568 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2569 tsmall->OutStatus.DataVersion);
2570 tdc->dflags |= DFEntryMod;
2571 afs_indexFlags[tdc->index] |= IFEverUsed;
2572 ConvertWToSLock(&tdc->lock);
2573 } /*Data version numbers don't match */
2576 * Data version numbers match.
2578 afs_stats_cmperf.dcacheHits++;
2579 } /*Data version numbers match */
2581 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2585 * avc->lock(R) if setLocks && !slowPass
2586 * avc->lock(W) if !setLocks || slowPass
2587 * tdc->lock(S) if tdc
2591 * See if this was a reference to a file in the local cell.
2593 if (afs_IsPrimaryCellNum(avc->fid.Cell))
2594 afs_stats_cmperf.dlocalAccesses++;
2596 afs_stats_cmperf.dremoteAccesses++;
2598 /* Fix up LRU info */
2601 MObtainWriteLock(&afs_xdcache, 602);
2602 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2603 hadd32(afs_indexCounter, 1);
2604 MReleaseWriteLock(&afs_xdcache);
2606 /* return the data */
2607 if (vType(avc) == VDIR)
2610 *aoffset = AFS_CHUNKOFFSET(abyte);
2611 *alen = (tdc->f.chunkBytes - *aoffset);
2612 ReleaseSharedLock(&tdc->lock);
2617 * avc->lock(R) if setLocks && !slowPass
2618 * avc->lock(W) if !setLocks || slowPass
2621 /* Fix up the callback and status values in the vcache */
2623 if (setLocks && !slowPass) {
2626 * This is our dirty little secret to parallel fetches.
2627 * We don't write-lock the vcache while doing the fetch,
2628 * but potentially we'll need to update the vcache after
2629 * the fetch is done.
2631 * Drop the read lock and try to re-obtain the write
2632 * lock. If the vcache still has the same DV, it's
2633 * ok to go ahead and install the new data.
2635 afs_hyper_t currentDV, statusDV;
2637 hset(currentDV, avc->m.DataVersion);
2639 if (setNewCallback && avc->callback != newCallback)
2643 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2644 tsmall->OutStatus.DataVersion);
2646 if (setVcacheStatus && avc->m.Length != tsmall->OutStatus.Length)
2648 if (setVcacheStatus && !hsame(currentDV, statusDV))
2652 ReleaseReadLock(&avc->lock);
2654 if (doVcacheUpdate) {
2655 ObtainWriteLock(&avc->lock, 615);
2656 if (!hsame(avc->m.DataVersion, currentDV)) {
2657 /* We lose. Someone will beat us to it. */
2659 ReleaseWriteLock(&avc->lock);
2664 /* With slow pass, we've already done all the updates */
2666 ReleaseWriteLock(&avc->lock);
2669 /* Check if we need to perform any last-minute fixes with a write-lock */
2670 if (!setLocks || doVcacheUpdate) {
2672 avc->callback = newCallback;
2673 if (tsmall && setVcacheStatus)
2674 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2676 ReleaseWriteLock(&avc->lock);
2680 osi_FreeLargeSpace(tsmall);
2683 } /*afs_GetDCache */
2687 * afs_WriteThroughDSlots
2690 * Sweep through the dcache slots and write out any modified
2691 * in-memory data back on to our caching store.
2697 * The afs_xdcache is write-locked through this whole affair.
2700 afs_WriteThroughDSlots(void)
2702 register struct dcache *tdc;
2703 register afs_int32 i, touchedit = 0;
2705 struct afs_q DirtyQ, *tq;
2707 AFS_STATCNT(afs_WriteThroughDSlots);
2710 * Because of lock ordering, we can't grab dcache locks while
2711 * holding afs_xdcache. So we enter xdcache, get a reference
2712 * for every dcache entry, and exit xdcache.
2714 MObtainWriteLock(&afs_xdcache, 283);
2716 for (i = 0; i < afs_cacheFiles; i++) {
2717 tdc = afs_indexTable[i];
2719 /* Grab tlock in case the existing refcount isn't zero */
2720 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2721 ObtainWriteLock(&tdc->tlock, 623);
2723 ReleaseWriteLock(&tdc->tlock);
2725 QAdd(&DirtyQ, &tdc->dirty);
2728 MReleaseWriteLock(&afs_xdcache);
2731 * Now, for each dcache entry we found, check if it's dirty.
2732 * If so, get write-lock, get afs_xdcache, which protects
2733 * afs_cacheInodep, and flush it. Don't forget to put back
2737 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2739 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2741 if (tdc->dflags & DFEntryMod) {
2744 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2746 /* Now that we have the write lock, double-check */
2747 if (wrLock && (tdc->dflags & DFEntryMod)) {
2748 tdc->dflags &= ~DFEntryMod;
2749 MObtainWriteLock(&afs_xdcache, 620);
2750 afs_WriteDCache(tdc, 1);
2751 MReleaseWriteLock(&afs_xdcache);
2755 ReleaseWriteLock(&tdc->lock);
2761 MObtainWriteLock(&afs_xdcache, 617);
2762 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2763 /* Touch the file to make sure that the mtime on the file is kept
2764 * up-to-date to avoid losing cached files on cold starts because
2765 * their mtime seems old...
2767 struct afs_fheader theader;
2769 theader.magic = AFS_FHMAGIC;
2770 theader.firstCSize = AFS_FIRSTCSIZE;
2771 theader.otherCSize = AFS_OTHERCSIZE;
2772 theader.version = AFS_CI_VERSION;
2773 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2775 MReleaseWriteLock(&afs_xdcache);
2782 * Return a pointer to an freshly initialized dcache entry using
2783 * a memory-based cache. The tlock will be read-locked.
2786 * aslot : Dcache slot to look at.
2787 * tmpdc : Ptr to dcache entry.
2790 * Must be called with afs_xdcache write-locked.
2794 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2796 register struct dcache *tdc;
2799 AFS_STATCNT(afs_MemGetDSlot);
2800 if (CheckLock(&afs_xdcache) != -1)
2801 osi_Panic("getdslot nolock");
2802 if (aslot < 0 || aslot >= afs_cacheFiles)
2803 osi_Panic("getdslot slot");
2804 tdc = afs_indexTable[aslot];
2806 QRemove(&tdc->lruq); /* move to queue head */
2807 QAdd(&afs_DLRU, &tdc->lruq);
2808 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2809 ObtainWriteLock(&tdc->tlock, 624);
2811 ConvertWToRLock(&tdc->tlock);
2814 if (tmpdc == NULL) {
2815 if (!afs_freeDSList)
2816 afs_GetDownDSlot(4);
2817 if (!afs_freeDSList) {
2818 /* none free, making one is better than a panic */
2819 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2820 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2821 #ifdef KERNEL_HAVE_PIN
2822 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2825 tdc = afs_freeDSList;
2826 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2829 tdc->dflags = 0; /* up-to-date, not in free q */
2831 QAdd(&afs_DLRU, &tdc->lruq);
2832 if (tdc->lruq.prev == &tdc->lruq)
2833 osi_Panic("lruq 3");
2839 /* initialize entry */
2840 tdc->f.fid.Cell = 0;
2841 tdc->f.fid.Fid.Volume = 0;
2843 hones(tdc->f.versionNo);
2844 tdc->f.inode = aslot;
2845 tdc->dflags |= DFEntryMod;
2848 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2851 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2852 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2853 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2856 RWLOCK_INIT(&tdc->lock, "dcache lock");
2857 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2858 RWLOCK_INIT(&tdc->mflock, "dcache flock");
2859 ObtainReadLock(&tdc->tlock);
2862 afs_indexTable[aslot] = tdc;
2865 } /*afs_MemGetDSlot */
2867 unsigned int last_error = 0, lasterrtime = 0;
2873 * Return a pointer to an freshly initialized dcache entry using
2874 * a UFS-based disk cache. The dcache tlock will be read-locked.
2877 * aslot : Dcache slot to look at.
2878 * tmpdc : Ptr to dcache entry.
2881 * afs_xdcache lock write-locked.
2884 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2886 register afs_int32 code;
2887 register struct dcache *tdc;
2891 AFS_STATCNT(afs_UFSGetDSlot);
2892 if (CheckLock(&afs_xdcache) != -1)
2893 osi_Panic("getdslot nolock");
2894 if (aslot < 0 || aslot >= afs_cacheFiles)
2895 osi_Panic("getdslot slot");
2896 tdc = afs_indexTable[aslot];
2898 QRemove(&tdc->lruq); /* move to queue head */
2899 QAdd(&afs_DLRU, &tdc->lruq);
2900 /* Grab tlock in case refCount != 0 */
2901 ObtainWriteLock(&tdc->tlock, 625);
2903 ConvertWToRLock(&tdc->tlock);
2906 /* otherwise we should read it in from the cache file */
2908 * If we weren't passed an in-memory region to place the file info,
2909 * we have to allocate one.
2911 if (tmpdc == NULL) {
2912 if (!afs_freeDSList)
2913 afs_GetDownDSlot(4);
2914 if (!afs_freeDSList) {
2915 /* none free, making one is better than a panic */
2916 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2917 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2918 #ifdef KERNEL_HAVE_PIN
2919 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2922 tdc = afs_freeDSList;
2923 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2926 tdc->dflags = 0; /* up-to-date, not in free q */
2928 QAdd(&afs_DLRU, &tdc->lruq);
2929 if (tdc->lruq.prev == &tdc->lruq)
2930 osi_Panic("lruq 3");
2937 * Seek to the aslot'th entry and read it in.
2940 afs_osi_Read(afs_cacheInodep,
2941 sizeof(struct fcache) * aslot +
2942 sizeof(struct afs_fheader), (char *)(&tdc->f),
2943 sizeof(struct fcache));
2945 if (code != sizeof(struct fcache))
2947 if (!afs_CellNumValid(tdc->f.fid.Cell))
2951 tdc->f.fid.Cell = 0;
2952 tdc->f.fid.Fid.Volume = 0;
2954 hones(tdc->f.versionNo);
2955 tdc->dflags |= DFEntryMod;
2956 #if defined(KERNEL_HAVE_UERROR)
2957 last_error = getuerror();
2959 lasterrtime = osi_Time();
2960 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2961 tdc->f.states &= ~(DRO|DBackup|DRW);
2962 afs_DCMoveBucket(tdc, 0, 0);
2965 if (tdc->f.states & DRO) {
2966 afs_DCMoveBucket(tdc, 0, 2);
2967 } else if (tdc->f.states & DBackup) {
2968 afs_DCMoveBucket(tdc, 0, 1);
2970 afs_DCMoveBucket(tdc, 0, 1);
2976 if (tdc->f.chunk >= 0)
2977 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2982 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2983 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2984 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2987 RWLOCK_INIT(&tdc->lock, "dcache lock");
2988 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2989 RWLOCK_INIT(&tdc->mflock, "dcache flock");
2990 ObtainReadLock(&tdc->tlock);
2993 * If we didn't read into a temporary dcache region, update the
2994 * slot pointer table.
2997 afs_indexTable[aslot] = tdc;
3000 } /*afs_UFSGetDSlot */
3008 * write a particular dcache entry back to its home in the
3012 * adc : Pointer to the dcache entry to write.
3013 * atime : If true, set the modtime on the file to the current time.
3016 * Must be called with the afs_xdcache lock at least read-locked,
3017 * and dcache entry at least read-locked.
3018 * The reference count is not changed.
3022 afs_WriteDCache(register struct dcache *adc, int atime)
3024 register afs_int32 code;
3026 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
3028 AFS_STATCNT(afs_WriteDCache);
3030 adc->f.modTime = osi_Time();
3032 * Seek to the right dcache slot and write the in-memory image out to disk.
3034 afs_cellname_write();
3036 afs_osi_Write(afs_cacheInodep,
3037 sizeof(struct fcache) * adc->index +
3038 sizeof(struct afs_fheader), (char *)(&adc->f),
3039 sizeof(struct fcache));
3040 if (code != sizeof(struct fcache))
3051 * Wake up users of a particular file waiting for stores to take
3055 * avc : Ptr to related vcache entry.
3058 * Nothing interesting.
3062 afs_wakeup(register struct vcache *avc)
3065 register struct brequest *tb;
3067 AFS_STATCNT(afs_wakeup);
3068 for (i = 0; i < NBRS; i++, tb++) {
3069 /* if request is valid and for this file, we've found it */
3070 if (tb->refCount > 0 && avc == tb->vc) {
3073 * If CSafeStore is on, then we don't awaken the guy
3074 * waiting for the store until the whole store has finished.
3075 * Otherwise, we do it now. Note that if CSafeStore is on,
3076 * the BStore routine actually wakes up the user, instead
3078 * I think this is redundant now because this sort of thing
3079 * is already being handled by the higher-level code.
3081 if ((avc->states & CSafeStore) == 0) {
3083 tb->flags |= BUVALID;
3084 if (tb->flags & BUWAIT) {
3085 tb->flags &= ~BUWAIT;
3100 * Given a file name and inode, set up that file to be an
3101 * active member in the AFS cache. This also involves checking
3102 * the usability of its data.
3105 * afile : Name of the cache file to initialize.
3106 * ainode : Inode of the file.
3109 * This function is called only during initialization.
3113 afs_InitCacheFile(char *afile, ino_t ainode)
3115 register afs_int32 code;
3116 #if defined(AFS_LINUX22_ENV)
3117 struct dentry *filevp;
3119 struct vnode *filevp;
3123 struct osi_file *tfile;
3124 struct osi_stat tstat;
3125 register struct dcache *tdc;
3127 AFS_STATCNT(afs_InitCacheFile);
3128 index = afs_stats_cmperf.cacheNumEntries;
3129 if (index >= afs_cacheFiles)
3132 MObtainWriteLock(&afs_xdcache, 282);
3133 tdc = afs_GetDSlot(index, NULL);
3134 ReleaseReadLock(&tdc->tlock);
3135 MReleaseWriteLock(&afs_xdcache);
3137 ObtainWriteLock(&tdc->lock, 621);
3138 MObtainWriteLock(&afs_xdcache, 622);
3140 code = gop_lookupname(afile, AFS_UIOSYS, 0, &filevp);
3142 ReleaseWriteLock(&afs_xdcache);
3143 ReleaseWriteLock(&tdc->lock);
3148 * We have a VN_HOLD on filevp. Get the useful info out and
3149 * return. We make use of the fact that the cache is in the
3150 * UFS file system, and just record the inode number.
3152 #ifdef AFS_LINUX22_ENV
3153 tdc->f.inode = VTOI(filevp->d_inode)->i_number;
3156 tdc->f.inode = afs_vnodeToInumber(filevp);
3158 #endif /* AFS_LINUX22_ENV */
3160 tdc->f.inode = ainode;
3163 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3165 tfile = osi_UFSOpen(tdc->f.inode);
3166 code = afs_osi_Stat(tfile, &tstat);
3168 osi_Panic("initcachefile stat");
3171 * If file size doesn't match the cache info file, it's probably bad.
3173 if (tdc->f.chunkBytes != tstat.size)
3175 tdc->f.chunkBytes = 0;
3178 * If file changed within T (120?) seconds of cache info file, it's
3179 * probably bad. In addition, if slot changed within last T seconds,
3180 * the cache info file may be incorrectly identified, and so slot
3183 if (cacheInfoModTime < tstat.mtime + 120)
3185 if (cacheInfoModTime < tdc->f.modTime + 120)
3187 /* In case write through is behind, make sure cache items entry is
3188 * at least as new as the chunk.
3190 if (tdc->f.modTime < tstat.mtime)
3193 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3194 if (tstat.size != 0)
3195 osi_UFSTruncate(tfile, 0);
3196 tdc->f.states &= ~(DRO|DBackup|DRW);
3197 afs_DCMoveBucket(tdc, 0, 0);
3198 /* put entry in free cache slot list */
3199 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3200 afs_freeDCList = index;
3202 afs_indexFlags[index] |= IFFree;
3203 afs_indexUnique[index] = 0;
3206 * We must put this entry in the appropriate hash tables.
3207 * Note that i is still set from the above DCHash call
3209 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3210 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3211 afs_dchashTbl[code] = tdc->index;
3212 code = DVHash(&tdc->f.fid);
3213 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3214 afs_dvhashTbl[code] = tdc->index;
3215 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3217 /* has nontrivial amt of data */
3218 afs_indexFlags[index] |= IFEverUsed;
3219 afs_stats_cmperf.cacheFilesReused++;
3221 * Initialize index times to file's mod times; init indexCounter
3224 hset32(afs_indexTimes[index], tstat.atime);
3225 if (hgetlo(afs_indexCounter) < tstat.atime) {
3226 hset32(afs_indexCounter, tstat.atime);
3228 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3229 } /*File is not bad */
3231 osi_UFSClose(tfile);
3232 tdc->f.states &= ~DWriting;
3233 tdc->dflags &= ~DFEntryMod;
3234 /* don't set f.modTime; we're just cleaning up */
3235 afs_WriteDCache(tdc, 0);
3236 ReleaseWriteLock(&afs_xdcache);
3237 ReleaseWriteLock(&tdc->lock);
3239 afs_stats_cmperf.cacheNumEntries++;
3244 /*Max # of struct dcache's resident at any time*/
3246 * If 'dchint' is enabled then in-memory dcache min is increased because of
3255 * Initialize dcache related variables.
3258 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3260 register struct dcache *tdp;
3264 afs_freeDCList = NULLIDX;
3265 afs_discardDCList = NULLIDX;
3266 afs_freeDCCount = 0;
3267 afs_freeDSList = NULL;
3268 hzero(afs_indexCounter);
3270 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3276 if (achunk < 0 || achunk > 30)
3277 achunk = 13; /* Use default */
3278 AFS_SETCHUNKSIZE(achunk);
3284 if (aflags & AFSCALL_INIT_MEMCACHE) {
3286 * Use a memory cache instead of a disk cache
3288 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3289 afs_cacheType = &afs_MemCacheOps;
3290 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3291 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3292 /* ablocks is reported in 1K blocks */
3293 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3295 printf("afsd: memory cache too large for available memory.\n");
3296 printf("afsd: AFS files cannot be accessed.\n\n");
3298 afiles = ablocks = 0;
3300 printf("Memory cache: Allocating %d dcache entries...",
3303 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3304 afs_cacheType = &afs_UfsCacheOps;
3307 if (aDentries > 512)
3308 afs_dhashsize = 2048;
3309 /* initialize hash tables */
3311 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3313 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3314 for (i = 0; i < afs_dhashsize; i++) {
3315 afs_dvhashTbl[i] = NULLIDX;
3316 afs_dchashTbl[i] = NULLIDX;
3318 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3319 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3320 for (i = 0; i < afiles; i++) {
3321 afs_dvnextTbl[i] = NULLIDX;
3322 afs_dcnextTbl[i] = NULLIDX;
3325 /* Allocate and zero the pointer array to the dcache entries */
3326 afs_indexTable = (struct dcache **)
3327 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3328 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3330 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3331 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3333 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3334 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3335 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3336 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3338 /* Allocate and thread the struct dcache entries themselves */
3339 tdp = afs_Initial_freeDSList =
3340 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3341 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3342 #ifdef KERNEL_HAVE_PIN
3343 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3344 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3345 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3346 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3347 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3348 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3349 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3350 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3351 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3354 afs_freeDSList = &tdp[0];
3355 for (i = 0; i < aDentries - 1; i++) {
3356 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3357 RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3358 RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3359 RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3361 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3362 RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3363 RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3364 RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3366 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3367 afs_cacheBlocks = ablocks;
3368 afs_ComputeCacheParms(); /* compute parms based on cache size */
3370 afs_dcentries = aDentries;
3372 afs_stats_cmperf.cacheBucket0_Discarded =
3373 afs_stats_cmperf.cacheBucket1_Discarded =
3374 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3384 shutdown_dcache(void)
3388 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3389 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3390 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3391 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3392 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3393 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3394 afs_osi_Free(afs_Initial_freeDSList,
3395 afs_dcentries * sizeof(struct dcache));
3396 #ifdef KERNEL_HAVE_PIN
3397 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3398 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3399 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3400 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3401 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3402 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3403 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3407 for (i = 0; i < afs_dhashsize; i++) {
3408 afs_dvhashTbl[i] = NULLIDX;
3409 afs_dchashTbl[i] = NULLIDX;
3412 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3413 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3415 afs_blocksUsed = afs_dcentries = 0;
3416 afs_stats_cmperf.cacheBucket0_Discarded =
3417 afs_stats_cmperf.cacheBucket1_Discarded =
3418 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3419 hzero(afs_indexCounter);
3421 afs_freeDCCount = 0;
3422 afs_freeDCList = NULLIDX;
3423 afs_discardDCList = NULLIDX;
3424 afs_freeDSList = afs_Initial_freeDSList = 0;
3426 LOCK_INIT(&afs_xdcache, "afs_xdcache");