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 / (afs_cacheBlocks / 100);
231 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
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 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
351 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
353 osi_GetuTime(&CTD_stats.CTD_afterSleep);
354 afs_TruncateDaemonRunning = 1;
356 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
357 MObtainWriteLock(&afs_xdcache, 266);
358 if (afs_CacheTooFull) {
359 int space_needed, slots_needed;
360 /* if we get woken up, we should try to clean something out */
361 for (counter = 0; counter < 10; counter++) {
363 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
365 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
366 afs_GetDownD(slots_needed, &space_needed, 0);
367 if ((space_needed <= 0) && (slots_needed <= 0)) {
370 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
373 if (!afs_CacheIsTooFull())
374 afs_CacheTooFull = 0;
376 MReleaseWriteLock(&afs_xdcache);
379 * This is a defensive check to try to avoid starving threads
380 * that may need the global lock so thay can help free some
381 * cache space. If this thread won't be sleeping or truncating
382 * any cache files then give up the global lock so other
383 * threads get a chance to run.
385 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
386 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
387 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
391 * This is where we free the discarded cache elements.
393 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
394 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
395 afs_FreeDiscardedDCache();
398 /* See if we need to continue to run. Someone may have
399 * signalled us while we were executing.
401 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
402 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
403 /* Collect statistics on truncate daemon. */
404 CTD_stats.CTD_nSleeps++;
405 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
406 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
407 CTD_stats.CTD_beforeSleep);
408 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
410 afs_TruncateDaemonRunning = 0;
411 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
412 afs_TruncateDaemonRunning = 1;
414 osi_GetuTime(&CTD_stats.CTD_afterSleep);
415 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
416 CTD_stats.CTD_afterSleep);
417 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
419 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
421 afs_termState = AFSOP_STOP_AFSDB;
423 afs_termState = AFSOP_STOP_RXEVENT;
425 afs_osi_Wakeup(&afs_termState);
436 * Make adjustment for the new size in the disk cache entry
438 * Major Assumptions Here:
439 * Assumes that frag size is an integral power of two, less one,
440 * and that this is a two's complement machine. I don't
441 * know of any filesystems which violate this assumption...
444 * adc : Ptr to dcache entry.
445 * anewsize : New size desired.
449 afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
451 register afs_int32 oldSize;
453 AFS_STATCNT(afs_AdjustSize);
455 adc->dflags |= DFEntryMod;
456 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
457 adc->f.chunkBytes = newSize;
460 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
461 afs_DCAdjustSize(adc, oldSize, newSize);
462 if (newSize > oldSize) {
463 /* We're growing the file, wakeup the daemon */
464 afs_MaybeWakeupTruncateDaemon();
466 afs_blocksUsed += (newSize - oldSize);
467 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
475 * This routine is responsible for moving at least one entry (but up
476 * to some number of them) from the LRU queue to the free queue.
479 * anumber : Number of entries that should ideally be moved.
480 * aneedSpace : How much space we need (1K blocks);
483 * The anumber parameter is just a hint; at least one entry MUST be
484 * moved, or we'll panic. We must be called with afs_xdcache
485 * write-locked. We should try to satisfy both anumber and aneedspace,
486 * whichever is more demanding - need to do several things:
487 * 1. only grab up to anumber victims if aneedSpace <= 0, not
488 * the whole set of MAXATONCE.
489 * 2. dynamically choose MAXATONCE to reflect severity of
490 * demand: something like (*aneedSpace >> (logChunk - 9))
491 * N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
492 * indicates that the cache is not properly configured/tuned or
493 * something. We should be able to automatically correct that problem.
496 #define MAXATONCE 16 /* max we can obtain at once */
498 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
502 struct VenusFid *afid;
506 register struct vcache *tvc;
507 afs_uint32 victims[MAXATONCE];
508 struct dcache *victimDCs[MAXATONCE];
509 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
510 afs_uint32 victimPtr; /* next free item in victim arrays */
511 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
512 afs_uint32 maxVictimPtr; /* where it is */
516 AFS_STATCNT(afs_GetDownD);
517 if (CheckLock(&afs_xdcache) != -1)
518 osi_Panic("getdownd nolock");
519 /* decrement anumber first for all dudes in free list */
520 /* SHOULD always decrement anumber first, even if aneedSpace >0,
521 * because we should try to free space even if anumber <=0 */
522 if (!aneedSpace || *aneedSpace <= 0) {
523 anumber -= afs_freeDCCount;
525 return; /* enough already free */
527 /* bounds check parameter */
528 if (anumber > MAXATONCE)
529 anumber = MAXATONCE; /* all we can do */
531 /* rewrite so phases include a better eligiblity for gc test*/
533 * The phase variable manages reclaims. Set to 0, the first pass,
534 * we don't reclaim active entries, or other than target bucket.
535 * Set to 1, we reclaim even active ones in target bucket.
536 * Set to 2, we reclaim any inactive one.
537 * Set to 3, we reclaim even active ones.
545 for (i = 0; i < afs_cacheFiles; i++)
546 /* turn off all flags */
547 afs_indexFlags[i] &= ~IFFlag;
549 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
550 /* find oldest entries for reclamation */
551 maxVictimPtr = victimPtr = 0;
552 hzero(maxVictimTime);
553 curbucket = afs_DCWhichBucket(phase, buckethint);
554 /* select victims from access time array */
555 for (i = 0; i < afs_cacheFiles; i++) {
556 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
557 /* skip if dirty or already free */
560 tdc = afs_indexTable[i];
561 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
563 /* Wrong bucket; can't use it! */
566 if (tdc && (tdc->refCount != 0)) {
567 /* Referenced; can't use it! */
570 hset(vtime, afs_indexTimes[i]);
572 /* if we've already looked at this one, skip it */
573 if (afs_indexFlags[i] & IFFlag)
576 if (victimPtr < MAXATONCE) {
577 /* if there's at least one free victim slot left */
578 victims[victimPtr] = i;
579 hset(victimTimes[victimPtr], vtime);
580 if (hcmp(vtime, maxVictimTime) > 0) {
581 hset(maxVictimTime, vtime);
582 maxVictimPtr = victimPtr;
585 } else if (hcmp(vtime, maxVictimTime) < 0) {
587 * We're older than youngest victim, so we replace at
590 /* find youngest (largest LRU) victim */
593 osi_Panic("getdownd local");
595 hset(victimTimes[j], vtime);
596 /* recompute maxVictimTime */
597 hset(maxVictimTime, vtime);
598 for (j = 0; j < victimPtr; j++)
599 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
600 hset(maxVictimTime, victimTimes[j]);
606 /* now really reclaim the victims */
607 j = 0; /* flag to track if we actually got any of the victims */
608 /* first, hold all the victims, since we're going to release the lock
609 * during the truncate operation.
611 for (i = 0; i < victimPtr; i++) {
612 tdc = afs_GetDSlot(victims[i], 0);
613 /* We got tdc->tlock(R) here */
614 if (tdc->refCount == 1)
618 ReleaseReadLock(&tdc->tlock);
622 for (i = 0; i < victimPtr; i++) {
623 /* q is first elt in dcache entry */
625 /* now, since we're dropping the afs_xdcache lock below, we
626 * have to verify, before proceeding, that there are no other
627 * references to this dcache entry, even now. Note that we
628 * compare with 1, since we bumped it above when we called
629 * afs_GetDSlot to preserve the entry's identity.
631 if (tdc && tdc->refCount == 1) {
632 unsigned char chunkFlags;
633 afs_size_t tchunkoffset = 0;
635 /* xdcache is lower than the xvcache lock */
636 MReleaseWriteLock(&afs_xdcache);
637 MObtainReadLock(&afs_xvcache);
638 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
639 MReleaseReadLock(&afs_xvcache);
640 MObtainWriteLock(&afs_xdcache, 527);
642 if (tdc->refCount > 1)
645 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
646 chunkFlags = afs_indexFlags[tdc->index];
647 if (((phase & 1) == 0) && osi_Active(tvc))
649 if (((phase & 1) == 1) && osi_Active(tvc)
650 && (tvc->states & CDCLock)
651 && (chunkFlags & IFAnyPages))
653 if (chunkFlags & IFDataMod)
655 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
656 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
657 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
658 ICL_HANDLE_OFFSET(tchunkoffset));
660 #if defined(AFS_SUN5_ENV)
662 * Now we try to invalidate pages. We do this only for
663 * Solaris. For other platforms, it's OK to recycle a
664 * dcache entry out from under a page, because the strategy
665 * function can call afs_GetDCache().
667 if (!skip && (chunkFlags & IFAnyPages)) {
670 MReleaseWriteLock(&afs_xdcache);
671 MObtainWriteLock(&tvc->vlock, 543);
672 if (tvc->multiPage) {
676 /* block locking pages */
677 tvc->vstates |= VPageCleaning;
678 /* block getting new pages */
680 MReleaseWriteLock(&tvc->vlock);
681 /* One last recheck */
682 MObtainWriteLock(&afs_xdcache, 333);
683 chunkFlags = afs_indexFlags[tdc->index];
684 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
685 || (osi_Active(tvc) && (tvc->states & CDCLock)
686 && (chunkFlags & IFAnyPages))) {
688 MReleaseWriteLock(&afs_xdcache);
691 MReleaseWriteLock(&afs_xdcache);
693 code = osi_VM_GetDownD(tvc, tdc);
695 MObtainWriteLock(&afs_xdcache, 269);
696 /* we actually removed all pages, clean and dirty */
698 afs_indexFlags[tdc->index] &=
699 ~(IFDirtyPages | IFAnyPages);
702 MReleaseWriteLock(&afs_xdcache);
704 MObtainWriteLock(&tvc->vlock, 544);
705 if (--tvc->activeV == 0
706 && (tvc->vstates & VRevokeWait)) {
707 tvc->vstates &= ~VRevokeWait;
708 afs_osi_Wakeup((char *)&tvc->vstates);
711 if (tvc->vstates & VPageCleaning) {
712 tvc->vstates &= ~VPageCleaning;
713 afs_osi_Wakeup((char *)&tvc->vstates);
716 MReleaseWriteLock(&tvc->vlock);
718 #endif /* AFS_SUN5_ENV */
720 MReleaseWriteLock(&afs_xdcache);
723 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
724 MObtainWriteLock(&afs_xdcache, 528);
725 if (afs_indexFlags[tdc->index] &
726 (IFDataMod | IFDirtyPages | IFAnyPages))
728 if (tdc->refCount > 1)
731 #if defined(AFS_SUN5_ENV)
733 /* no vnode, so IFDirtyPages is spurious (we don't
734 * sweep dcaches on vnode recycling, so we can have
735 * DIRTYPAGES set even when all pages are gone). Just
737 * Hold vcache lock to prevent vnode from being
738 * created while we're clearing IFDirtyPages.
740 afs_indexFlags[tdc->index] &=
741 ~(IFDirtyPages | IFAnyPages);
745 /* skip this guy and mark him as recently used */
746 afs_indexFlags[tdc->index] |= IFFlag;
747 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
748 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
749 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
750 ICL_HANDLE_OFFSET(tchunkoffset));
752 /* flush this dude from the data cache and reclaim;
753 * first, make sure no one will care that we damage
754 * it, by removing it from all hash tables. Then,
755 * melt it down for parts. Note that any concurrent
756 * (new possibility!) calls to GetDownD won't touch
757 * this guy because his reference count is > 0. */
758 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
759 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
760 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
761 ICL_HANDLE_OFFSET(tchunkoffset));
762 AFS_STATCNT(afs_gget);
763 afs_HashOutDCache(tdc);
764 if (tdc->f.chunkBytes != 0) {
768 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
773 afs_DiscardDCache(tdc);
778 j = 1; /* we reclaimed at least one victim */
785 /* Phase is 0 and no one was found, so try phase 1 (ignore
786 * osi_Active flag) */
789 for (i = 0; i < afs_cacheFiles; i++)
790 /* turn off all flags */
791 afs_indexFlags[i] &= ~IFFlag;
794 /* found no one in phases 0-5, we're hosed */
798 } /* big while loop */
805 * Description: remove adc from any hash tables that would allow it to be located
806 * again by afs_FindDCache or afs_GetDCache.
808 * Parameters: adc -- pointer to dcache entry to remove from hash tables.
810 * Locks: Must have the afs_xdcache lock write-locked to call this function.
813 afs_HashOutDCache(struct dcache *adc)
817 AFS_STATCNT(afs_glink);
818 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
820 /* if this guy is in the hash table, pull him out */
821 if (adc->f.fid.Fid.Volume != 0) {
822 /* remove entry from first hash chains */
823 i = DCHash(&adc->f.fid, adc->f.chunk);
824 us = afs_dchashTbl[i];
825 if (us == adc->index) {
826 /* first dude in the list */
827 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
829 /* somewhere on the chain */
830 while (us != NULLIDX) {
831 if (afs_dcnextTbl[us] == adc->index) {
832 /* found item pointing at the one to delete */
833 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
836 us = afs_dcnextTbl[us];
839 osi_Panic("dcache hc");
841 /* remove entry from *other* hash chain */
842 i = DVHash(&adc->f.fid);
843 us = afs_dvhashTbl[i];
844 if (us == adc->index) {
845 /* first dude in the list */
846 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
848 /* somewhere on the chain */
849 while (us != NULLIDX) {
850 if (afs_dvnextTbl[us] == adc->index) {
851 /* found item pointing at the one to delete */
852 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
855 us = afs_dvnextTbl[us];
858 osi_Panic("dcache hv");
862 /* prevent entry from being found on a reboot (it is already out of
863 * the hash table, but after a crash, we just look at fid fields of
864 * stable (old) entries).
866 adc->f.fid.Fid.Volume = 0; /* invalid */
868 /* mark entry as modified */
869 adc->dflags |= DFEntryMod;
873 } /*afs_HashOutDCache */
880 * Flush the given dcache entry, pulling it from hash chains
881 * and truncating the associated cache file.
884 * adc: Ptr to dcache entry to flush.
887 * This routine must be called with the afs_xdcache lock held
892 afs_FlushDCache(register struct dcache *adc)
894 AFS_STATCNT(afs_FlushDCache);
896 * Bump the number of cache files flushed.
898 afs_stats_cmperf.cacheFlushes++;
900 /* remove from all hash tables */
901 afs_HashOutDCache(adc);
903 /* Free its space; special case null operation, since truncate operation
904 * in UFS is slow even in this case, and this allows us to pre-truncate
905 * these files at more convenient times with fewer locks set
906 * (see afs_GetDownD).
908 if (adc->f.chunkBytes != 0) {
909 afs_DiscardDCache(adc);
910 afs_MaybeWakeupTruncateDaemon();
915 if (afs_WaitForCacheDrain) {
916 if (afs_blocksUsed <=
917 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
918 afs_WaitForCacheDrain = 0;
919 afs_osi_Wakeup(&afs_WaitForCacheDrain);
922 } /*afs_FlushDCache */
928 * Description: put a dcache entry on the free dcache entry list.
930 * Parameters: adc -- dcache entry to free
932 * Environment: called with afs_xdcache lock write-locked.
935 afs_FreeDCache(register struct dcache *adc)
937 /* Thread on free list, update free list count and mark entry as
938 * freed in its indexFlags element. Also, ensure DCache entry gets
939 * written out (set DFEntryMod).
942 afs_dvnextTbl[adc->index] = afs_freeDCList;
943 afs_freeDCList = adc->index;
945 afs_indexFlags[adc->index] |= IFFree;
946 adc->dflags |= DFEntryMod;
948 if (afs_WaitForCacheDrain) {
949 if ((afs_blocksUsed - afs_blocksDiscarded) <=
950 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
951 afs_WaitForCacheDrain = 0;
952 afs_osi_Wakeup(&afs_WaitForCacheDrain);
961 * Discard the cache element by moving it to the discardDCList.
962 * This puts the cache element into a quasi-freed state, where
963 * the space may be reused, but the file has not been truncated.
965 * Major Assumptions Here:
966 * Assumes that frag size is an integral power of two, less one,
967 * and that this is a two's complement machine. I don't
968 * know of any filesystems which violate this assumption...
971 * adc : Ptr to dcache entry.
974 * Must be called with afs_xdcache write-locked.
978 afs_DiscardDCache(register struct dcache *adc)
980 register afs_int32 size;
982 AFS_STATCNT(afs_DiscardDCache);
984 osi_Assert(adc->refCount == 1);
986 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
987 afs_blocksDiscarded += size;
988 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
990 afs_dvnextTbl[adc->index] = afs_discardDCList;
991 afs_discardDCList = adc->index;
992 afs_discardDCCount++;
994 adc->f.fid.Fid.Volume = 0;
995 adc->dflags |= DFEntryMod;
996 afs_indexFlags[adc->index] |= IFDiscarded;
998 if (afs_WaitForCacheDrain) {
999 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1000 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1001 afs_WaitForCacheDrain = 0;
1002 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1006 } /*afs_DiscardDCache */
1009 * afs_FreeDiscardedDCache
1012 * Free the next element on the list of discarded cache elements.
1015 afs_FreeDiscardedDCache(void)
1017 register struct dcache *tdc;
1018 register struct osi_file *tfile;
1019 register afs_int32 size;
1021 AFS_STATCNT(afs_FreeDiscardedDCache);
1023 MObtainWriteLock(&afs_xdcache, 510);
1024 if (!afs_blocksDiscarded) {
1025 MReleaseWriteLock(&afs_xdcache);
1030 * Get an entry from the list of discarded cache elements
1032 tdc = afs_GetDSlot(afs_discardDCList, 0);
1033 osi_Assert(tdc->refCount == 1);
1034 ReleaseReadLock(&tdc->tlock);
1036 afs_discardDCList = afs_dvnextTbl[tdc->index];
1037 afs_dvnextTbl[tdc->index] = NULLIDX;
1038 afs_discardDCCount--;
1039 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1040 afs_blocksDiscarded -= size;
1041 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1042 /* We can lock because we just took it off the free list */
1043 ObtainWriteLock(&tdc->lock, 626);
1044 MReleaseWriteLock(&afs_xdcache);
1047 * Truncate the element to reclaim its space
1049 tfile = afs_CFileOpen(tdc->f.inode);
1050 afs_CFileTruncate(tfile, 0);
1051 afs_CFileClose(tfile);
1052 afs_AdjustSize(tdc, 0);
1053 tdc->f.states &= ~(DRO|DBackup|DRW);
1054 afs_DCMoveBucket(tdc, 0, 0);
1057 * Free the element we just truncated
1059 MObtainWriteLock(&afs_xdcache, 511);
1060 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1061 afs_FreeDCache(tdc);
1062 ReleaseWriteLock(&tdc->lock);
1064 MReleaseWriteLock(&afs_xdcache);
1068 * afs_MaybeFreeDiscardedDCache
1071 * Free as many entries from the list of discarded cache elements
1072 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1078 afs_MaybeFreeDiscardedDCache(void)
1081 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1083 while (afs_blocksDiscarded
1084 && (afs_blocksUsed >
1085 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1086 afs_FreeDiscardedDCache();
1095 * Try to free up a certain number of disk slots.
1098 * anumber : Targeted number of disk slots to free up.
1101 * Must be called with afs_xdcache write-locked.
1104 afs_GetDownDSlot(int anumber)
1106 struct afs_q *tq, *nq;
1111 AFS_STATCNT(afs_GetDownDSlot);
1112 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1113 osi_Panic("diskless getdowndslot");
1115 if (CheckLock(&afs_xdcache) != -1)
1116 osi_Panic("getdowndslot nolock");
1118 /* decrement anumber first for all dudes in free list */
1119 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1122 return; /* enough already free */
1124 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1126 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1127 nq = QPrev(tq); /* in case we remove it */
1128 if (tdc->refCount == 0) {
1129 if ((ix = tdc->index) == NULLIDX)
1130 osi_Panic("getdowndslot");
1131 /* pull the entry out of the lruq and put it on the free list */
1132 QRemove(&tdc->lruq);
1134 /* write-through if modified */
1135 if (tdc->dflags & DFEntryMod) {
1136 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1138 * ask proxy to do this for us - we don't have the stack space
1140 while (tdc->dflags & DFEntryMod) {
1143 s = SPLOCK(afs_sgibklock);
1144 if (afs_sgibklist == NULL) {
1145 /* if slot is free, grab it. */
1146 afs_sgibklist = tdc;
1147 SV_SIGNAL(&afs_sgibksync);
1149 /* wait for daemon to (start, then) finish. */
1150 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1154 tdc->dflags &= ~DFEntryMod;
1155 afs_WriteDCache(tdc, 1);
1159 /* finally put the entry in the free list */
1160 afs_indexTable[ix] = NULL;
1161 afs_indexFlags[ix] &= ~IFEverUsed;
1162 tdc->index = NULLIDX;
1163 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1164 afs_freeDSList = tdc;
1168 } /*afs_GetDownDSlot */
1175 * Increment the reference count on a disk cache entry,
1176 * which already has a non-zero refcount. In order to
1177 * increment the refcount of a zero-reference entry, you
1178 * have to hold afs_xdcache.
1181 * adc : Pointer to the dcache entry to increment.
1184 * Nothing interesting.
1187 afs_RefDCache(struct dcache *adc)
1189 ObtainWriteLock(&adc->tlock, 627);
1190 if (adc->refCount < 0)
1191 osi_Panic("RefDCache: negative refcount");
1193 ReleaseWriteLock(&adc->tlock);
1202 * Decrement the reference count on a disk cache entry.
1205 * ad : Ptr to the dcache entry to decrement.
1208 * Nothing interesting.
1211 afs_PutDCache(register struct dcache *adc)
1213 AFS_STATCNT(afs_PutDCache);
1214 ObtainWriteLock(&adc->tlock, 276);
1215 if (adc->refCount <= 0)
1216 osi_Panic("putdcache");
1218 ReleaseWriteLock(&adc->tlock);
1227 * Try to discard all data associated with this file from the
1231 * avc : Pointer to the cache info for the file.
1234 * Both pvnLock and lock are write held.
1237 afs_TryToSmush(register struct vcache *avc, struct AFS_UCRED *acred, int sync)
1239 register struct dcache *tdc;
1242 AFS_STATCNT(afs_TryToSmush);
1243 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1244 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length));
1245 sync = 1; /* XX Temp testing XX */
1247 #if defined(AFS_SUN5_ENV)
1248 ObtainWriteLock(&avc->vlock, 573);
1249 avc->activeV++; /* block new getpages */
1250 ReleaseWriteLock(&avc->vlock);
1253 /* Flush VM pages */
1254 osi_VM_TryToSmush(avc, acred, sync);
1257 * Get the hash chain containing all dce's for this fid
1259 i = DVHash(&avc->fid);
1260 MObtainWriteLock(&afs_xdcache, 277);
1261 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1262 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1263 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1264 int releaseTlock = 1;
1265 tdc = afs_GetDSlot(index, NULL);
1266 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1268 if ((afs_indexFlags[index] & IFDataMod) == 0
1269 && tdc->refCount == 1) {
1270 ReleaseReadLock(&tdc->tlock);
1272 afs_FlushDCache(tdc);
1275 afs_indexTable[index] = 0;
1278 ReleaseReadLock(&tdc->tlock);
1282 #if defined(AFS_SUN5_ENV)
1283 ObtainWriteLock(&avc->vlock, 545);
1284 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1285 avc->vstates &= ~VRevokeWait;
1286 afs_osi_Wakeup((char *)&avc->vstates);
1288 ReleaseWriteLock(&avc->vlock);
1290 MReleaseWriteLock(&afs_xdcache);
1292 * It's treated like a callback so that when we do lookups we'll invalidate the unique bit if any
1293 * trytoSmush occured during the lookup call
1302 * Given the cached info for a file and a byte offset into the
1303 * file, make sure the dcache entry for that file and containing
1304 * the given byte is available, returning it to our caller.
1307 * avc : Pointer to the (held) vcache entry to look in.
1308 * abyte : Which byte we want to get to.
1311 * Pointer to the dcache entry covering the file & desired byte,
1312 * or NULL if not found.
1315 * The vcache entry is held upon entry.
1319 afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1322 register afs_int32 i, index;
1323 register struct dcache *tdc = NULL;
1325 AFS_STATCNT(afs_FindDCache);
1326 chunk = AFS_CHUNK(abyte);
1329 * Hash on the [fid, chunk] and get the corresponding dcache index
1330 * after write-locking the dcache.
1332 i = DCHash(&avc->fid, chunk);
1333 MObtainWriteLock(&afs_xdcache, 278);
1334 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1335 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1336 tdc = afs_GetDSlot(index, NULL);
1337 ReleaseReadLock(&tdc->tlock);
1338 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1339 break; /* leaving refCount high for caller */
1344 index = afs_dcnextTbl[index];
1346 MReleaseWriteLock(&afs_xdcache);
1347 if (index != NULLIDX) {
1348 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1349 hadd32(afs_indexCounter, 1);
1354 } /*afs_FindDCache */
1358 * afs_UFSCacheStoreProc
1361 * Called upon store.
1364 * acall : Ptr to the Rx call structure involved.
1365 * afile : Ptr to the related file descriptor.
1366 * alen : Size of the file in bytes.
1367 * avc : Ptr to the vcache entry.
1368 * shouldWake : is it "safe" to return early from close() ?
1369 * abytesToXferP : Set to the number of bytes to xfer.
1370 * NOTE: This parameter is only used if AFS_NOSTATS
1372 * abytesXferredP : Set to the number of bytes actually xferred.
1373 * NOTE: This parameter is only used if AFS_NOSTATS
1377 * Nothing interesting.
1380 afs_UFSCacheStoreProc(register struct rx_call *acall, struct osi_file *afile,
1381 register afs_int32 alen, struct vcache *avc,
1382 int *shouldWake, afs_size_t * abytesToXferP,
1383 afs_size_t * abytesXferredP)
1385 afs_int32 code, got;
1386 register char *tbuffer;
1389 AFS_STATCNT(UFS_CacheStoreProc);
1393 * In this case, alen is *always* the amount of data we'll be trying
1396 (*abytesToXferP) = alen;
1397 (*abytesXferredP) = 0;
1398 #endif /* AFS_NOSTATS */
1400 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1401 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1402 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1403 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1405 tlen = (alen > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : alen);
1406 got = afs_osi_Read(afile, -1, tbuffer, tlen);
1408 #if defined(KERNEL_HAVE_UERROR)
1409 || (got != tlen && getuerror())
1412 osi_FreeLargeSpace(tbuffer);
1415 afs_Trace2(afs_iclSetp, CM_TRACE_STOREPROC2, ICL_TYPE_OFFSET,
1416 ICL_HANDLE_OFFSET(*tbuffer), ICL_TYPE_INT32, got);
1418 code = rx_Write(acall, tbuffer, got); /* writing 0 bytes will
1419 * push a short packet. Is that really what we want, just because the
1420 * data didn't come back from the disk yet? Let's try it and see. */
1423 (*abytesXferredP) += code;
1424 #endif /* AFS_NOSTATS */
1426 code = rx_Error(acall);
1427 osi_FreeLargeSpace(tbuffer);
1428 return code ? code : -33;
1432 * If file has been locked on server, we can allow the store
1435 if (shouldWake && *shouldWake && (rx_GetRemoteStatus(acall) & 1)) {
1436 *shouldWake = 0; /* only do this once */
1440 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1441 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1442 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1443 osi_FreeLargeSpace(tbuffer);
1446 } /* afs_UFSCacheStoreProc */
1450 * afs_UFSCacheFetchProc
1453 * Routine called on fetch; also tells people waiting for data
1454 * that more has arrived.
1457 * acall : Ptr to the Rx call structure.
1458 * afile : File descriptor for the cache file.
1459 * abase : Base offset to fetch.
1460 * adc : Ptr to the dcache entry for the file, write-locked.
1461 * avc : Ptr to the vcache entry for the file.
1462 * abytesToXferP : Set to the number of bytes to xfer.
1463 * NOTE: This parameter is only used if AFS_NOSTATS
1465 * abytesXferredP : Set to the number of bytes actually xferred.
1466 * NOTE: This parameter is only used if AFS_NOSTATS
1470 * Nothing interesting.
1474 afs_UFSCacheFetchProc(register struct rx_call *acall, struct osi_file *afile,
1475 afs_size_t abase, struct dcache *adc,
1476 struct vcache *avc, afs_size_t * abytesToXferP,
1477 afs_size_t * abytesXferredP, afs_int32 lengthFound)
1480 register afs_int32 code;
1481 register char *tbuffer;
1485 AFS_STATCNT(UFS_CacheFetchProc);
1486 osi_Assert(WriteLocked(&adc->lock));
1487 afile->offset = 0; /* Each time start from the beginning */
1488 length = lengthFound;
1490 (*abytesToXferP) = 0;
1491 (*abytesXferredP) = 0;
1492 #endif /* AFS_NOSTATS */
1493 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1494 adc->validPos = abase;
1498 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
1500 length = ntohl(length);
1501 if (code != sizeof(afs_int32)) {
1502 osi_FreeLargeSpace(tbuffer);
1503 code = rx_Error(acall);
1504 return (code ? code : -1); /* try to return code, not -1 */
1508 * The fetch protocol is extended for the AFS/DFS translator
1509 * to allow multiple blocks of data, each with its own length,
1510 * to be returned. As long as the top bit is set, there are more
1513 * We do not do this for AFS file servers because they sometimes
1514 * return large negative numbers as the transfer size.
1516 if (avc->states & CForeign) {
1517 moredata = length & 0x80000000;
1518 length &= ~0x80000000;
1523 (*abytesToXferP) += length;
1524 #endif /* AFS_NOSTATS */
1525 while (length > 0) {
1526 tlen = (length > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : length);
1527 #ifdef RX_KERNEL_TRACE
1528 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1532 code = rx_Read(acall, tbuffer, tlen);
1534 #ifdef RX_KERNEL_TRACE
1535 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1539 (*abytesXferredP) += code;
1540 #endif /* AFS_NOSTATS */
1542 osi_FreeLargeSpace(tbuffer);
1543 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64READ,
1544 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
1545 ICL_TYPE_INT32, length);
1548 code = afs_osi_Write(afile, -1, tbuffer, tlen);
1550 osi_FreeLargeSpace(tbuffer);
1555 adc->validPos = abase;
1556 if (afs_osi_Wakeup(&adc->validPos) == 0)
1557 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
1558 __FILE__, ICL_TYPE_INT32, __LINE__,
1559 ICL_TYPE_POINTER, adc, ICL_TYPE_INT32,
1563 osi_FreeLargeSpace(tbuffer);
1566 } /* afs_UFSCacheFetchProc */
1572 * This function is called to obtain a reference to data stored in
1573 * the disk cache, locating a chunk of data containing the desired
1574 * byte and returning a reference to the disk cache entry, with its
1575 * reference count incremented.
1579 * avc : Ptr to a vcache entry (unlocked)
1580 * abyte : Byte position in the file desired
1581 * areq : Request structure identifying the requesting user.
1582 * aflags : Settings as follows:
1584 * 2 : Return after creating entry.
1585 * 4 : called from afs_vnop_write.c
1586 * *alen contains length of data to be written.
1588 * aoffset : Set to the offset within the chunk where the resident
1590 * alen : Set to the number of bytes of data after the desired
1591 * byte (including the byte itself) which can be read
1595 * The vcache entry pointed to by avc is unlocked upon entry.
1599 struct AFSVolSync tsync;
1600 struct AFSFetchStatus OutStatus;
1601 struct AFSCallBack CallBack;
1605 * Update the vnode-to-dcache hint if we can get the vnode lock
1606 * right away. Assumes dcache entry is at least read-locked.
1609 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1611 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1612 if (hsame(v->m.DataVersion, d->f.versionNo) && v->callback)
1615 ReleaseWriteLock(&v->lock);
1619 /* avc - Write-locked unless aflags & 1 */
1621 afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1622 register struct vrequest *areq, afs_size_t * aoffset,
1623 afs_size_t * alen, int aflags)
1625 register afs_int32 i, code, code1 = 0, shortcut;
1626 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1627 register afs_int32 adjustsize = 0;
1633 afs_size_t maxGoodLength; /* amount of good data at server */
1634 struct rx_call *tcall;
1635 afs_size_t Position = 0;
1636 #ifdef AFS_64BIT_CLIENT
1638 afs_size_t lengthFound; /* as returned from server */
1639 #endif /* AFS_64BIT_CLIENT */
1640 afs_int32 size, tlen; /* size of segment to transfer */
1641 struct tlocal1 *tsmall = 0;
1642 register struct dcache *tdc;
1643 register struct osi_file *file;
1644 register struct conn *tc;
1646 struct server *newCallback = NULL;
1647 char setNewCallback;
1648 char setVcacheStatus;
1649 char doVcacheUpdate;
1651 int doAdjustSize = 0;
1652 int doReallyAdjustSize = 0;
1653 int overWriteWholeChunk = 0;
1657 struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
1658 osi_timeval_t xferStartTime, /*FS xfer start time */
1659 xferStopTime; /*FS xfer stop time */
1660 afs_size_t bytesToXfer; /* # bytes to xfer */
1661 afs_size_t bytesXferred; /* # bytes actually xferred */
1662 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1663 int fromReplica; /*Are we reading from a replica? */
1664 int numFetchLoops; /*# times around the fetch/analyze loop */
1665 #endif /* AFS_NOSTATS */
1667 AFS_STATCNT(afs_GetDCache);
1672 setLocks = aflags & 1;
1675 * Determine the chunk number and offset within the chunk corresponding
1676 * to the desired byte.
1678 if (avc->fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1681 chunk = AFS_CHUNK(abyte);
1684 /* come back to here if we waited for the cache to drain. */
1687 setNewCallback = setVcacheStatus = 0;
1691 ObtainWriteLock(&avc->lock, 616);
1693 ObtainReadLock(&avc->lock);
1698 * avc->lock(R) if setLocks && !slowPass
1699 * avc->lock(W) if !setLocks || slowPass
1704 /* check hints first! (might could use bcmp or some such...) */
1705 if ((tdc = avc->dchint)) {
1709 * The locking order between afs_xdcache and dcache lock matters.
1710 * The hint dcache entry could be anywhere, even on the free list.
1711 * Locking afs_xdcache ensures that noone is trying to pull dcache
1712 * entries from the free list, and thereby assuming them to be not
1713 * referenced and not locked.
1715 MObtainReadLock(&afs_xdcache);
1716 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1718 if (dcLocked && (tdc->index != NULLIDX)
1719 && !FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk
1720 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1721 /* got the right one. It might not be the right version, and it
1722 * might be fetching, but it's the right dcache entry.
1724 /* All this code should be integrated better with what follows:
1725 * I can save a good bit more time under a write lock if I do..
1727 ObtainWriteLock(&tdc->tlock, 603);
1729 ReleaseWriteLock(&tdc->tlock);
1731 MReleaseReadLock(&afs_xdcache);
1734 if (hsame(tdc->f.versionNo, avc->m.DataVersion)
1735 && !(tdc->dflags & DFFetching)) {
1737 afs_stats_cmperf.dcacheHits++;
1738 MObtainWriteLock(&afs_xdcache, 559);
1739 QRemove(&tdc->lruq);
1740 QAdd(&afs_DLRU, &tdc->lruq);
1741 MReleaseWriteLock(&afs_xdcache);
1744 * avc->lock(R) if setLocks && !slowPass
1745 * avc->lock(W) if !setLocks || slowPass
1752 ReleaseSharedLock(&tdc->lock);
1753 MReleaseReadLock(&afs_xdcache);
1761 * avc->lock(R) if setLocks && !slowPass
1762 * avc->lock(W) if !setLocks || slowPass
1763 * tdc->lock(S) if tdc
1766 if (!tdc) { /* If the hint wasn't the right dcache entry */
1768 * Hash on the [fid, chunk] and get the corresponding dcache index
1769 * after write-locking the dcache.
1774 * avc->lock(R) if setLocks && !slowPass
1775 * avc->lock(W) if !setLocks || slowPass
1778 i = DCHash(&avc->fid, chunk);
1779 /* check to make sure our space is fine */
1780 afs_MaybeWakeupTruncateDaemon();
1782 MObtainWriteLock(&afs_xdcache, 280);
1784 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1785 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1786 tdc = afs_GetDSlot(index, NULL);
1787 ReleaseReadLock(&tdc->tlock);
1790 * avc->lock(R) if setLocks && !slowPass
1791 * avc->lock(W) if !setLocks || slowPass
1794 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1795 /* Move it up in the beginning of the list */
1796 if (afs_dchashTbl[i] != index) {
1797 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1798 afs_dcnextTbl[index] = afs_dchashTbl[i];
1799 afs_dchashTbl[i] = index;
1801 MReleaseWriteLock(&afs_xdcache);
1802 ObtainSharedLock(&tdc->lock, 606);
1803 break; /* leaving refCount high for caller */
1809 index = afs_dcnextTbl[index];
1813 * If we didn't find the entry, we'll create one.
1815 if (index == NULLIDX) {
1818 * avc->lock(R) if setLocks
1819 * avc->lock(W) if !setLocks
1822 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1823 avc, ICL_TYPE_INT32, chunk);
1825 /* Make sure there is a free dcache entry for us to use */
1826 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1829 avc->states |= CDCLock;
1830 /* just need slots */
1831 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1833 avc->states &= ~CDCLock;
1834 if (afs_discardDCList != NULLIDX
1835 || afs_freeDCList != NULLIDX)
1837 /* If we can't get space for 5 mins we give up and panic */
1838 if (++downDCount > 300)
1839 osi_Panic("getdcache");
1840 MReleaseWriteLock(&afs_xdcache);
1843 * avc->lock(R) if setLocks
1844 * avc->lock(W) if !setLocks
1846 afs_osi_Wait(1000, 0, 0);
1851 if (afs_discardDCList == NULLIDX
1852 || ((aflags & 2) && afs_freeDCList != NULLIDX)) {
1854 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1855 tdc = afs_GetDSlot(afs_freeDCList, 0);
1856 osi_Assert(tdc->refCount == 1);
1857 ReleaseReadLock(&tdc->tlock);
1858 ObtainWriteLock(&tdc->lock, 604);
1859 afs_freeDCList = afs_dvnextTbl[tdc->index];
1862 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1863 tdc = afs_GetDSlot(afs_discardDCList, 0);
1864 osi_Assert(tdc->refCount == 1);
1865 ReleaseReadLock(&tdc->tlock);
1866 ObtainWriteLock(&tdc->lock, 605);
1867 afs_discardDCList = afs_dvnextTbl[tdc->index];
1868 afs_discardDCCount--;
1870 ((tdc->f.chunkBytes +
1871 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1872 tdc->f.states &= ~(DRO|DBackup|DRW);
1873 afs_DCMoveBucket(tdc, size, 0);
1874 afs_blocksDiscarded -= size;
1875 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1877 /* Truncate the chunk so zeroes get filled properly */
1878 file = afs_CFileOpen(tdc->f.inode);
1879 afs_CFileTruncate(file, 0);
1880 afs_CFileClose(file);
1881 afs_AdjustSize(tdc, 0);
1887 * avc->lock(R) if setLocks
1888 * avc->lock(W) if !setLocks
1894 * Fill in the newly-allocated dcache record.
1896 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1897 tdc->f.fid = avc->fid;
1898 if (avc->states & CRO)
1899 tdc->f.states = DRO;
1900 else if (avc->states & CBackup)
1901 tdc->f.states = DBackup;
1903 tdc->f.states = DRW;
1904 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1905 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1906 hones(tdc->f.versionNo); /* invalid value */
1907 tdc->f.chunk = chunk;
1908 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1910 if (tdc->lruq.prev == &tdc->lruq)
1911 osi_Panic("lruq 1");
1914 * Now add to the two hash chains - note that i is still set
1915 * from the above DCHash call.
1917 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1918 afs_dchashTbl[i] = tdc->index;
1919 i = DVHash(&avc->fid);
1920 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1921 afs_dvhashTbl[i] = tdc->index;
1922 tdc->dflags = DFEntryMod;
1924 afs_MaybeWakeupTruncateDaemon();
1925 MReleaseWriteLock(&afs_xdcache);
1926 ConvertWToSLock(&tdc->lock);
1931 /* vcache->dcache hint failed */
1934 * avc->lock(R) if setLocks && !slowPass
1935 * avc->lock(W) if !setLocks || slowPass
1938 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1939 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1940 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1941 hgetlo(avc->m.DataVersion));
1943 * Here we have the entry in tdc, with its refCount incremented.
1944 * Note: we don't use the S-lock on avc; it costs concurrency when
1945 * storing a file back to the server.
1949 * Not a newly created file so we need to check the file's length and
1950 * compare data versions since someone could have changed the data or we're
1951 * reading a file written elsewhere. We only want to bypass doing no-op
1952 * read rpcs on newly created files (dv of 0) since only then we guarantee
1953 * that this chunk's data hasn't been filled by another client.
1955 size = AFS_CHUNKSIZE(abyte);
1956 if (aflags & 4) /* called from write */
1958 else /* called from read */
1959 tlen = tdc->validPos - abyte;
1960 Position = AFS_CHUNKTOBASE(chunk);
1961 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1962 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1963 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1964 ICL_HANDLE_OFFSET(Position));
1965 if ((aflags & 4) && (hiszero(avc->m.DataVersion)))
1967 if ((AFS_CHUNKTOBASE(chunk) >= avc->m.Length) ||
1968 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
1969 overWriteWholeChunk = 1;
1970 if (doAdjustSize || overWriteWholeChunk) {
1971 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1973 #ifdef AFS_SGI64_ENV
1976 #else /* AFS_SGI64_ENV */
1979 #endif /* AFS_SGI64_ENV */
1980 #else /* AFS_SGI_ENV */
1983 #endif /* AFS_SGI_ENV */
1984 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->m.Length &&
1985 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1986 #if defined(AFS_SUN5_ENV) || defined(AFS_OSF_ENV)
1987 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
1989 if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
1991 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1992 !hsame(avc->m.DataVersion, tdc->f.versionNo))
1993 doReallyAdjustSize = 1;
1995 if (doReallyAdjustSize || overWriteWholeChunk) {
1996 /* no data in file to read at this position */
1997 UpgradeSToWLock(&tdc->lock, 607);
1999 file = afs_CFileOpen(tdc->f.inode);
2000 afs_CFileTruncate(file, 0);
2001 afs_CFileClose(file);
2002 afs_AdjustSize(tdc, 0);
2003 hset(tdc->f.versionNo, avc->m.DataVersion);
2004 tdc->dflags |= DFEntryMod;
2006 ConvertWToSLock(&tdc->lock);
2011 * We must read in the whole chunk if the version number doesn't
2015 /* don't need data, just a unique dcache entry */
2016 ObtainWriteLock(&afs_xdcache, 608);
2017 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2018 hadd32(afs_indexCounter, 1);
2019 ReleaseWriteLock(&afs_xdcache);
2021 updateV2DC(setLocks, avc, tdc, 553);
2022 if (vType(avc) == VDIR)
2025 *aoffset = AFS_CHUNKOFFSET(abyte);
2026 if (tdc->validPos < abyte)
2027 *alen = (afs_size_t) 0;
2029 *alen = tdc->validPos - abyte;
2030 ReleaseSharedLock(&tdc->lock);
2033 ReleaseWriteLock(&avc->lock);
2035 ReleaseReadLock(&avc->lock);
2037 return tdc; /* check if we're done */
2042 * avc->lock(R) if setLocks && !slowPass
2043 * avc->lock(W) if !setLocks || slowPass
2046 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2048 setNewCallback = setVcacheStatus = 0;
2052 * avc->lock(R) if setLocks && !slowPass
2053 * avc->lock(W) if !setLocks || slowPass
2056 if (!hsame(avc->m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2058 * Version number mismatch.
2060 UpgradeSToWLock(&tdc->lock, 609);
2063 * If data ever existed for this vnode, and this is a text object,
2064 * do some clearing. Now, you'd think you need only do the flush
2065 * when VTEXT is on, but VTEXT is turned off when the text object
2066 * is freed, while pages are left lying around in memory marked
2067 * with this vnode. If we would reactivate (create a new text
2068 * object from) this vnode, we could easily stumble upon some of
2069 * these old pages in pagein. So, we always flush these guys.
2070 * Sun has a wonderful lack of useful invariants in this system.
2072 * avc->flushDV is the data version # of the file at the last text
2073 * flush. Clearly, at least, we don't have to flush the file more
2074 * often than it changes
2076 if (hcmp(avc->flushDV, avc->m.DataVersion) < 0) {
2078 * By here, the cache entry is always write-locked. We can
2079 * deadlock if we call osi_Flush with the cache entry locked...
2080 * Unlock the dcache too.
2082 ReleaseWriteLock(&tdc->lock);
2083 if (setLocks && !slowPass)
2084 ReleaseReadLock(&avc->lock);
2086 ReleaseWriteLock(&avc->lock);
2090 * Call osi_FlushPages in open, read/write, and map, since it
2091 * is too hard here to figure out if we should lock the
2094 if (setLocks && !slowPass)
2095 ObtainReadLock(&avc->lock);
2097 ObtainWriteLock(&avc->lock, 66);
2098 ObtainWriteLock(&tdc->lock, 610);
2103 * avc->lock(R) if setLocks && !slowPass
2104 * avc->lock(W) if !setLocks || slowPass
2108 /* Watch for standard race condition around osi_FlushText */
2109 if (hsame(avc->m.DataVersion, tdc->f.versionNo)) {
2110 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2111 afs_stats_cmperf.dcacheHits++;
2112 ConvertWToSLock(&tdc->lock);
2116 /* Sleep here when cache needs to be drained. */
2117 if (setLocks && !slowPass
2118 && (afs_blocksUsed >
2119 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2120 /* Make sure truncate daemon is running */
2121 afs_MaybeWakeupTruncateDaemon();
2122 ObtainWriteLock(&tdc->tlock, 614);
2123 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2124 ReleaseWriteLock(&tdc->tlock);
2125 ReleaseWriteLock(&tdc->lock);
2126 ReleaseReadLock(&avc->lock);
2127 while ((afs_blocksUsed - afs_blocksDiscarded) >
2128 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2129 afs_WaitForCacheDrain = 1;
2130 afs_osi_Sleep(&afs_WaitForCacheDrain);
2132 afs_MaybeFreeDiscardedDCache();
2133 /* need to check if someone else got the chunk first. */
2134 goto RetryGetDCache;
2137 /* Do not fetch data beyond truncPos. */
2138 maxGoodLength = avc->m.Length;
2139 if (avc->truncPos < maxGoodLength)
2140 maxGoodLength = avc->truncPos;
2141 Position = AFS_CHUNKBASE(abyte);
2142 if (vType(avc) == VDIR) {
2143 size = avc->m.Length;
2144 if (size > tdc->f.chunkBytes) {
2145 /* pre-reserve space for file */
2146 afs_AdjustSize(tdc, size);
2148 size = 999999999; /* max size for transfer */
2150 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2151 /* don't read past end of good data on server */
2152 if (Position + size > maxGoodLength)
2153 size = maxGoodLength - Position;
2155 size = 0; /* Handle random races */
2156 if (size > tdc->f.chunkBytes) {
2157 /* pre-reserve space for file */
2158 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2159 /* max size for transfer still in size */
2162 if (afs_mariner && !tdc->f.chunk)
2163 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2165 * Right now, we only have one tool, and it's a hammer. So, we
2166 * fetch the whole file.
2168 DZap(tdc); /* pages in cache may be old */
2169 file = afs_CFileOpen(tdc->f.inode);
2170 afs_RemoveVCB(&avc->fid);
2171 tdc->f.states |= DWriting;
2172 tdc->dflags |= DFFetching;
2173 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2174 if (tdc->mflags & DFFetchReq) {
2175 tdc->mflags &= ~DFFetchReq;
2176 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2177 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2178 __FILE__, ICL_TYPE_INT32, __LINE__,
2179 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2183 (struct tlocal1 *)osi_AllocLargeSpace(sizeof(struct tlocal1));
2184 setVcacheStatus = 0;
2187 * Remember if we are doing the reading from a replicated volume,
2188 * and how many times we've zipped around the fetch/analyze loop.
2190 fromReplica = (avc->states & CRO) ? 1 : 0;
2192 accP = &(afs_stats_cmfullperf.accessinf);
2194 (accP->replicatedRefs)++;
2196 (accP->unreplicatedRefs)++;
2197 #endif /* AFS_NOSTATS */
2198 /* this is a cache miss */
2199 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2200 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
2201 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2204 afs_stats_cmperf.dcacheMisses++;
2207 * Dynamic root support: fetch data from local memory.
2209 if (afs_IsDynroot(avc)) {
2213 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2215 dynrootDir += Position;
2216 dynrootLen -= Position;
2217 if (size > dynrootLen)
2221 code = afs_CFileWrite(file, 0, dynrootDir, size);
2229 tdc->validPos = Position + size;
2230 afs_CFileTruncate(file, size); /* prune it */
2231 } else if (afs_IsDynrootMount(avc)) {
2235 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2237 dynrootDir += Position;
2238 dynrootLen -= Position;
2239 if (size > dynrootLen)
2243 code = afs_CFileWrite(file, 0, dynrootDir, size);
2251 tdc->validPos = Position + size;
2252 afs_CFileTruncate(file, size); /* prune it */
2253 } else if (afs_IsDynrootMount(avc)) {
2257 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2259 dynrootDir += Position;
2260 dynrootLen -= Position;
2261 if (size > dynrootLen)
2265 code = afs_CFileWrite(file, 0, dynrootDir, size);
2273 tdc->validPos = Position + size;
2274 afs_CFileTruncate(file, size); /* prune it */
2277 * Not a dynamic vnode: do the real fetch.
2282 * avc->lock(R) if setLocks && !slowPass
2283 * avc->lock(W) if !setLocks || slowPass
2287 tc = afs_Conn(&avc->fid, areq, SHARED_LOCK);
2289 afs_int32 length_hi, length, bytes;
2293 (accP->numReplicasAccessed)++;
2295 #endif /* AFS_NOSTATS */
2296 if (!setLocks || slowPass) {
2297 avc->callback = tc->srvr->server;
2299 newCallback = tc->srvr->server;
2304 tcall = rx_NewCall(tc->id);
2307 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
2308 #ifdef AFS_64BIT_CLIENT
2309 length_hi = code = 0;
2310 if (!afs_serverHasNo64Bit(tc)) {
2314 StartRXAFS_FetchData64(tcall,
2315 (struct AFSFid *)&avc->fid.
2316 Fid, Position, tsize);
2319 afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
2320 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32,
2324 rx_Read(tcall, (char *)&length_hi,
2327 if (bytes == sizeof(afs_int32)) {
2328 length_hi = ntohl(length_hi);
2331 code = rx_Error(tcall);
2333 code1 = rx_EndCall(tcall, code);
2335 tcall = (struct rx_call *)0;
2339 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
2340 if (Position > 0x7FFFFFFF) {
2347 tcall = rx_NewCall(tc->id);
2349 StartRXAFS_FetchData(tcall, (struct AFSFid *)
2354 afs_serverSetNo64Bit(tc);
2359 rx_Read(tcall, (char *)&length,
2362 if (bytes == sizeof(afs_int32)) {
2363 length = ntohl(length);
2365 code = rx_Error(tcall);
2368 FillInt64(lengthFound, length_hi, length);
2369 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64LENG,
2370 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
2372 ICL_HANDLE_OFFSET(lengthFound));
2373 #else /* AFS_64BIT_CLIENT */
2376 StartRXAFS_FetchData(tcall,
2377 (struct AFSFid *)&avc->fid.Fid,
2383 rx_Read(tcall, (char *)&length,
2386 if (bytes == sizeof(afs_int32)) {
2387 length = ntohl(length);
2389 code = rx_Error(tcall);
2392 #endif /* AFS_64BIT_CLIENT */
2397 &(afs_stats_cmfullperf.rpc.
2398 fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
2399 osi_GetuTime(&xferStartTime);
2402 afs_CacheFetchProc(tcall, file,
2403 (afs_size_t) Position, tdc,
2405 &bytesXferred, length);
2407 osi_GetuTime(&xferStopTime);
2408 (xferP->numXfers)++;
2410 (xferP->numSuccesses)++;
2411 afs_stats_XferSumBytes
2412 [AFS_STATS_FS_XFERIDX_FETCHDATA] +=
2414 (xferP->sumBytes) +=
2415 (afs_stats_XferSumBytes
2416 [AFS_STATS_FS_XFERIDX_FETCHDATA] >> 10);
2417 afs_stats_XferSumBytes
2418 [AFS_STATS_FS_XFERIDX_FETCHDATA] &= 0x3FF;
2419 if (bytesXferred < xferP->minBytes)
2420 xferP->minBytes = bytesXferred;
2421 if (bytesXferred > xferP->maxBytes)
2422 xferP->maxBytes = bytesXferred;
2425 * Tally the size of the object. Note: we tally the actual size,
2426 * NOT the number of bytes that made it out over the wire.
2428 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
2429 (xferP->count[0])++;
2430 else if (bytesToXfer <=
2431 AFS_STATS_MAXBYTES_BUCKET1)
2432 (xferP->count[1])++;
2433 else if (bytesToXfer <=
2434 AFS_STATS_MAXBYTES_BUCKET2)
2435 (xferP->count[2])++;
2436 else if (bytesToXfer <=
2437 AFS_STATS_MAXBYTES_BUCKET3)
2438 (xferP->count[3])++;
2439 else if (bytesToXfer <=
2440 AFS_STATS_MAXBYTES_BUCKET4)
2441 (xferP->count[4])++;
2442 else if (bytesToXfer <=
2443 AFS_STATS_MAXBYTES_BUCKET5)
2444 (xferP->count[5])++;
2445 else if (bytesToXfer <=
2446 AFS_STATS_MAXBYTES_BUCKET6)
2447 (xferP->count[6])++;
2448 else if (bytesToXfer <=
2449 AFS_STATS_MAXBYTES_BUCKET7)
2450 (xferP->count[7])++;
2452 (xferP->count[8])++;
2454 afs_stats_GetDiff(elapsedTime, xferStartTime,
2456 afs_stats_AddTo((xferP->sumTime), elapsedTime);
2457 afs_stats_SquareAddTo((xferP->sqrTime),
2459 if (afs_stats_TimeLessThan
2460 (elapsedTime, (xferP->minTime))) {
2461 afs_stats_TimeAssign((xferP->minTime),
2464 if (afs_stats_TimeGreaterThan
2465 (elapsedTime, (xferP->maxTime))) {
2466 afs_stats_TimeAssign((xferP->maxTime),
2472 afs_CacheFetchProc(tcall, file, Position, tdc,
2474 #endif /* AFS_NOSTATS */
2479 EndRXAFS_FetchData(tcall, &tsmall->OutStatus,
2487 code1 = rx_EndCall(tcall, code);
2496 /* callback could have been broken (or expired) in a race here,
2497 * but we return the data anyway. It's as good as we knew about
2498 * when we started. */
2500 * validPos is updated by CacheFetchProc, and can only be
2501 * modifed under a dcache write lock, which we've blocked out
2503 size = tdc->validPos - Position; /* actual segment size */
2506 afs_CFileTruncate(file, size); /* prune it */
2508 if (!setLocks || slowPass) {
2509 ObtainWriteLock(&afs_xcbhash, 453);
2510 afs_DequeueCallback(avc);
2511 avc->states &= ~(CStatd | CUnique);
2512 avc->callback = NULL;
2513 ReleaseWriteLock(&afs_xcbhash);
2514 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2515 osi_dnlc_purgedp(avc);
2517 /* Something lost. Forget about performance, and go
2518 * back with a vcache write lock.
2520 afs_CFileTruncate(file, 0);
2521 afs_AdjustSize(tdc, 0);
2522 afs_CFileClose(file);
2523 osi_FreeLargeSpace(tsmall);
2525 ReleaseWriteLock(&tdc->lock);
2528 ReleaseReadLock(&avc->lock);
2530 goto RetryGetDCache;
2534 } while (afs_Analyze
2535 (tc, code, &avc->fid, areq,
2536 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2540 * avc->lock(R) if setLocks && !slowPass
2541 * avc->lock(W) if !setLocks || slowPass
2547 * In the case of replicated access, jot down info on the number of
2548 * attempts it took before we got through or gave up.
2551 if (numFetchLoops <= 1)
2552 (accP->refFirstReplicaOK)++;
2553 if (numFetchLoops > accP->maxReplicasPerRef)
2554 accP->maxReplicasPerRef = numFetchLoops;
2556 #endif /* AFS_NOSTATS */
2558 tdc->dflags &= ~DFFetching;
2559 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2560 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2561 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2562 tdc, ICL_TYPE_INT32, tdc->dflags);
2563 if (avc->execsOrWriters == 0)
2564 tdc->f.states &= ~DWriting;
2566 /* now, if code != 0, we have an error and should punt.
2567 * note that we have the vcache write lock, either because
2568 * !setLocks or slowPass.
2571 afs_CFileTruncate(file, 0);
2572 afs_AdjustSize(tdc, 0);
2573 afs_CFileClose(file);
2574 ZapDCE(tdc); /* sets DFEntryMod */
2575 if (vType(avc) == VDIR) {
2578 tdc->f.states &= ~(DRO|DBackup|DRW);
2579 afs_DCMoveBucket(tdc, 0, 0);
2580 ReleaseWriteLock(&tdc->lock);
2582 ObtainWriteLock(&afs_xcbhash, 454);
2583 afs_DequeueCallback(avc);
2584 avc->states &= ~(CStatd | CUnique);
2585 ReleaseWriteLock(&afs_xcbhash);
2586 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2587 osi_dnlc_purgedp(avc);
2590 * avc->lock(W); assert(!setLocks || slowPass)
2592 osi_Assert(!setLocks || slowPass);
2593 tdc->f.states &= ~(DRO|DBackup|DRW);
2594 afs_DCMoveBucket(tdc, 0, 0);
2599 /* otherwise we copy in the just-fetched info */
2600 afs_CFileClose(file);
2601 afs_AdjustSize(tdc, size); /* new size */
2603 * Copy appropriate fields into vcache. Status is
2604 * copied later where we selectively acquire the
2605 * vcache write lock.
2608 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2610 setVcacheStatus = 1;
2611 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2612 tsmall->OutStatus.DataVersion);
2613 tdc->dflags |= DFEntryMod;
2614 afs_indexFlags[tdc->index] |= IFEverUsed;
2615 ConvertWToSLock(&tdc->lock);
2616 } /*Data version numbers don't match */
2619 * Data version numbers match.
2621 afs_stats_cmperf.dcacheHits++;
2622 } /*Data version numbers match */
2624 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2628 * avc->lock(R) if setLocks && !slowPass
2629 * avc->lock(W) if !setLocks || slowPass
2630 * tdc->lock(S) if tdc
2634 * See if this was a reference to a file in the local cell.
2636 if (afs_IsPrimaryCellNum(avc->fid.Cell))
2637 afs_stats_cmperf.dlocalAccesses++;
2639 afs_stats_cmperf.dremoteAccesses++;
2641 /* Fix up LRU info */
2644 MObtainWriteLock(&afs_xdcache, 602);
2645 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2646 hadd32(afs_indexCounter, 1);
2647 MReleaseWriteLock(&afs_xdcache);
2649 /* return the data */
2650 if (vType(avc) == VDIR)
2653 *aoffset = AFS_CHUNKOFFSET(abyte);
2654 *alen = (tdc->f.chunkBytes - *aoffset);
2655 ReleaseSharedLock(&tdc->lock);
2660 * avc->lock(R) if setLocks && !slowPass
2661 * avc->lock(W) if !setLocks || slowPass
2664 /* Fix up the callback and status values in the vcache */
2666 if (setLocks && !slowPass) {
2669 * This is our dirty little secret to parallel fetches.
2670 * We don't write-lock the vcache while doing the fetch,
2671 * but potentially we'll need to update the vcache after
2672 * the fetch is done.
2674 * Drop the read lock and try to re-obtain the write
2675 * lock. If the vcache still has the same DV, it's
2676 * ok to go ahead and install the new data.
2678 afs_hyper_t currentDV, statusDV;
2680 hset(currentDV, avc->m.DataVersion);
2682 if (setNewCallback && avc->callback != newCallback)
2686 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2687 tsmall->OutStatus.DataVersion);
2689 if (setVcacheStatus && avc->m.Length != tsmall->OutStatus.Length)
2691 if (setVcacheStatus && !hsame(currentDV, statusDV))
2695 ReleaseReadLock(&avc->lock);
2697 if (doVcacheUpdate) {
2698 ObtainWriteLock(&avc->lock, 615);
2699 if (!hsame(avc->m.DataVersion, currentDV)) {
2700 /* We lose. Someone will beat us to it. */
2702 ReleaseWriteLock(&avc->lock);
2707 /* With slow pass, we've already done all the updates */
2709 ReleaseWriteLock(&avc->lock);
2712 /* Check if we need to perform any last-minute fixes with a write-lock */
2713 if (!setLocks || doVcacheUpdate) {
2715 avc->callback = newCallback;
2716 if (tsmall && setVcacheStatus)
2717 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2719 ReleaseWriteLock(&avc->lock);
2723 osi_FreeLargeSpace(tsmall);
2726 } /*afs_GetDCache */
2730 * afs_WriteThroughDSlots
2733 * Sweep through the dcache slots and write out any modified
2734 * in-memory data back on to our caching store.
2740 * The afs_xdcache is write-locked through this whole affair.
2743 afs_WriteThroughDSlots(void)
2745 register struct dcache *tdc;
2746 register afs_int32 i, touchedit = 0;
2748 struct afs_q DirtyQ, *tq;
2750 AFS_STATCNT(afs_WriteThroughDSlots);
2753 * Because of lock ordering, we can't grab dcache locks while
2754 * holding afs_xdcache. So we enter xdcache, get a reference
2755 * for every dcache entry, and exit xdcache.
2757 MObtainWriteLock(&afs_xdcache, 283);
2759 for (i = 0; i < afs_cacheFiles; i++) {
2760 tdc = afs_indexTable[i];
2762 /* Grab tlock in case the existing refcount isn't zero */
2763 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2764 ObtainWriteLock(&tdc->tlock, 623);
2766 ReleaseWriteLock(&tdc->tlock);
2768 QAdd(&DirtyQ, &tdc->dirty);
2771 MReleaseWriteLock(&afs_xdcache);
2774 * Now, for each dcache entry we found, check if it's dirty.
2775 * If so, get write-lock, get afs_xdcache, which protects
2776 * afs_cacheInodep, and flush it. Don't forget to put back
2780 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2782 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2784 if (tdc->dflags & DFEntryMod) {
2787 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2789 /* Now that we have the write lock, double-check */
2790 if (wrLock && (tdc->dflags & DFEntryMod)) {
2791 tdc->dflags &= ~DFEntryMod;
2792 MObtainWriteLock(&afs_xdcache, 620);
2793 afs_WriteDCache(tdc, 1);
2794 MReleaseWriteLock(&afs_xdcache);
2798 ReleaseWriteLock(&tdc->lock);
2804 MObtainWriteLock(&afs_xdcache, 617);
2805 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2806 /* Touch the file to make sure that the mtime on the file is kept
2807 * up-to-date to avoid losing cached files on cold starts because
2808 * their mtime seems old...
2810 struct afs_fheader theader;
2812 theader.magic = AFS_FHMAGIC;
2813 theader.firstCSize = AFS_FIRSTCSIZE;
2814 theader.otherCSize = AFS_OTHERCSIZE;
2815 theader.version = AFS_CI_VERSION;
2816 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2818 MReleaseWriteLock(&afs_xdcache);
2825 * Return a pointer to an freshly initialized dcache entry using
2826 * a memory-based cache. The tlock will be read-locked.
2829 * aslot : Dcache slot to look at.
2830 * tmpdc : Ptr to dcache entry.
2833 * Must be called with afs_xdcache write-locked.
2837 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2839 register struct dcache *tdc;
2842 AFS_STATCNT(afs_MemGetDSlot);
2843 if (CheckLock(&afs_xdcache) != -1)
2844 osi_Panic("getdslot nolock");
2845 if (aslot < 0 || aslot >= afs_cacheFiles)
2846 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2847 tdc = afs_indexTable[aslot];
2849 QRemove(&tdc->lruq); /* move to queue head */
2850 QAdd(&afs_DLRU, &tdc->lruq);
2851 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2852 ObtainWriteLock(&tdc->tlock, 624);
2854 ConvertWToRLock(&tdc->tlock);
2857 if (tmpdc == NULL) {
2858 if (!afs_freeDSList)
2859 afs_GetDownDSlot(4);
2860 if (!afs_freeDSList) {
2861 /* none free, making one is better than a panic */
2862 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2863 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2864 #ifdef KERNEL_HAVE_PIN
2865 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2868 tdc = afs_freeDSList;
2869 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2872 tdc->dflags = 0; /* up-to-date, not in free q */
2874 QAdd(&afs_DLRU, &tdc->lruq);
2875 if (tdc->lruq.prev == &tdc->lruq)
2876 osi_Panic("lruq 3");
2882 /* initialize entry */
2883 tdc->f.fid.Cell = 0;
2884 tdc->f.fid.Fid.Volume = 0;
2886 hones(tdc->f.versionNo);
2887 tdc->f.inode = aslot;
2888 tdc->dflags |= DFEntryMod;
2891 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2894 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2895 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2896 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2899 RWLOCK_INIT(&tdc->lock, "dcache lock");
2900 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2901 RWLOCK_INIT(&tdc->mflock, "dcache flock");
2902 ObtainReadLock(&tdc->tlock);
2905 afs_indexTable[aslot] = tdc;
2908 } /*afs_MemGetDSlot */
2910 unsigned int last_error = 0, lasterrtime = 0;
2916 * Return a pointer to an freshly initialized dcache entry using
2917 * a UFS-based disk cache. The dcache tlock will be read-locked.
2920 * aslot : Dcache slot to look at.
2921 * tmpdc : Ptr to dcache entry.
2924 * afs_xdcache lock write-locked.
2927 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2929 register afs_int32 code;
2930 register struct dcache *tdc;
2934 AFS_STATCNT(afs_UFSGetDSlot);
2935 if (CheckLock(&afs_xdcache) != -1)
2936 osi_Panic("getdslot nolock");
2937 if (aslot < 0 || aslot >= afs_cacheFiles)
2938 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2939 tdc = afs_indexTable[aslot];
2941 QRemove(&tdc->lruq); /* move to queue head */
2942 QAdd(&afs_DLRU, &tdc->lruq);
2943 /* Grab tlock in case refCount != 0 */
2944 ObtainWriteLock(&tdc->tlock, 625);
2946 ConvertWToRLock(&tdc->tlock);
2949 /* otherwise we should read it in from the cache file */
2951 * If we weren't passed an in-memory region to place the file info,
2952 * we have to allocate one.
2954 if (tmpdc == NULL) {
2955 if (!afs_freeDSList)
2956 afs_GetDownDSlot(4);
2957 if (!afs_freeDSList) {
2958 /* none free, making one is better than a panic */
2959 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2960 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2961 #ifdef KERNEL_HAVE_PIN
2962 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2965 tdc = afs_freeDSList;
2966 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2969 tdc->dflags = 0; /* up-to-date, not in free q */
2971 QAdd(&afs_DLRU, &tdc->lruq);
2972 if (tdc->lruq.prev == &tdc->lruq)
2973 osi_Panic("lruq 3");
2980 * Seek to the aslot'th entry and read it in.
2983 afs_osi_Read(afs_cacheInodep,
2984 sizeof(struct fcache) * aslot +
2985 sizeof(struct afs_fheader), (char *)(&tdc->f),
2986 sizeof(struct fcache));
2988 if (code != sizeof(struct fcache))
2990 if (!afs_CellNumValid(tdc->f.fid.Cell))
2994 tdc->f.fid.Cell = 0;
2995 tdc->f.fid.Fid.Volume = 0;
2997 hones(tdc->f.versionNo);
2998 tdc->dflags |= DFEntryMod;
2999 #if defined(KERNEL_HAVE_UERROR)
3000 last_error = getuerror();
3002 lasterrtime = osi_Time();
3003 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
3004 tdc->f.states &= ~(DRO|DBackup|DRW);
3005 afs_DCMoveBucket(tdc, 0, 0);
3008 if (tdc->f.states & DRO) {
3009 afs_DCMoveBucket(tdc, 0, 2);
3010 } else if (tdc->f.states & DBackup) {
3011 afs_DCMoveBucket(tdc, 0, 1);
3013 afs_DCMoveBucket(tdc, 0, 1);
3019 if (tdc->f.chunk >= 0)
3020 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
3025 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
3026 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
3027 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
3030 RWLOCK_INIT(&tdc->lock, "dcache lock");
3031 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
3032 RWLOCK_INIT(&tdc->mflock, "dcache flock");
3033 ObtainReadLock(&tdc->tlock);
3036 * If we didn't read into a temporary dcache region, update the
3037 * slot pointer table.
3040 afs_indexTable[aslot] = tdc;
3043 } /*afs_UFSGetDSlot */
3051 * write a particular dcache entry back to its home in the
3055 * adc : Pointer to the dcache entry to write.
3056 * atime : If true, set the modtime on the file to the current time.
3059 * Must be called with the afs_xdcache lock at least read-locked,
3060 * and dcache entry at least read-locked.
3061 * The reference count is not changed.
3065 afs_WriteDCache(register struct dcache *adc, int atime)
3067 register afs_int32 code;
3069 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
3071 AFS_STATCNT(afs_WriteDCache);
3073 adc->f.modTime = osi_Time();
3075 * Seek to the right dcache slot and write the in-memory image out to disk.
3077 afs_cellname_write();
3079 afs_osi_Write(afs_cacheInodep,
3080 sizeof(struct fcache) * adc->index +
3081 sizeof(struct afs_fheader), (char *)(&adc->f),
3082 sizeof(struct fcache));
3083 if (code != sizeof(struct fcache))
3094 * Wake up users of a particular file waiting for stores to take
3098 * avc : Ptr to related vcache entry.
3101 * Nothing interesting.
3105 afs_wakeup(register struct vcache *avc)
3108 register struct brequest *tb;
3110 AFS_STATCNT(afs_wakeup);
3111 for (i = 0; i < NBRS; i++, tb++) {
3112 /* if request is valid and for this file, we've found it */
3113 if (tb->refCount > 0 && avc == tb->vc) {
3116 * If CSafeStore is on, then we don't awaken the guy
3117 * waiting for the store until the whole store has finished.
3118 * Otherwise, we do it now. Note that if CSafeStore is on,
3119 * the BStore routine actually wakes up the user, instead
3121 * I think this is redundant now because this sort of thing
3122 * is already being handled by the higher-level code.
3124 if ((avc->states & CSafeStore) == 0) {
3126 tb->flags |= BUVALID;
3127 if (tb->flags & BUWAIT) {
3128 tb->flags &= ~BUWAIT;
3143 * Given a file name and inode, set up that file to be an
3144 * active member in the AFS cache. This also involves checking
3145 * the usability of its data.
3148 * afile : Name of the cache file to initialize.
3149 * ainode : Inode of the file.
3152 * This function is called only during initialization.
3156 afs_InitCacheFile(char *afile, ino_t ainode)
3158 register afs_int32 code;
3159 #if defined(AFS_LINUX22_ENV)
3160 struct dentry *filevp;
3162 struct vnode *filevp;
3166 struct osi_file *tfile;
3167 struct osi_stat tstat;
3168 register struct dcache *tdc;
3170 AFS_STATCNT(afs_InitCacheFile);
3171 index = afs_stats_cmperf.cacheNumEntries;
3172 if (index >= afs_cacheFiles)
3175 MObtainWriteLock(&afs_xdcache, 282);
3176 tdc = afs_GetDSlot(index, NULL);
3177 ReleaseReadLock(&tdc->tlock);
3178 MReleaseWriteLock(&afs_xdcache);
3180 ObtainWriteLock(&tdc->lock, 621);
3181 MObtainWriteLock(&afs_xdcache, 622);
3183 code = gop_lookupname(afile, AFS_UIOSYS, 0, &filevp);
3185 ReleaseWriteLock(&afs_xdcache);
3186 ReleaseWriteLock(&tdc->lock);
3191 * We have a VN_HOLD on filevp. Get the useful info out and
3192 * return. We make use of the fact that the cache is in the
3193 * UFS file system, and just record the inode number.
3195 #ifdef AFS_LINUX22_ENV
3196 tdc->f.inode = VTOI(filevp->d_inode)->i_number;
3199 tdc->f.inode = afs_vnodeToInumber(filevp);
3201 #endif /* AFS_LINUX22_ENV */
3203 tdc->f.inode = ainode;
3206 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3208 tfile = osi_UFSOpen(tdc->f.inode);
3209 code = afs_osi_Stat(tfile, &tstat);
3211 osi_Panic("initcachefile stat");
3214 * If file size doesn't match the cache info file, it's probably bad.
3216 if (tdc->f.chunkBytes != tstat.size)
3218 tdc->f.chunkBytes = 0;
3221 * If file changed within T (120?) seconds of cache info file, it's
3222 * probably bad. In addition, if slot changed within last T seconds,
3223 * the cache info file may be incorrectly identified, and so slot
3226 if (cacheInfoModTime < tstat.mtime + 120)
3228 if (cacheInfoModTime < tdc->f.modTime + 120)
3230 /* In case write through is behind, make sure cache items entry is
3231 * at least as new as the chunk.
3233 if (tdc->f.modTime < tstat.mtime)
3236 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3237 if (tstat.size != 0)
3238 osi_UFSTruncate(tfile, 0);
3239 tdc->f.states &= ~(DRO|DBackup|DRW);
3240 afs_DCMoveBucket(tdc, 0, 0);
3241 /* put entry in free cache slot list */
3242 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3243 afs_freeDCList = index;
3245 afs_indexFlags[index] |= IFFree;
3246 afs_indexUnique[index] = 0;
3249 * We must put this entry in the appropriate hash tables.
3250 * Note that i is still set from the above DCHash call
3252 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3253 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3254 afs_dchashTbl[code] = tdc->index;
3255 code = DVHash(&tdc->f.fid);
3256 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3257 afs_dvhashTbl[code] = tdc->index;
3258 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3260 /* has nontrivial amt of data */
3261 afs_indexFlags[index] |= IFEverUsed;
3262 afs_stats_cmperf.cacheFilesReused++;
3264 * Initialize index times to file's mod times; init indexCounter
3267 hset32(afs_indexTimes[index], tstat.atime);
3268 if (hgetlo(afs_indexCounter) < tstat.atime) {
3269 hset32(afs_indexCounter, tstat.atime);
3271 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3272 } /*File is not bad */
3274 osi_UFSClose(tfile);
3275 tdc->f.states &= ~DWriting;
3276 tdc->dflags &= ~DFEntryMod;
3277 /* don't set f.modTime; we're just cleaning up */
3278 afs_WriteDCache(tdc, 0);
3279 ReleaseWriteLock(&afs_xdcache);
3280 ReleaseWriteLock(&tdc->lock);
3282 afs_stats_cmperf.cacheNumEntries++;
3287 /*Max # of struct dcache's resident at any time*/
3289 * If 'dchint' is enabled then in-memory dcache min is increased because of
3298 * Initialize dcache related variables.
3301 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3303 register struct dcache *tdp;
3307 afs_freeDCList = NULLIDX;
3308 afs_discardDCList = NULLIDX;
3309 afs_freeDCCount = 0;
3310 afs_freeDSList = NULL;
3311 hzero(afs_indexCounter);
3313 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3319 if (achunk < 0 || achunk > 30)
3320 achunk = 13; /* Use default */
3321 AFS_SETCHUNKSIZE(achunk);
3327 if (aflags & AFSCALL_INIT_MEMCACHE) {
3329 * Use a memory cache instead of a disk cache
3331 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3332 afs_cacheType = &afs_MemCacheOps;
3333 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3334 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3335 /* ablocks is reported in 1K blocks */
3336 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3338 printf("afsd: memory cache too large for available memory.\n");
3339 printf("afsd: AFS files cannot be accessed.\n\n");
3341 afiles = ablocks = 0;
3343 printf("Memory cache: Allocating %d dcache entries...",
3346 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3347 afs_cacheType = &afs_UfsCacheOps;
3350 if (aDentries > 512)
3351 afs_dhashsize = 2048;
3352 /* initialize hash tables */
3354 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3356 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3357 for (i = 0; i < afs_dhashsize; i++) {
3358 afs_dvhashTbl[i] = NULLIDX;
3359 afs_dchashTbl[i] = NULLIDX;
3361 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3362 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3363 for (i = 0; i < afiles; i++) {
3364 afs_dvnextTbl[i] = NULLIDX;
3365 afs_dcnextTbl[i] = NULLIDX;
3368 /* Allocate and zero the pointer array to the dcache entries */
3369 afs_indexTable = (struct dcache **)
3370 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3371 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3373 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3374 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3376 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3377 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3378 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3379 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3381 /* Allocate and thread the struct dcache entries themselves */
3382 tdp = afs_Initial_freeDSList =
3383 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3384 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3385 #ifdef KERNEL_HAVE_PIN
3386 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3387 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3388 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3389 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3390 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3391 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3392 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3393 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3394 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3397 afs_freeDSList = &tdp[0];
3398 for (i = 0; i < aDentries - 1; i++) {
3399 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3400 RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3401 RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3402 RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3404 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3405 RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3406 RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3407 RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3409 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3410 afs_cacheBlocks = ablocks;
3411 afs_ComputeCacheParms(); /* compute parms based on cache size */
3413 afs_dcentries = aDentries;
3415 afs_stats_cmperf.cacheBucket0_Discarded =
3416 afs_stats_cmperf.cacheBucket1_Discarded =
3417 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3427 shutdown_dcache(void)
3431 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3432 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3433 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3434 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3435 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3436 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3437 afs_osi_Free(afs_Initial_freeDSList,
3438 afs_dcentries * sizeof(struct dcache));
3439 #ifdef KERNEL_HAVE_PIN
3440 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3441 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3442 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3443 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3444 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3445 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3446 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3450 for (i = 0; i < afs_dhashsize; i++) {
3451 afs_dvhashTbl[i] = NULLIDX;
3452 afs_dchashTbl[i] = NULLIDX;
3455 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3456 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3458 afs_blocksUsed = afs_dcentries = 0;
3459 afs_stats_cmperf.cacheBucket0_Discarded =
3460 afs_stats_cmperf.cacheBucket1_Discarded =
3461 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3462 hzero(afs_indexCounter);
3464 afs_freeDCCount = 0;
3465 afs_freeDCList = NULLIDX;
3466 afs_discardDCList = NULLIDX;
3467 afs_freeDSList = afs_Initial_freeDSList = 0;
3469 LOCK_INIT(&afs_xdcache, "afs_xdcache");