2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
13 #include <afsconfig.h>
14 #include "afs/param.h"
17 #include "afs/sysincludes.h" /*Standard vendor system headers */
18 #include "afsincludes.h" /*AFS-based standard headers */
19 #include "afs/afs_stats.h" /* statistics */
20 #include "afs/afs_cbqueue.h"
21 #include "afs/afs_osidnlc.h"
23 /* Forward declarations. */
24 static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
25 static void afs_FreeDiscardedDCache(void);
26 static void afs_DiscardDCache(struct dcache *);
27 static void afs_FreeDCache(struct dcache *);
29 static afs_int32 afs_DCGetBucket(struct vcache *);
30 static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
31 static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
32 static void afs_DCSizeInit(void);
33 static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);
36 * --------------------- Exported definitions ---------------------
39 afs_int32 afs_blocksUsed_0; /*!< 1K blocks in cache - in theory is zero */
40 afs_int32 afs_blocksUsed_1; /*!< 1K blocks in cache */
41 afs_int32 afs_blocksUsed_2; /*!< 1K blocks in cache */
42 afs_int32 afs_pct1 = -1;
43 afs_int32 afs_pct2 = -1;
44 afs_uint32 afs_tpct1 = 0;
45 afs_uint32 afs_tpct2 = 0;
46 afs_uint32 splitdcache = 0;
48 afs_lock_t afs_xdcache; /*!< Lock: alloc new disk cache entries */
49 afs_int32 afs_freeDCList; /*!< Free list for disk cache entries */
50 afs_int32 afs_freeDCCount; /*!< Count of elts in freeDCList */
51 afs_int32 afs_discardDCList; /*!< Discarded disk cache entries */
52 afs_int32 afs_discardDCCount; /*!< Count of elts in discardDCList */
53 struct dcache *afs_freeDSList; /*!< Free list for disk slots */
54 struct dcache *afs_Initial_freeDSList; /*!< Initial list for above */
55 afs_dcache_id_t cacheInode; /*!< Inode for CacheItems file */
56 struct osi_file *afs_cacheInodep = 0; /*!< file for CacheItems inode */
57 struct afs_q afs_DLRU; /*!< dcache LRU */
58 afs_int32 afs_dhashsize = 1024;
59 afs_int32 *afs_dvhashTbl; /*!< Data cache hash table: hashed by FID + chunk number. */
60 afs_int32 *afs_dchashTbl; /*!< Data cache hash table: hashed by FID. */
61 afs_int32 *afs_dvnextTbl; /*!< Dcache hash table links */
62 afs_int32 *afs_dcnextTbl; /*!< Dcache hash table links */
63 struct dcache **afs_indexTable; /*!< Pointers to dcache entries */
64 afs_hyper_t *afs_indexTimes; /*!< Dcache entry Access times */
65 afs_int32 *afs_indexUnique; /*!< dcache entry Fid.Unique */
66 unsigned char *afs_indexFlags; /*!< (only one) Is there data there? */
67 afs_hyper_t afs_indexCounter; /*!< Fake time for marking index
69 afs_int32 afs_cacheFiles = 0; /*!< Size of afs_indexTable */
70 afs_int32 afs_cacheBlocks; /*!< 1K blocks in cache */
71 afs_int32 afs_cacheStats; /*!< Stat entries in cache */
72 afs_int32 afs_blocksUsed; /*!< Number of blocks in use */
73 afs_int32 afs_blocksDiscarded; /*!<Blocks freed but not truncated */
74 afs_int32 afs_fsfragsize = AFS_MIN_FRAGSIZE; /*!< Underlying Filesystem minimum unit
75 *of disk allocation usually 1K
76 *this value is (truefrag -1 ) to
77 *save a bunch of subtracts... */
78 #ifdef AFS_64BIT_CLIENT
79 #ifdef AFS_VM_RDWR_ENV
80 afs_size_t afs_vmMappingEnd; /* !< For large files (>= 2GB) the VM
81 * mapping an 32bit addressing machines
82 * can only be used below the 2 GB
83 * line. From this point upwards we
84 * must do direct I/O into the cache
85 * files. The value should be on a
87 #endif /* AFS_VM_RDWR_ENV */
88 #endif /* AFS_64BIT_CLIENT */
90 /* The following is used to ensure that new dcache's aren't obtained when
91 * the cache is nearly full.
93 int afs_WaitForCacheDrain = 0;
94 int afs_TruncateDaemonRunning = 0;
95 int afs_CacheTooFull = 0;
97 afs_int32 afs_dcentries; /*!< In-memory dcache entries */
100 int dcacheDisabled = 0;
102 struct afs_cacheOps afs_UfsCacheOps = {
103 #ifndef HAVE_STRUCT_LABEL_SUPPORT
116 .truncate = osi_UFSTruncate,
117 .fread = afs_osi_Read,
118 .fwrite = afs_osi_Write,
119 .close = osi_UFSClose,
120 .vreadUIO = afs_UFSReadUIO,
121 .vwriteUIO = afs_UFSWriteUIO,
122 .GetDSlot = afs_UFSGetDSlot,
123 .GetVolSlot = afs_UFSGetVolSlot,
124 .HandleLink = afs_UFSHandleLink,
128 struct afs_cacheOps afs_MemCacheOps = {
129 #ifndef HAVE_STRUCT_LABEL_SUPPORT
131 afs_MemCacheTruncate,
141 .open = afs_MemCacheOpen,
142 .truncate = afs_MemCacheTruncate,
143 .fread = afs_MemReadBlk,
144 .fwrite = afs_MemWriteBlk,
145 .close = afs_MemCacheClose,
146 .vreadUIO = afs_MemReadUIO,
147 .vwriteUIO = afs_MemWriteUIO,
148 .GetDSlot = afs_MemGetDSlot,
149 .GetVolSlot = afs_MemGetVolSlot,
150 .HandleLink = afs_MemHandleLink,
154 int cacheDiskType; /*Type of backing disk for cache */
155 struct afs_cacheOps *afs_cacheType;
158 * Where is this vcache's entry associated dcache located/
159 * \param avc The vcache entry.
160 * \return Bucket index:
165 afs_DCGetBucket(struct vcache *avc)
170 /* This should be replaced with some sort of user configurable function */
171 if (avc->f.states & CRO) {
173 } else if (avc->f.states & CBackup) {
183 * Readjust a dcache's size.
185 * \param adc The dcache to be adjusted.
186 * \param oldSize Old size for the dcache.
187 * \param newSize The new size to be adjusted to.
191 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
193 afs_int32 adjustSize = newSize - oldSize;
201 afs_blocksUsed_0 += adjustSize;
202 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
205 afs_blocksUsed_1 += adjustSize;
206 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
209 afs_blocksUsed_2 += adjustSize;
210 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
218 * Move a dcache from one bucket to another.
220 * \param adc Operate on this dcache.
221 * \param size Size in bucket (?).
222 * \param newBucket Destination bucket.
226 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
231 /* Substract size from old bucket. */
235 afs_blocksUsed_0 -= size;
238 afs_blocksUsed_1 -= size;
241 afs_blocksUsed_2 -= size;
245 /* Set new bucket and increase destination bucket size. */
246 adc->bucket = newBucket;
251 afs_blocksUsed_0 += size;
254 afs_blocksUsed_1 += size;
257 afs_blocksUsed_2 += size;
265 * Init split caches size.
270 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
279 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
284 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
285 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
287 /* Short cut: if we don't know about it, try to kill it */
288 if (phase < 2 && afs_blocksUsed_0)
291 if (afs_pct1 > afs_tpct1)
293 if (afs_pct2 > afs_tpct2)
295 return 0; /* unlikely */
300 * Warn about failing to store a file.
302 * \param acode Associated error code.
303 * \param avolume Volume involved.
304 * \param aflags How to handle the output:
305 * aflags & 1: Print out on console
306 * aflags & 2: Print out on controlling tty
308 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
312 afs_StoreWarn(afs_int32 acode, afs_int32 avolume,
315 static char problem_fmt[] =
316 "afs: failed to store file in volume %d (%s)\n";
317 static char problem_fmt_w_error[] =
318 "afs: failed to store file in volume %d (error %d)\n";
319 static char netproblems[] = "network problems";
320 static char partfull[] = "partition full";
321 static char overquota[] = "over quota";
323 AFS_STATCNT(afs_StoreWarn);
329 afs_warn(problem_fmt, avolume, netproblems);
331 afs_warnuser(problem_fmt, avolume, netproblems);
332 } else if (acode == ENOSPC) {
337 afs_warn(problem_fmt, avolume, partfull);
339 afs_warnuser(problem_fmt, avolume, partfull);
342 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
343 * Instead ENOSPC will be sent...
345 if (acode == EDQUOT) {
350 afs_warn(problem_fmt, avolume, overquota);
352 afs_warnuser(problem_fmt, avolume, overquota);
360 afs_warn(problem_fmt_w_error, avolume, acode);
362 afs_warnuser(problem_fmt_w_error, avolume, acode);
367 * Try waking up truncation daemon, if it's worth it.
370 afs_MaybeWakeupTruncateDaemon(void)
372 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
373 afs_CacheTooFull = 1;
374 if (!afs_TruncateDaemonRunning)
375 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
376 } else if (!afs_TruncateDaemonRunning
377 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
378 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
385 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
386 * struct so we need only export one symbol for AIX.
388 static struct CTD_stats {
389 osi_timeval_t CTD_beforeSleep;
390 osi_timeval_t CTD_afterSleep;
391 osi_timeval_t CTD_sleepTime;
392 osi_timeval_t CTD_runTime;
396 u_int afs_min_cache = 0;
399 * Keeps the cache clean and free by truncating uneeded files, when used.
404 afs_CacheTruncateDaemon(void)
406 osi_timeval_t CTD_tmpTime;
410 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
412 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
414 osi_GetuTime(&CTD_stats.CTD_afterSleep);
415 afs_TruncateDaemonRunning = 1;
417 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
418 ObtainWriteLock(&afs_xdcache, 266);
419 if (afs_CacheTooFull) {
420 int space_needed, slots_needed;
421 /* if we get woken up, we should try to clean something out */
422 for (counter = 0; counter < 10; counter++) {
424 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
426 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
427 afs_GetDownD(slots_needed, &space_needed, 0);
428 if ((space_needed <= 0) && (slots_needed <= 0)) {
431 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
434 if (!afs_CacheIsTooFull())
435 afs_CacheTooFull = 0;
436 } /* end of cache cleanup */
437 ReleaseWriteLock(&afs_xdcache);
440 * This is a defensive check to try to avoid starving threads
441 * that may need the global lock so thay can help free some
442 * cache space. If this thread won't be sleeping or truncating
443 * any cache files then give up the global lock so other
444 * threads get a chance to run.
446 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
447 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
448 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
452 * This is where we free the discarded cache elements.
454 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
455 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
456 afs_FreeDiscardedDCache();
459 /* See if we need to continue to run. Someone may have
460 * signalled us while we were executing.
462 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
463 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
464 /* Collect statistics on truncate daemon. */
465 CTD_stats.CTD_nSleeps++;
466 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
467 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
468 CTD_stats.CTD_beforeSleep);
469 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
471 afs_TruncateDaemonRunning = 0;
472 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
473 afs_TruncateDaemonRunning = 1;
475 osi_GetuTime(&CTD_stats.CTD_afterSleep);
476 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
477 CTD_stats.CTD_afterSleep);
478 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
480 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
481 afs_termState = AFSOP_STOP_AFSDB;
482 afs_osi_Wakeup(&afs_termState);
490 * Make adjustment for the new size in the disk cache entry
492 * \note Major Assumptions Here:
493 * Assumes that frag size is an integral power of two, less one,
494 * and that this is a two's complement machine. I don't
495 * know of any filesystems which violate this assumption...
497 * \param adc Ptr to dcache entry.
498 * \param anewsize New size desired.
503 afs_AdjustSize(struct dcache *adc, afs_int32 newSize)
507 AFS_STATCNT(afs_AdjustSize);
509 adc->dflags |= DFEntryMod;
510 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
511 adc->f.chunkBytes = newSize;
514 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
515 afs_DCAdjustSize(adc, oldSize, newSize);
516 if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {
518 /* We're growing the file, wakeup the daemon */
519 afs_MaybeWakeupTruncateDaemon();
521 afs_blocksUsed += (newSize - oldSize);
522 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
527 * This routine is responsible for moving at least one entry (but up
528 * to some number of them) from the LRU queue to the free queue.
530 * \param anumber Number of entries that should ideally be moved.
531 * \param aneedSpace How much space we need (1K blocks);
534 * The anumber parameter is just a hint; at least one entry MUST be
535 * moved, or we'll panic. We must be called with afs_xdcache
536 * write-locked. We should try to satisfy both anumber and aneedspace,
537 * whichever is more demanding - need to do several things:
538 * 1. only grab up to anumber victims if aneedSpace <= 0, not
539 * the whole set of MAXATONCE.
540 * 2. dynamically choose MAXATONCE to reflect severity of
541 * demand: something like (*aneedSpace >> (logChunk - 9))
543 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
544 * indicates that the cache is not properly configured/tuned or
545 * something. We should be able to automatically correct that problem.
548 #define MAXATONCE 16 /* max we can obtain at once */
550 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
554 struct VenusFid *afid;
559 afs_uint32 victims[MAXATONCE];
560 struct dcache *victimDCs[MAXATONCE];
561 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
562 afs_uint32 victimPtr; /* next free item in victim arrays */
563 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
564 afs_uint32 maxVictimPtr; /* where it is */
568 AFS_STATCNT(afs_GetDownD);
570 if (CheckLock(&afs_xdcache) != -1)
571 osi_Panic("getdownd nolock");
572 /* decrement anumber first for all dudes in free list */
573 /* SHOULD always decrement anumber first, even if aneedSpace >0,
574 * because we should try to free space even if anumber <=0 */
575 if (!aneedSpace || *aneedSpace <= 0) {
576 anumber -= afs_freeDCCount;
578 return; /* enough already free */
582 /* bounds check parameter */
583 if (anumber > MAXATONCE)
584 anumber = MAXATONCE; /* all we can do */
586 /* rewrite so phases include a better eligiblity for gc test*/
588 * The phase variable manages reclaims. Set to 0, the first pass,
589 * we don't reclaim active entries, or other than target bucket.
590 * Set to 1, we reclaim even active ones in target bucket.
591 * Set to 2, we reclaim any inactive one.
592 * Set to 3, we reclaim even active ones. On Solaris, we also reclaim
593 * entries whose corresponding vcache has a nonempty multiPage list, when
602 for (i = 0; i < afs_cacheFiles; i++)
603 /* turn off all flags */
604 afs_indexFlags[i] &= ~IFFlag;
606 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
607 /* find oldest entries for reclamation */
608 maxVictimPtr = victimPtr = 0;
609 hzero(maxVictimTime);
610 curbucket = afs_DCWhichBucket(phase, buckethint);
611 /* select victims from access time array */
612 for (i = 0; i < afs_cacheFiles; i++) {
613 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
614 /* skip if dirty or already free */
617 tdc = afs_indexTable[i];
618 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
620 /* Wrong bucket; can't use it! */
623 if (tdc && (tdc->refCount != 0)) {
624 /* Referenced; can't use it! */
627 hset(vtime, afs_indexTimes[i]);
629 /* if we've already looked at this one, skip it */
630 if (afs_indexFlags[i] & IFFlag)
633 if (victimPtr < MAXATONCE) {
634 /* if there's at least one free victim slot left */
635 victims[victimPtr] = i;
636 hset(victimTimes[victimPtr], vtime);
637 if (hcmp(vtime, maxVictimTime) > 0) {
638 hset(maxVictimTime, vtime);
639 maxVictimPtr = victimPtr;
642 } else if (hcmp(vtime, maxVictimTime) < 0) {
644 * We're older than youngest victim, so we replace at
647 /* find youngest (largest LRU) victim */
650 osi_Panic("getdownd local");
652 hset(victimTimes[j], vtime);
653 /* recompute maxVictimTime */
654 hset(maxVictimTime, vtime);
655 for (j = 0; j < victimPtr; j++)
656 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
657 hset(maxVictimTime, victimTimes[j]);
663 /* now really reclaim the victims */
664 j = 0; /* flag to track if we actually got any of the victims */
665 /* first, hold all the victims, since we're going to release the lock
666 * during the truncate operation.
668 for (i = 0; i < victimPtr; i++) {
669 tdc = afs_GetDSlot(victims[i], 0);
670 /* We got tdc->tlock(R) here */
671 if (tdc->refCount == 1)
675 ReleaseReadLock(&tdc->tlock);
679 for (i = 0; i < victimPtr; i++) {
680 /* q is first elt in dcache entry */
682 /* now, since we're dropping the afs_xdcache lock below, we
683 * have to verify, before proceeding, that there are no other
684 * references to this dcache entry, even now. Note that we
685 * compare with 1, since we bumped it above when we called
686 * afs_GetDSlot to preserve the entry's identity.
688 if (tdc && tdc->refCount == 1) {
689 unsigned char chunkFlags;
690 afs_size_t tchunkoffset = 0;
692 /* xdcache is lower than the xvcache lock */
693 ReleaseWriteLock(&afs_xdcache);
694 ObtainReadLock(&afs_xvcache);
695 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
696 ReleaseReadLock(&afs_xvcache);
697 ObtainWriteLock(&afs_xdcache, 527);
699 if (tdc->refCount > 1)
702 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
703 chunkFlags = afs_indexFlags[tdc->index];
704 if (((phase & 1) == 0) && osi_Active(tvc))
706 if (((phase & 1) == 1) && osi_Active(tvc)
707 && (tvc->f.states & CDCLock)
708 && (chunkFlags & IFAnyPages))
710 if (chunkFlags & IFDataMod)
712 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
713 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
714 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
715 ICL_HANDLE_OFFSET(tchunkoffset));
717 #if defined(AFS_SUN5_ENV)
719 * Now we try to invalidate pages. We do this only for
720 * Solaris. For other platforms, it's OK to recycle a
721 * dcache entry out from under a page, because the strategy
722 * function can call afs_GetDCache().
724 if (!skip && (chunkFlags & IFAnyPages)) {
727 ReleaseWriteLock(&afs_xdcache);
728 ObtainWriteLock(&tvc->vlock, 543);
729 if (!QEmpty(&tvc->multiPage)) {
730 if (phase < 3 || osi_VM_MultiPageConflict(tvc, tdc)) {
735 /* block locking pages */
736 tvc->vstates |= VPageCleaning;
737 /* block getting new pages */
739 ReleaseWriteLock(&tvc->vlock);
740 /* One last recheck */
741 ObtainWriteLock(&afs_xdcache, 333);
742 chunkFlags = afs_indexFlags[tdc->index];
743 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
744 || (osi_Active(tvc) && (tvc->f.states & CDCLock)
745 && (chunkFlags & IFAnyPages))) {
747 ReleaseWriteLock(&afs_xdcache);
750 ReleaseWriteLock(&afs_xdcache);
752 code = osi_VM_GetDownD(tvc, tdc);
754 ObtainWriteLock(&afs_xdcache, 269);
755 /* we actually removed all pages, clean and dirty */
757 afs_indexFlags[tdc->index] &=
758 ~(IFDirtyPages | IFAnyPages);
761 ReleaseWriteLock(&afs_xdcache);
763 ObtainWriteLock(&tvc->vlock, 544);
764 if (--tvc->activeV == 0
765 && (tvc->vstates & VRevokeWait)) {
766 tvc->vstates &= ~VRevokeWait;
767 afs_osi_Wakeup((char *)&tvc->vstates);
770 if (tvc->vstates & VPageCleaning) {
771 tvc->vstates &= ~VPageCleaning;
772 afs_osi_Wakeup((char *)&tvc->vstates);
775 ReleaseWriteLock(&tvc->vlock);
777 #endif /* AFS_SUN5_ENV */
779 ReleaseWriteLock(&afs_xdcache);
782 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
783 ObtainWriteLock(&afs_xdcache, 528);
784 if (afs_indexFlags[tdc->index] &
785 (IFDataMod | IFDirtyPages | IFAnyPages))
787 if (tdc->refCount > 1)
790 #if defined(AFS_SUN5_ENV)
792 /* no vnode, so IFDirtyPages is spurious (we don't
793 * sweep dcaches on vnode recycling, so we can have
794 * DIRTYPAGES set even when all pages are gone). Just
796 * Hold vcache lock to prevent vnode from being
797 * created while we're clearing IFDirtyPages.
799 afs_indexFlags[tdc->index] &=
800 ~(IFDirtyPages | IFAnyPages);
804 /* skip this guy and mark him as recently used */
805 afs_indexFlags[tdc->index] |= IFFlag;
806 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
807 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
808 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
809 ICL_HANDLE_OFFSET(tchunkoffset));
811 /* flush this dude from the data cache and reclaim;
812 * first, make sure no one will care that we damage
813 * it, by removing it from all hash tables. Then,
814 * melt it down for parts. Note that any concurrent
815 * (new possibility!) calls to GetDownD won't touch
816 * this guy because his reference count is > 0. */
817 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
818 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
819 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
820 ICL_HANDLE_OFFSET(tchunkoffset));
821 AFS_STATCNT(afs_gget);
822 afs_HashOutDCache(tdc, 1);
823 if (tdc->f.chunkBytes != 0) {
827 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
832 afs_DiscardDCache(tdc);
837 j = 1; /* we reclaimed at least one victim */
841 } /* end of for victims loop */
844 /* Phase is 0 and no one was found, so try phase 1 (ignore
845 * osi_Active flag) */
848 for (i = 0; i < afs_cacheFiles; i++)
849 /* turn off all flags */
850 afs_indexFlags[i] &= ~IFFlag;
853 /* found no one in phases 0-5, we're hosed */
857 } /* big while loop */
865 * Remove adc from any hash tables that would allow it to be located
866 * again by afs_FindDCache or afs_GetDCache.
868 * \param adc Pointer to dcache entry to remove from hash tables.
870 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
874 afs_HashOutDCache(struct dcache *adc, int zap)
878 AFS_STATCNT(afs_glink);
880 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
882 /* if this guy is in the hash table, pull him out */
883 if (adc->f.fid.Fid.Volume != 0) {
884 /* remove entry from first hash chains */
885 i = DCHash(&adc->f.fid, adc->f.chunk);
886 us = afs_dchashTbl[i];
887 if (us == adc->index) {
888 /* first dude in the list */
889 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
891 /* somewhere on the chain */
892 while (us != NULLIDX) {
893 if (afs_dcnextTbl[us] == adc->index) {
894 /* found item pointing at the one to delete */
895 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
898 us = afs_dcnextTbl[us];
901 osi_Panic("dcache hc");
903 /* remove entry from *other* hash chain */
904 i = DVHash(&adc->f.fid);
905 us = afs_dvhashTbl[i];
906 if (us == adc->index) {
907 /* first dude in the list */
908 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
910 /* somewhere on the chain */
911 while (us != NULLIDX) {
912 if (afs_dvnextTbl[us] == adc->index) {
913 /* found item pointing at the one to delete */
914 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
917 us = afs_dvnextTbl[us];
920 osi_Panic("dcache hv");
925 /* prevent entry from being found on a reboot (it is already out of
926 * the hash table, but after a crash, we just look at fid fields of
927 * stable (old) entries).
929 adc->f.fid.Fid.Volume = 0; /* invalid */
931 /* mark entry as modified */
932 adc->dflags |= DFEntryMod;
937 } /*afs_HashOutDCache */
940 * Flush the given dcache entry, pulling it from hash chains
941 * and truncating the associated cache file.
943 * \param adc Ptr to dcache entry to flush.
946 * This routine must be called with the afs_xdcache lock held
950 afs_FlushDCache(struct dcache *adc)
952 AFS_STATCNT(afs_FlushDCache);
954 * Bump the number of cache files flushed.
956 afs_stats_cmperf.cacheFlushes++;
958 /* remove from all hash tables */
959 afs_HashOutDCache(adc, 1);
961 /* Free its space; special case null operation, since truncate operation
962 * in UFS is slow even in this case, and this allows us to pre-truncate
963 * these files at more convenient times with fewer locks set
964 * (see afs_GetDownD).
966 if (adc->f.chunkBytes != 0) {
967 afs_DiscardDCache(adc);
968 afs_MaybeWakeupTruncateDaemon();
973 if (afs_WaitForCacheDrain) {
974 if (afs_blocksUsed <=
975 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
976 afs_WaitForCacheDrain = 0;
977 afs_osi_Wakeup(&afs_WaitForCacheDrain);
980 } /*afs_FlushDCache */
984 * Put a dcache entry on the free dcache entry list.
986 * \param adc dcache entry to free.
988 * \note Environment: called with afs_xdcache lock write-locked.
991 afs_FreeDCache(struct dcache *adc)
993 /* Thread on free list, update free list count and mark entry as
994 * freed in its indexFlags element. Also, ensure DCache entry gets
995 * written out (set DFEntryMod).
998 afs_dvnextTbl[adc->index] = afs_freeDCList;
999 afs_freeDCList = adc->index;
1001 afs_indexFlags[adc->index] |= IFFree;
1002 adc->dflags |= DFEntryMod;
1004 if (afs_WaitForCacheDrain) {
1005 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1006 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1007 afs_WaitForCacheDrain = 0;
1008 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1011 } /* afs_FreeDCache */
1014 * Discard the cache element by moving it to the discardDCList.
1015 * This puts the cache element into a quasi-freed state, where
1016 * the space may be reused, but the file has not been truncated.
1018 * \note Major Assumptions Here:
1019 * Assumes that frag size is an integral power of two, less one,
1020 * and that this is a two's complement machine. I don't
1021 * know of any filesystems which violate this assumption...
1023 * \param adr Ptr to dcache entry.
1025 * \note Environment:
1026 * Must be called with afs_xdcache write-locked.
1030 afs_DiscardDCache(struct dcache *adc)
1034 AFS_STATCNT(afs_DiscardDCache);
1036 osi_Assert(adc->refCount == 1);
1038 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1039 afs_blocksDiscarded += size;
1040 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1042 afs_dvnextTbl[adc->index] = afs_discardDCList;
1043 afs_discardDCList = adc->index;
1044 afs_discardDCCount++;
1046 adc->f.fid.Fid.Volume = 0;
1047 adc->dflags |= DFEntryMod;
1048 afs_indexFlags[adc->index] |= IFDiscarded;
1050 if (afs_WaitForCacheDrain) {
1051 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1052 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1053 afs_WaitForCacheDrain = 0;
1054 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1058 } /*afs_DiscardDCache */
1061 * Free the next element on the list of discarded cache elements.
1064 afs_FreeDiscardedDCache(void)
1067 struct osi_file *tfile;
1070 AFS_STATCNT(afs_FreeDiscardedDCache);
1072 ObtainWriteLock(&afs_xdcache, 510);
1073 if (!afs_blocksDiscarded) {
1074 ReleaseWriteLock(&afs_xdcache);
1079 * Get an entry from the list of discarded cache elements
1081 tdc = afs_GetDSlot(afs_discardDCList, 0);
1082 osi_Assert(tdc->refCount == 1);
1083 ReleaseReadLock(&tdc->tlock);
1085 afs_discardDCList = afs_dvnextTbl[tdc->index];
1086 afs_dvnextTbl[tdc->index] = NULLIDX;
1087 afs_discardDCCount--;
1088 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1089 afs_blocksDiscarded -= size;
1090 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1091 /* We can lock because we just took it off the free list */
1092 ObtainWriteLock(&tdc->lock, 626);
1093 ReleaseWriteLock(&afs_xdcache);
1096 * Truncate the element to reclaim its space
1098 tfile = afs_CFileOpen(&tdc->f.inode);
1099 afs_CFileTruncate(tfile, 0);
1100 afs_CFileClose(tfile);
1101 afs_AdjustSize(tdc, 0);
1102 afs_DCMoveBucket(tdc, 0, 0);
1105 * Free the element we just truncated
1107 ObtainWriteLock(&afs_xdcache, 511);
1108 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1109 afs_FreeDCache(tdc);
1110 tdc->f.states &= ~(DRO|DBackup|DRW);
1111 ReleaseWriteLock(&tdc->lock);
1113 ReleaseWriteLock(&afs_xdcache);
1117 * Free as many entries from the list of discarded cache elements
1118 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1123 afs_MaybeFreeDiscardedDCache(void)
1126 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1128 while (afs_blocksDiscarded
1129 && (afs_blocksUsed >
1130 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1131 afs_FreeDiscardedDCache();
1137 * Try to free up a certain number of disk slots.
1139 * \param anumber Targeted number of disk slots to free up.
1141 * \note Environment:
1142 * Must be called with afs_xdcache write-locked.
1146 afs_GetDownDSlot(int anumber)
1148 struct afs_q *tq, *nq;
1153 AFS_STATCNT(afs_GetDownDSlot);
1154 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1155 osi_Panic("diskless getdowndslot");
1157 if (CheckLock(&afs_xdcache) != -1)
1158 osi_Panic("getdowndslot nolock");
1160 /* decrement anumber first for all dudes in free list */
1161 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1164 return; /* enough already free */
1166 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1168 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1169 nq = QPrev(tq); /* in case we remove it */
1170 if (tdc->refCount == 0) {
1171 if ((ix = tdc->index) == NULLIDX)
1172 osi_Panic("getdowndslot");
1173 /* pull the entry out of the lruq and put it on the free list */
1174 QRemove(&tdc->lruq);
1176 /* write-through if modified */
1177 if (tdc->dflags & DFEntryMod) {
1178 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1180 * ask proxy to do this for us - we don't have the stack space
1182 while (tdc->dflags & DFEntryMod) {
1185 s = SPLOCK(afs_sgibklock);
1186 if (afs_sgibklist == NULL) {
1187 /* if slot is free, grab it. */
1188 afs_sgibklist = tdc;
1189 SV_SIGNAL(&afs_sgibksync);
1191 /* wait for daemon to (start, then) finish. */
1192 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1196 tdc->dflags &= ~DFEntryMod;
1197 afs_WriteDCache(tdc, 1);
1201 /* finally put the entry in the free list */
1202 afs_indexTable[ix] = NULL;
1203 afs_indexFlags[ix] &= ~IFEverUsed;
1204 tdc->index = NULLIDX;
1205 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1206 afs_freeDSList = tdc;
1210 } /*afs_GetDownDSlot */
1217 * Increment the reference count on a disk cache entry,
1218 * which already has a non-zero refcount. In order to
1219 * increment the refcount of a zero-reference entry, you
1220 * have to hold afs_xdcache.
1223 * adc : Pointer to the dcache entry to increment.
1226 * Nothing interesting.
1229 afs_RefDCache(struct dcache *adc)
1231 ObtainWriteLock(&adc->tlock, 627);
1232 if (adc->refCount < 0)
1233 osi_Panic("RefDCache: negative refcount");
1235 ReleaseWriteLock(&adc->tlock);
1244 * Decrement the reference count on a disk cache entry.
1247 * ad : Ptr to the dcache entry to decrement.
1250 * Nothing interesting.
1253 afs_PutDCache(struct dcache *adc)
1255 AFS_STATCNT(afs_PutDCache);
1256 ObtainWriteLock(&adc->tlock, 276);
1257 if (adc->refCount <= 0)
1258 osi_Panic("putdcache");
1260 ReleaseWriteLock(&adc->tlock);
1269 * Try to discard all data associated with this file from the
1273 * avc : Pointer to the cache info for the file.
1276 * Both pvnLock and lock are write held.
1279 afs_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
1284 AFS_STATCNT(afs_TryToSmush);
1285 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1286 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
1287 sync = 1; /* XX Temp testing XX */
1289 #if defined(AFS_SUN5_ENV)
1290 ObtainWriteLock(&avc->vlock, 573);
1291 avc->activeV++; /* block new getpages */
1292 ReleaseWriteLock(&avc->vlock);
1295 /* Flush VM pages */
1296 osi_VM_TryToSmush(avc, acred, sync);
1299 * Get the hash chain containing all dce's for this fid
1301 i = DVHash(&avc->f.fid);
1302 ObtainWriteLock(&afs_xdcache, 277);
1303 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1304 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1305 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1306 int releaseTlock = 1;
1307 tdc = afs_GetDSlot(index, NULL);
1308 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1310 if ((afs_indexFlags[index] & IFDataMod) == 0
1311 && tdc->refCount == 1) {
1312 ReleaseReadLock(&tdc->tlock);
1314 afs_FlushDCache(tdc);
1317 afs_indexTable[index] = 0;
1320 ReleaseReadLock(&tdc->tlock);
1324 #if defined(AFS_SUN5_ENV)
1325 ObtainWriteLock(&avc->vlock, 545);
1326 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1327 avc->vstates &= ~VRevokeWait;
1328 afs_osi_Wakeup((char *)&avc->vstates);
1330 ReleaseWriteLock(&avc->vlock);
1332 ReleaseWriteLock(&afs_xdcache);
1334 * It's treated like a callback so that when we do lookups we'll
1335 * invalidate the unique bit if any
1336 * trytoSmush occured during the lookup call
1342 * afs_DCacheMissingChunks
1345 * Given the cached info for a file, return the number of chunks that
1346 * are not available from the dcache.
1349 * avc: Pointer to the (held) vcache entry to look in.
1352 * The number of chunks which are not currently cached.
1355 * The vcache entry is held upon entry.
1359 afs_DCacheMissingChunks(struct vcache *avc)
1362 afs_size_t totalLength = 0;
1363 afs_uint32 totalChunks = 0;
1366 totalLength = avc->f.m.Length;
1367 if (avc->f.truncPos < totalLength)
1368 totalLength = avc->f.truncPos;
1370 /* Length is 0, no chunk missing. */
1371 if (totalLength == 0)
1374 /* If totalLength is a multiple of chunksize, the last byte appears
1375 * as being part of the next chunk, which does not exist.
1376 * Decrementing totalLength by one fixes that.
1379 totalChunks = (AFS_CHUNK(totalLength) + 1);
1381 /* If we're a directory, we only ever have one chunk, regardless of
1382 * the size of the dir.
1384 if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
1388 printf("Should have %d chunks for %u bytes\n",
1389 totalChunks, (totalLength + 1));
1391 i = DVHash(&avc->f.fid);
1392 ObtainWriteLock(&afs_xdcache, 1001);
1393 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1394 i = afs_dvnextTbl[index];
1395 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1396 tdc = afs_GetDSlot(index, NULL);
1397 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1400 ReleaseReadLock(&tdc->tlock);
1404 ReleaseWriteLock(&afs_xdcache);
1406 /*printf("Missing %d chunks\n", totalChunks);*/
1408 return (totalChunks);
1415 * Given the cached info for a file and a byte offset into the
1416 * file, make sure the dcache entry for that file and containing
1417 * the given byte is available, returning it to our caller.
1420 * avc : Pointer to the (held) vcache entry to look in.
1421 * abyte : Which byte we want to get to.
1424 * Pointer to the dcache entry covering the file & desired byte,
1425 * or NULL if not found.
1428 * The vcache entry is held upon entry.
1432 afs_FindDCache(struct vcache *avc, afs_size_t abyte)
1436 struct dcache *tdc = NULL;
1438 AFS_STATCNT(afs_FindDCache);
1439 chunk = AFS_CHUNK(abyte);
1442 * Hash on the [fid, chunk] and get the corresponding dcache index
1443 * after write-locking the dcache.
1445 i = DCHash(&avc->f.fid, chunk);
1446 ObtainWriteLock(&afs_xdcache, 278);
1447 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1448 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1449 tdc = afs_GetDSlot(index, NULL);
1450 ReleaseReadLock(&tdc->tlock);
1451 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1452 break; /* leaving refCount high for caller */
1456 index = afs_dcnextTbl[index];
1458 if (index != NULLIDX) {
1459 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1460 hadd32(afs_indexCounter, 1);
1461 ReleaseWriteLock(&afs_xdcache);
1464 ReleaseWriteLock(&afs_xdcache);
1466 } /*afs_FindDCache */
1470 * Get a fresh dcache from the free or discarded list.
1472 * \param avc Who's dcache is this going to be?
1473 * \param chunk The position where it will be placed in.
1474 * \param lock How are locks held.
1475 * \param ashFid If this dcache going to be used for a shadow dir,
1478 * \note Required locks:
1480 * - avc (R if (lock & 1) set and W otherwise)
1481 * \note It write locks the new dcache. The caller must unlock it.
1483 * \return The new dcache.
1486 afs_AllocDCache(struct vcache *avc, afs_int32 chunk, afs_int32 lock,
1487 struct VenusFid *ashFid)
1489 struct dcache *tdc = NULL;
1490 afs_uint32 size = 0;
1491 struct osi_file *file;
1493 if (afs_discardDCList == NULLIDX
1494 || ((lock & 2) && afs_freeDCList != NULLIDX)) {
1496 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1497 tdc = afs_GetDSlot(afs_freeDCList, 0);
1498 osi_Assert(tdc->refCount == 1);
1499 ReleaseReadLock(&tdc->tlock);
1500 ObtainWriteLock(&tdc->lock, 604);
1501 afs_freeDCList = afs_dvnextTbl[tdc->index];
1504 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1505 tdc = afs_GetDSlot(afs_discardDCList, 0);
1506 osi_Assert(tdc->refCount == 1);
1507 ReleaseReadLock(&tdc->tlock);
1508 ObtainWriteLock(&tdc->lock, 605);
1509 afs_discardDCList = afs_dvnextTbl[tdc->index];
1510 afs_discardDCCount--;
1512 ((tdc->f.chunkBytes +
1513 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1514 tdc->f.states &= ~(DRO|DBackup|DRW);
1515 afs_DCMoveBucket(tdc, size, 0);
1516 afs_blocksDiscarded -= size;
1517 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1519 /* Truncate the chunk so zeroes get filled properly */
1520 file = afs_CFileOpen(&tdc->f.inode);
1521 afs_CFileTruncate(file, 0);
1522 afs_CFileClose(file);
1523 afs_AdjustSize(tdc, 0);
1529 * avc->lock(R) if setLocks
1530 * avc->lock(W) if !setLocks
1536 * Fill in the newly-allocated dcache record.
1538 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1540 /* Use shadow fid if provided. */
1541 tdc->f.fid = *ashFid;
1543 /* Use normal vcache's fid otherwise. */
1544 tdc->f.fid = avc->f.fid;
1545 if (avc->f.states & CRO)
1546 tdc->f.states = DRO;
1547 else if (avc->f.states & CBackup)
1548 tdc->f.states = DBackup;
1550 tdc->f.states = DRW;
1551 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1552 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1554 hones(tdc->f.versionNo); /* invalid value */
1555 tdc->f.chunk = chunk;
1556 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1558 if (tdc->lruq.prev == &tdc->lruq)
1559 osi_Panic("lruq 1");
1568 * This function is called to obtain a reference to data stored in
1569 * the disk cache, locating a chunk of data containing the desired
1570 * byte and returning a reference to the disk cache entry, with its
1571 * reference count incremented.
1575 * avc : Ptr to a vcache entry (unlocked)
1576 * abyte : Byte position in the file desired
1577 * areq : Request structure identifying the requesting user.
1578 * aflags : Settings as follows:
1580 * 2 : Return after creating entry.
1581 * 4 : called from afs_vnop_write.c
1582 * *alen contains length of data to be written.
1584 * aoffset : Set to the offset within the chunk where the resident
1586 * alen : Set to the number of bytes of data after the desired
1587 * byte (including the byte itself) which can be read
1591 * The vcache entry pointed to by avc is unlocked upon entry.
1595 * Update the vnode-to-dcache hint if we can get the vnode lock
1596 * right away. Assumes dcache entry is at least read-locked.
1599 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1601 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1602 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1605 ReleaseWriteLock(&v->lock);
1609 /* avc - Write-locked unless aflags & 1 */
1611 afs_GetDCache(struct vcache *avc, afs_size_t abyte,
1612 struct vrequest *areq, afs_size_t * aoffset,
1613 afs_size_t * alen, int aflags)
1615 afs_int32 i, code, shortcut;
1616 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1617 afs_int32 adjustsize = 0;
1623 afs_size_t maxGoodLength; /* amount of good data at server */
1624 afs_size_t Position = 0;
1625 afs_int32 size, tlen; /* size of segment to transfer */
1626 struct afs_FetchOutput *tsmall = 0;
1628 struct osi_file *file;
1629 struct afs_conn *tc;
1631 struct server *newCallback = NULL;
1632 char setNewCallback;
1633 char setVcacheStatus;
1634 char doVcacheUpdate;
1636 int doAdjustSize = 0;
1637 int doReallyAdjustSize = 0;
1638 int overWriteWholeChunk = 0;
1639 struct rx_connection *rxconn;
1642 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1643 int fromReplica; /*Are we reading from a replica? */
1644 int numFetchLoops; /*# times around the fetch/analyze loop */
1645 #endif /* AFS_NOSTATS */
1647 AFS_STATCNT(afs_GetDCache);
1651 setLocks = aflags & 1;
1654 * Determine the chunk number and offset within the chunk corresponding
1655 * to the desired byte.
1657 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1660 chunk = AFS_CHUNK(abyte);
1663 /* come back to here if we waited for the cache to drain. */
1666 setNewCallback = setVcacheStatus = 0;
1670 ObtainWriteLock(&avc->lock, 616);
1672 ObtainReadLock(&avc->lock);
1677 * avc->lock(R) if setLocks && !slowPass
1678 * avc->lock(W) if !setLocks || slowPass
1683 /* check hints first! (might could use bcmp or some such...) */
1684 if ((tdc = avc->dchint)) {
1688 * The locking order between afs_xdcache and dcache lock matters.
1689 * The hint dcache entry could be anywhere, even on the free list.
1690 * Locking afs_xdcache ensures that noone is trying to pull dcache
1691 * entries from the free list, and thereby assuming them to be not
1692 * referenced and not locked.
1694 ObtainReadLock(&afs_xdcache);
1695 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1697 if (dcLocked && (tdc->index != NULLIDX)
1698 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1699 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1700 /* got the right one. It might not be the right version, and it
1701 * might be fetching, but it's the right dcache entry.
1703 /* All this code should be integrated better with what follows:
1704 * I can save a good bit more time under a write lock if I do..
1706 ObtainWriteLock(&tdc->tlock, 603);
1708 ReleaseWriteLock(&tdc->tlock);
1710 ReleaseReadLock(&afs_xdcache);
1713 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1714 && !(tdc->dflags & DFFetching)) {
1716 afs_stats_cmperf.dcacheHits++;
1717 ObtainWriteLock(&afs_xdcache, 559);
1718 QRemove(&tdc->lruq);
1719 QAdd(&afs_DLRU, &tdc->lruq);
1720 ReleaseWriteLock(&afs_xdcache);
1723 * avc->lock(R) if setLocks && !slowPass
1724 * avc->lock(W) if !setLocks || slowPass
1731 ReleaseSharedLock(&tdc->lock);
1732 ReleaseReadLock(&afs_xdcache);
1740 * avc->lock(R) if setLocks && !slowPass
1741 * avc->lock(W) if !setLocks || slowPass
1742 * tdc->lock(S) if tdc
1745 if (!tdc) { /* If the hint wasn't the right dcache entry */
1747 * Hash on the [fid, chunk] and get the corresponding dcache index
1748 * after write-locking the dcache.
1753 * avc->lock(R) if setLocks && !slowPass
1754 * avc->lock(W) if !setLocks || slowPass
1757 i = DCHash(&avc->f.fid, chunk);
1758 /* check to make sure our space is fine */
1759 afs_MaybeWakeupTruncateDaemon();
1761 ObtainWriteLock(&afs_xdcache, 280);
1763 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1764 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1765 tdc = afs_GetDSlot(index, NULL);
1766 ReleaseReadLock(&tdc->tlock);
1769 * avc->lock(R) if setLocks && !slowPass
1770 * avc->lock(W) if !setLocks || slowPass
1773 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1774 /* Move it up in the beginning of the list */
1775 if (afs_dchashTbl[i] != index) {
1776 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1777 afs_dcnextTbl[index] = afs_dchashTbl[i];
1778 afs_dchashTbl[i] = index;
1780 ReleaseWriteLock(&afs_xdcache);
1781 ObtainSharedLock(&tdc->lock, 606);
1782 break; /* leaving refCount high for caller */
1788 index = afs_dcnextTbl[index];
1792 * If we didn't find the entry, we'll create one.
1794 if (index == NULLIDX) {
1797 * avc->lock(R) if setLocks
1798 * avc->lock(W) if !setLocks
1801 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1802 avc, ICL_TYPE_INT32, chunk);
1804 /* Make sure there is a free dcache entry for us to use */
1805 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1808 avc->f.states |= CDCLock;
1809 /* just need slots */
1810 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1812 avc->f.states &= ~CDCLock;
1813 if (afs_discardDCList != NULLIDX
1814 || afs_freeDCList != NULLIDX)
1816 /* If we can't get space for 5 mins we give up and panic */
1817 if (++downDCount > 300) {
1818 osi_Panic("getdcache");
1820 ReleaseWriteLock(&afs_xdcache);
1823 * avc->lock(R) if setLocks
1824 * avc->lock(W) if !setLocks
1826 afs_osi_Wait(1000, 0, 0);
1831 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1834 * Now add to the two hash chains - note that i is still set
1835 * from the above DCHash call.
1837 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1838 afs_dchashTbl[i] = tdc->index;
1839 i = DVHash(&avc->f.fid);
1840 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1841 afs_dvhashTbl[i] = tdc->index;
1842 tdc->dflags = DFEntryMod;
1844 afs_MaybeWakeupTruncateDaemon();
1845 ReleaseWriteLock(&afs_xdcache);
1846 ConvertWToSLock(&tdc->lock);
1851 /* vcache->dcache hint failed */
1854 * avc->lock(R) if setLocks && !slowPass
1855 * avc->lock(W) if !setLocks || slowPass
1858 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1859 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1860 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1861 hgetlo(avc->f.m.DataVersion));
1863 * Here we have the entry in tdc, with its refCount incremented.
1864 * Note: we don't use the S-lock on avc; it costs concurrency when
1865 * storing a file back to the server.
1869 * Not a newly created file so we need to check the file's length and
1870 * compare data versions since someone could have changed the data or we're
1871 * reading a file written elsewhere. We only want to bypass doing no-op
1872 * read rpcs on newly created files (dv of 0) since only then we guarantee
1873 * that this chunk's data hasn't been filled by another client.
1875 size = AFS_CHUNKSIZE(abyte);
1876 if (aflags & 4) /* called from write */
1878 else /* called from read */
1879 tlen = tdc->validPos - abyte;
1880 Position = AFS_CHUNKTOBASE(chunk);
1881 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
1882 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
1883 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
1884 ICL_HANDLE_OFFSET(Position));
1885 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
1887 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
1888 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
1889 overWriteWholeChunk = 1;
1890 if (doAdjustSize || overWriteWholeChunk) {
1891 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1893 #ifdef AFS_SGI64_ENV
1896 #else /* AFS_SGI64_ENV */
1899 #endif /* AFS_SGI64_ENV */
1900 #else /* AFS_SGI_ENV */
1903 #endif /* AFS_SGI_ENV */
1904 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
1905 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1906 #if defined(AFS_SUN5_ENV)
1907 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
1909 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
1911 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
1912 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
1913 doReallyAdjustSize = 1;
1915 if (doReallyAdjustSize || overWriteWholeChunk) {
1916 /* no data in file to read at this position */
1917 UpgradeSToWLock(&tdc->lock, 607);
1918 file = afs_CFileOpen(&tdc->f.inode);
1919 afs_CFileTruncate(file, 0);
1920 afs_CFileClose(file);
1921 afs_AdjustSize(tdc, 0);
1922 hset(tdc->f.versionNo, avc->f.m.DataVersion);
1923 tdc->dflags |= DFEntryMod;
1925 ConvertWToSLock(&tdc->lock);
1930 * We must read in the whole chunk if the version number doesn't
1934 /* don't need data, just a unique dcache entry */
1935 ObtainWriteLock(&afs_xdcache, 608);
1936 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1937 hadd32(afs_indexCounter, 1);
1938 ReleaseWriteLock(&afs_xdcache);
1940 updateV2DC(setLocks, avc, tdc, 553);
1941 if (vType(avc) == VDIR)
1944 *aoffset = AFS_CHUNKOFFSET(abyte);
1945 if (tdc->validPos < abyte)
1946 *alen = (afs_size_t) 0;
1948 *alen = tdc->validPos - abyte;
1949 ReleaseSharedLock(&tdc->lock);
1952 ReleaseWriteLock(&avc->lock);
1954 ReleaseReadLock(&avc->lock);
1956 return tdc; /* check if we're done */
1961 * avc->lock(R) if setLocks && !slowPass
1962 * avc->lock(W) if !setLocks || slowPass
1965 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
1967 setNewCallback = setVcacheStatus = 0;
1971 * avc->lock(R) if setLocks && !slowPass
1972 * avc->lock(W) if !setLocks || slowPass
1975 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
1977 * Version number mismatch.
1980 * If we are disconnected, then we can't do much of anything
1981 * because the data doesn't match the file.
1983 if (AFS_IS_DISCONNECTED) {
1984 ReleaseSharedLock(&tdc->lock);
1987 ReleaseWriteLock(&avc->lock);
1989 ReleaseReadLock(&avc->lock);
1991 /* Flush the Dcache */
1996 UpgradeSToWLock(&tdc->lock, 609);
1999 * If data ever existed for this vnode, and this is a text object,
2000 * do some clearing. Now, you'd think you need only do the flush
2001 * when VTEXT is on, but VTEXT is turned off when the text object
2002 * is freed, while pages are left lying around in memory marked
2003 * with this vnode. If we would reactivate (create a new text
2004 * object from) this vnode, we could easily stumble upon some of
2005 * these old pages in pagein. So, we always flush these guys.
2006 * Sun has a wonderful lack of useful invariants in this system.
2008 * avc->flushDV is the data version # of the file at the last text
2009 * flush. Clearly, at least, we don't have to flush the file more
2010 * often than it changes
2012 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2014 * By here, the cache entry is always write-locked. We can
2015 * deadlock if we call osi_Flush with the cache entry locked...
2016 * Unlock the dcache too.
2018 ReleaseWriteLock(&tdc->lock);
2019 if (setLocks && !slowPass)
2020 ReleaseReadLock(&avc->lock);
2022 ReleaseWriteLock(&avc->lock);
2026 * Call osi_FlushPages in open, read/write, and map, since it
2027 * is too hard here to figure out if we should lock the
2030 if (setLocks && !slowPass)
2031 ObtainReadLock(&avc->lock);
2033 ObtainWriteLock(&avc->lock, 66);
2034 ObtainWriteLock(&tdc->lock, 610);
2039 * avc->lock(R) if setLocks && !slowPass
2040 * avc->lock(W) if !setLocks || slowPass
2044 /* Watch for standard race condition around osi_FlushText */
2045 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2046 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2047 afs_stats_cmperf.dcacheHits++;
2048 ConvertWToSLock(&tdc->lock);
2052 /* Sleep here when cache needs to be drained. */
2053 if (setLocks && !slowPass
2054 && (afs_blocksUsed >
2055 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2056 /* Make sure truncate daemon is running */
2057 afs_MaybeWakeupTruncateDaemon();
2058 ObtainWriteLock(&tdc->tlock, 614);
2059 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2060 ReleaseWriteLock(&tdc->tlock);
2061 ReleaseWriteLock(&tdc->lock);
2062 ReleaseReadLock(&avc->lock);
2063 while ((afs_blocksUsed - afs_blocksDiscarded) >
2064 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2065 afs_WaitForCacheDrain = 1;
2066 afs_osi_Sleep(&afs_WaitForCacheDrain);
2068 afs_MaybeFreeDiscardedDCache();
2069 /* need to check if someone else got the chunk first. */
2070 goto RetryGetDCache;
2073 /* Do not fetch data beyond truncPos. */
2074 maxGoodLength = avc->f.m.Length;
2075 if (avc->f.truncPos < maxGoodLength)
2076 maxGoodLength = avc->f.truncPos;
2077 Position = AFS_CHUNKBASE(abyte);
2078 if (vType(avc) == VDIR) {
2079 size = avc->f.m.Length;
2080 if (size > tdc->f.chunkBytes) {
2081 /* pre-reserve space for file */
2082 afs_AdjustSize(tdc, size);
2084 size = 999999999; /* max size for transfer */
2086 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2087 /* don't read past end of good data on server */
2088 if (Position + size > maxGoodLength)
2089 size = maxGoodLength - Position;
2091 size = 0; /* Handle random races */
2092 if (size > tdc->f.chunkBytes) {
2093 /* pre-reserve space for file */
2094 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2095 /* max size for transfer still in size */
2098 if (afs_mariner && !tdc->f.chunk)
2099 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2101 * Right now, we only have one tool, and it's a hammer. So, we
2102 * fetch the whole file.
2104 DZap(tdc); /* pages in cache may be old */
2105 file = afs_CFileOpen(&tdc->f.inode);
2106 afs_RemoveVCB(&avc->f.fid);
2107 tdc->f.states |= DWriting;
2108 tdc->dflags |= DFFetching;
2109 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2110 if (tdc->mflags & DFFetchReq) {
2111 tdc->mflags &= ~DFFetchReq;
2112 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2113 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2114 __FILE__, ICL_TYPE_INT32, __LINE__,
2115 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2119 (struct afs_FetchOutput *)osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2120 setVcacheStatus = 0;
2123 * Remember if we are doing the reading from a replicated volume,
2124 * and how many times we've zipped around the fetch/analyze loop.
2126 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2128 accP = &(afs_stats_cmfullperf.accessinf);
2130 (accP->replicatedRefs)++;
2132 (accP->unreplicatedRefs)++;
2133 #endif /* AFS_NOSTATS */
2134 /* this is a cache miss */
2135 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2136 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2137 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2140 afs_stats_cmperf.dcacheMisses++;
2143 * Dynamic root support: fetch data from local memory.
2145 if (afs_IsDynroot(avc)) {
2149 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2151 dynrootDir += Position;
2152 dynrootLen -= Position;
2153 if (size > dynrootLen)
2157 code = afs_CFileWrite(file, 0, dynrootDir, size);
2165 tdc->validPos = Position + size;
2166 afs_CFileTruncate(file, size); /* prune it */
2167 } else if (afs_IsDynrootMount(avc)) {
2171 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2173 dynrootDir += Position;
2174 dynrootLen -= Position;
2175 if (size > dynrootLen)
2179 code = afs_CFileWrite(file, 0, dynrootDir, size);
2187 tdc->validPos = Position + size;
2188 afs_CFileTruncate(file, size); /* prune it */
2191 * Not a dynamic vnode: do the real fetch.
2196 * avc->lock(R) if setLocks && !slowPass
2197 * avc->lock(W) if !setLocks || slowPass
2201 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
2206 (accP->numReplicasAccessed)++;
2208 #endif /* AFS_NOSTATS */
2209 if (!setLocks || slowPass) {
2210 avc->callback = tc->parent->srvr->server;
2212 newCallback = tc->parent->srvr->server;
2216 code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
2222 /* callback could have been broken (or expired) in a race here,
2223 * but we return the data anyway. It's as good as we knew about
2224 * when we started. */
2226 * validPos is updated by CacheFetchProc, and can only be
2227 * modifed under a dcache write lock, which we've blocked out
2229 size = tdc->validPos - Position; /* actual segment size */
2232 afs_CFileTruncate(file, size); /* prune it */
2234 if (!setLocks || slowPass) {
2235 ObtainWriteLock(&afs_xcbhash, 453);
2236 afs_DequeueCallback(avc);
2237 avc->f.states &= ~(CStatd | CUnique);
2238 avc->callback = NULL;
2239 ReleaseWriteLock(&afs_xcbhash);
2240 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2241 osi_dnlc_purgedp(avc);
2243 /* Something lost. Forget about performance, and go
2244 * back with a vcache write lock.
2246 afs_CFileTruncate(file, 0);
2247 afs_AdjustSize(tdc, 0);
2248 afs_CFileClose(file);
2249 osi_FreeLargeSpace(tsmall);
2251 ReleaseWriteLock(&tdc->lock);
2254 ReleaseReadLock(&avc->lock);
2256 goto RetryGetDCache;
2260 } while (afs_Analyze
2261 (tc, rxconn, code, &avc->f.fid, areq,
2262 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2266 * avc->lock(R) if setLocks && !slowPass
2267 * avc->lock(W) if !setLocks || slowPass
2273 * In the case of replicated access, jot down info on the number of
2274 * attempts it took before we got through or gave up.
2277 if (numFetchLoops <= 1)
2278 (accP->refFirstReplicaOK)++;
2279 if (numFetchLoops > accP->maxReplicasPerRef)
2280 accP->maxReplicasPerRef = numFetchLoops;
2282 #endif /* AFS_NOSTATS */
2284 tdc->dflags &= ~DFFetching;
2285 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2286 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2287 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2288 tdc, ICL_TYPE_INT32, tdc->dflags);
2289 if (avc->execsOrWriters == 0)
2290 tdc->f.states &= ~DWriting;
2292 /* now, if code != 0, we have an error and should punt.
2293 * note that we have the vcache write lock, either because
2294 * !setLocks or slowPass.
2297 afs_CFileTruncate(file, 0);
2298 afs_AdjustSize(tdc, 0);
2299 afs_CFileClose(file);
2300 ZapDCE(tdc); /* sets DFEntryMod */
2301 if (vType(avc) == VDIR) {
2304 tdc->f.states &= ~(DRO|DBackup|DRW);
2305 afs_DCMoveBucket(tdc, 0, 0);
2306 ReleaseWriteLock(&tdc->lock);
2308 if (!afs_IsDynroot(avc)) {
2309 ObtainWriteLock(&afs_xcbhash, 454);
2310 afs_DequeueCallback(avc);
2311 avc->f.states &= ~(CStatd | CUnique);
2312 ReleaseWriteLock(&afs_xcbhash);
2313 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2314 osi_dnlc_purgedp(avc);
2317 * avc->lock(W); assert(!setLocks || slowPass)
2319 osi_Assert(!setLocks || slowPass);
2325 /* otherwise we copy in the just-fetched info */
2326 afs_CFileClose(file);
2327 afs_AdjustSize(tdc, size); /* new size */
2329 * Copy appropriate fields into vcache. Status is
2330 * copied later where we selectively acquire the
2331 * vcache write lock.
2334 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2336 setVcacheStatus = 1;
2337 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2338 tsmall->OutStatus.DataVersion);
2339 tdc->dflags |= DFEntryMod;
2340 afs_indexFlags[tdc->index] |= IFEverUsed;
2341 ConvertWToSLock(&tdc->lock);
2342 } /*Data version numbers don't match */
2345 * Data version numbers match.
2347 afs_stats_cmperf.dcacheHits++;
2348 } /*Data version numbers match */
2350 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2354 * avc->lock(R) if setLocks && !slowPass
2355 * avc->lock(W) if !setLocks || slowPass
2356 * tdc->lock(S) if tdc
2360 * See if this was a reference to a file in the local cell.
2362 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2363 afs_stats_cmperf.dlocalAccesses++;
2365 afs_stats_cmperf.dremoteAccesses++;
2367 /* Fix up LRU info */
2370 ObtainWriteLock(&afs_xdcache, 602);
2371 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2372 hadd32(afs_indexCounter, 1);
2373 ReleaseWriteLock(&afs_xdcache);
2375 /* return the data */
2376 if (vType(avc) == VDIR)
2379 *aoffset = AFS_CHUNKOFFSET(abyte);
2380 *alen = (tdc->f.chunkBytes - *aoffset);
2381 ReleaseSharedLock(&tdc->lock);
2386 * avc->lock(R) if setLocks && !slowPass
2387 * avc->lock(W) if !setLocks || slowPass
2390 /* Fix up the callback and status values in the vcache */
2392 if (setLocks && !slowPass) {
2395 * This is our dirty little secret to parallel fetches.
2396 * We don't write-lock the vcache while doing the fetch,
2397 * but potentially we'll need to update the vcache after
2398 * the fetch is done.
2400 * Drop the read lock and try to re-obtain the write
2401 * lock. If the vcache still has the same DV, it's
2402 * ok to go ahead and install the new data.
2404 afs_hyper_t currentDV, statusDV;
2406 hset(currentDV, avc->f.m.DataVersion);
2408 if (setNewCallback && avc->callback != newCallback)
2412 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2413 tsmall->OutStatus.DataVersion);
2415 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2417 if (setVcacheStatus && !hsame(currentDV, statusDV))
2421 ReleaseReadLock(&avc->lock);
2423 if (doVcacheUpdate) {
2424 ObtainWriteLock(&avc->lock, 615);
2425 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2426 /* We lose. Someone will beat us to it. */
2428 ReleaseWriteLock(&avc->lock);
2433 /* With slow pass, we've already done all the updates */
2435 ReleaseWriteLock(&avc->lock);
2438 /* Check if we need to perform any last-minute fixes with a write-lock */
2439 if (!setLocks || doVcacheUpdate) {
2441 avc->callback = newCallback;
2442 if (tsmall && setVcacheStatus)
2443 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2445 ReleaseWriteLock(&avc->lock);
2449 osi_FreeLargeSpace(tsmall);
2452 } /*afs_GetDCache */
2456 * afs_WriteThroughDSlots
2459 * Sweep through the dcache slots and write out any modified
2460 * in-memory data back on to our caching store.
2466 * The afs_xdcache is write-locked through this whole affair.
2469 afs_WriteThroughDSlots(void)
2472 afs_int32 i, touchedit = 0;
2474 struct afs_q DirtyQ, *tq;
2476 AFS_STATCNT(afs_WriteThroughDSlots);
2479 * Because of lock ordering, we can't grab dcache locks while
2480 * holding afs_xdcache. So we enter xdcache, get a reference
2481 * for every dcache entry, and exit xdcache.
2483 ObtainWriteLock(&afs_xdcache, 283);
2485 for (i = 0; i < afs_cacheFiles; i++) {
2486 tdc = afs_indexTable[i];
2488 /* Grab tlock in case the existing refcount isn't zero */
2489 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2490 ObtainWriteLock(&tdc->tlock, 623);
2492 ReleaseWriteLock(&tdc->tlock);
2494 QAdd(&DirtyQ, &tdc->dirty);
2497 ReleaseWriteLock(&afs_xdcache);
2500 * Now, for each dcache entry we found, check if it's dirty.
2501 * If so, get write-lock, get afs_xdcache, which protects
2502 * afs_cacheInodep, and flush it. Don't forget to put back
2506 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2508 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2510 if (tdc->dflags & DFEntryMod) {
2513 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2515 /* Now that we have the write lock, double-check */
2516 if (wrLock && (tdc->dflags & DFEntryMod)) {
2517 tdc->dflags &= ~DFEntryMod;
2518 ObtainWriteLock(&afs_xdcache, 620);
2519 afs_WriteDCache(tdc, 1);
2520 ReleaseWriteLock(&afs_xdcache);
2524 ReleaseWriteLock(&tdc->lock);
2530 ObtainWriteLock(&afs_xdcache, 617);
2531 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2532 /* Touch the file to make sure that the mtime on the file is kept
2533 * up-to-date to avoid losing cached files on cold starts because
2534 * their mtime seems old...
2536 struct afs_fheader theader;
2538 theader.magic = AFS_FHMAGIC;
2539 theader.firstCSize = AFS_FIRSTCSIZE;
2540 theader.otherCSize = AFS_OTHERCSIZE;
2541 theader.version = AFS_CI_VERSION;
2542 theader.dataSize = sizeof(struct fcache);
2543 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2545 ReleaseWriteLock(&afs_xdcache);
2552 * Return a pointer to an freshly initialized dcache entry using
2553 * a memory-based cache. The tlock will be read-locked.
2556 * aslot : Dcache slot to look at.
2557 * tmpdc : Ptr to dcache entry.
2560 * Must be called with afs_xdcache write-locked.
2564 afs_MemGetDSlot(afs_int32 aslot, struct dcache *tmpdc)
2569 AFS_STATCNT(afs_MemGetDSlot);
2570 if (CheckLock(&afs_xdcache) != -1)
2571 osi_Panic("getdslot nolock");
2572 if (aslot < 0 || aslot >= afs_cacheFiles)
2573 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2574 tdc = afs_indexTable[aslot];
2576 QRemove(&tdc->lruq); /* move to queue head */
2577 QAdd(&afs_DLRU, &tdc->lruq);
2578 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2579 ObtainWriteLock(&tdc->tlock, 624);
2581 ConvertWToRLock(&tdc->tlock);
2584 if (tmpdc == NULL) {
2585 if (!afs_freeDSList)
2586 afs_GetDownDSlot(4);
2587 if (!afs_freeDSList) {
2588 /* none free, making one is better than a panic */
2589 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2590 tdc = afs_osi_Alloc(sizeof(struct dcache));
2591 osi_Assert(tdc != NULL);
2592 #ifdef KERNEL_HAVE_PIN
2593 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2596 tdc = afs_freeDSList;
2597 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2600 tdc->dflags = 0; /* up-to-date, not in free q */
2602 QAdd(&afs_DLRU, &tdc->lruq);
2603 if (tdc->lruq.prev == &tdc->lruq)
2604 osi_Panic("lruq 3");
2610 /* initialize entry */
2611 tdc->f.fid.Cell = 0;
2612 tdc->f.fid.Fid.Volume = 0;
2614 hones(tdc->f.versionNo);
2615 tdc->f.inode.mem = aslot;
2616 tdc->dflags |= DFEntryMod;
2619 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2622 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2623 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2624 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2627 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2628 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2629 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2630 ObtainReadLock(&tdc->tlock);
2633 afs_indexTable[aslot] = tdc;
2636 } /*afs_MemGetDSlot */
2638 unsigned int last_error = 0, lasterrtime = 0;
2644 * Return a pointer to an freshly initialized dcache entry using
2645 * a UFS-based disk cache. The dcache tlock will be read-locked.
2648 * aslot : Dcache slot to look at.
2649 * tmpdc : Ptr to dcache entry.
2652 * afs_xdcache lock write-locked.
2655 afs_UFSGetDSlot(afs_int32 aslot, struct dcache *tmpdc)
2662 AFS_STATCNT(afs_UFSGetDSlot);
2663 if (CheckLock(&afs_xdcache) != -1)
2664 osi_Panic("getdslot nolock");
2665 if (aslot < 0 || aslot >= afs_cacheFiles)
2666 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2667 tdc = afs_indexTable[aslot];
2669 QRemove(&tdc->lruq); /* move to queue head */
2670 QAdd(&afs_DLRU, &tdc->lruq);
2671 /* Grab tlock in case refCount != 0 */
2672 ObtainWriteLock(&tdc->tlock, 625);
2674 ConvertWToRLock(&tdc->tlock);
2677 /* otherwise we should read it in from the cache file */
2679 * If we weren't passed an in-memory region to place the file info,
2680 * we have to allocate one.
2682 if (tmpdc == NULL) {
2683 if (!afs_freeDSList)
2684 afs_GetDownDSlot(4);
2685 if (!afs_freeDSList) {
2686 /* none free, making one is better than a panic */
2687 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2688 tdc = afs_osi_Alloc(sizeof(struct dcache));
2689 osi_Assert(tdc != NULL);
2690 #ifdef KERNEL_HAVE_PIN
2691 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2694 tdc = afs_freeDSList;
2695 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2698 tdc->dflags = 0; /* up-to-date, not in free q */
2700 QAdd(&afs_DLRU, &tdc->lruq);
2701 if (tdc->lruq.prev == &tdc->lruq)
2702 osi_Panic("lruq 3");
2709 * Seek to the aslot'th entry and read it in.
2712 afs_osi_Read(afs_cacheInodep,
2713 sizeof(struct fcache) * aslot +
2714 sizeof(struct afs_fheader), (char *)(&tdc->f),
2715 sizeof(struct fcache));
2717 if (code != sizeof(struct fcache))
2719 if (!afs_CellNumValid(tdc->f.fid.Cell))
2723 tdc->f.fid.Cell = 0;
2724 tdc->f.fid.Fid.Volume = 0;
2726 hones(tdc->f.versionNo);
2727 tdc->dflags |= DFEntryMod;
2728 #if defined(KERNEL_HAVE_UERROR)
2729 last_error = getuerror();
2731 lasterrtime = osi_Time();
2732 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2733 tdc->f.states &= ~(DRO|DBackup|DRW);
2734 afs_DCMoveBucket(tdc, 0, 0);
2737 if (tdc->f.states & DRO) {
2738 afs_DCMoveBucket(tdc, 0, 2);
2739 } else if (tdc->f.states & DBackup) {
2740 afs_DCMoveBucket(tdc, 0, 1);
2742 afs_DCMoveBucket(tdc, 0, 1);
2748 if (tdc->f.chunk >= 0)
2749 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2754 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2755 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2756 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2759 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2760 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2761 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2762 ObtainReadLock(&tdc->tlock);
2765 * If we didn't read into a temporary dcache region, update the
2766 * slot pointer table.
2769 afs_indexTable[aslot] = tdc;
2772 } /*afs_UFSGetDSlot */
2777 * Write a particular dcache entry back to its home in the
2780 * \param adc Pointer to the dcache entry to write.
2781 * \param atime If true, set the modtime on the file to the current time.
2783 * \note Environment:
2784 * Must be called with the afs_xdcache lock at least read-locked,
2785 * and dcache entry at least read-locked.
2786 * The reference count is not changed.
2790 afs_WriteDCache(struct dcache *adc, int atime)
2794 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2796 AFS_STATCNT(afs_WriteDCache);
2797 osi_Assert(WriteLocked(&afs_xdcache));
2799 adc->f.modTime = osi_Time();
2801 * Seek to the right dcache slot and write the in-memory image out to disk.
2803 afs_cellname_write();
2805 afs_osi_Write(afs_cacheInodep,
2806 sizeof(struct fcache) * adc->index +
2807 sizeof(struct afs_fheader), (char *)(&adc->f),
2808 sizeof(struct fcache));
2809 if (code != sizeof(struct fcache))
2817 * Wake up users of a particular file waiting for stores to take
2820 * \param avc Ptr to related vcache entry.
2822 * \note Environment:
2823 * Nothing interesting.
2826 afs_wakeup(struct vcache *avc)
2829 struct brequest *tb;
2831 AFS_STATCNT(afs_wakeup);
2832 for (i = 0; i < NBRS; i++, tb++) {
2833 /* if request is valid and for this file, we've found it */
2834 if (tb->refCount > 0 && avc == tb->vc) {
2837 * If CSafeStore is on, then we don't awaken the guy
2838 * waiting for the store until the whole store has finished.
2839 * Otherwise, we do it now. Note that if CSafeStore is on,
2840 * the BStore routine actually wakes up the user, instead
2842 * I think this is redundant now because this sort of thing
2843 * is already being handled by the higher-level code.
2845 if ((avc->f.states & CSafeStore) == 0) {
2847 tb->flags |= BUVALID;
2848 if (tb->flags & BUWAIT) {
2849 tb->flags &= ~BUWAIT;
2861 * Given a file name and inode, set up that file to be an
2862 * active member in the AFS cache. This also involves checking
2863 * the usability of its data.
2865 * \param afile Name of the cache file to initialize.
2866 * \param ainode Inode of the file.
2868 * \note Environment:
2869 * This function is called only during initialization.
2872 afs_InitCacheFile(char *afile, ino_t ainode)
2877 struct osi_file *tfile;
2878 struct osi_stat tstat;
2881 AFS_STATCNT(afs_InitCacheFile);
2882 index = afs_stats_cmperf.cacheNumEntries;
2883 if (index >= afs_cacheFiles)
2886 ObtainWriteLock(&afs_xdcache, 282);
2887 tdc = afs_GetDSlot(index, NULL);
2888 ReleaseReadLock(&tdc->tlock);
2889 ReleaseWriteLock(&afs_xdcache);
2891 ObtainWriteLock(&tdc->lock, 621);
2892 ObtainWriteLock(&afs_xdcache, 622);
2894 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
2896 ReleaseWriteLock(&afs_xdcache);
2897 ReleaseWriteLock(&tdc->lock);
2902 /* Add any other 'complex' inode types here ... */
2903 #if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
2904 tdc->f.inode.ufs = ainode;
2906 osi_Panic("Can't init cache with inode numbers when complex inodes are "
2911 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
2913 tfile = osi_UFSOpen(&tdc->f.inode);
2914 code = afs_osi_Stat(tfile, &tstat);
2916 osi_Panic("initcachefile stat");
2919 * If file size doesn't match the cache info file, it's probably bad.
2921 if (tdc->f.chunkBytes != tstat.size)
2923 tdc->f.chunkBytes = 0;
2926 * If file changed within T (120?) seconds of cache info file, it's
2927 * probably bad. In addition, if slot changed within last T seconds,
2928 * the cache info file may be incorrectly identified, and so slot
2931 if (cacheInfoModTime < tstat.mtime + 120)
2933 if (cacheInfoModTime < tdc->f.modTime + 120)
2935 /* In case write through is behind, make sure cache items entry is
2936 * at least as new as the chunk.
2938 if (tdc->f.modTime < tstat.mtime)
2941 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
2942 if (tstat.size != 0)
2943 osi_UFSTruncate(tfile, 0);
2944 tdc->f.states &= ~(DRO|DBackup|DRW);
2945 afs_DCMoveBucket(tdc, 0, 0);
2946 /* put entry in free cache slot list */
2947 afs_dvnextTbl[tdc->index] = afs_freeDCList;
2948 afs_freeDCList = index;
2950 afs_indexFlags[index] |= IFFree;
2951 afs_indexUnique[index] = 0;
2954 * We must put this entry in the appropriate hash tables.
2955 * Note that i is still set from the above DCHash call
2957 code = DCHash(&tdc->f.fid, tdc->f.chunk);
2958 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
2959 afs_dchashTbl[code] = tdc->index;
2960 code = DVHash(&tdc->f.fid);
2961 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
2962 afs_dvhashTbl[code] = tdc->index;
2963 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
2965 /* has nontrivial amt of data */
2966 afs_indexFlags[index] |= IFEverUsed;
2967 afs_stats_cmperf.cacheFilesReused++;
2969 * Initialize index times to file's mod times; init indexCounter
2972 hset32(afs_indexTimes[index], tstat.atime);
2973 if (hgetlo(afs_indexCounter) < tstat.atime) {
2974 hset32(afs_indexCounter, tstat.atime);
2976 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
2977 } /*File is not bad */
2979 osi_UFSClose(tfile);
2980 tdc->f.states &= ~DWriting;
2981 tdc->dflags &= ~DFEntryMod;
2982 /* don't set f.modTime; we're just cleaning up */
2983 afs_WriteDCache(tdc, 0);
2984 ReleaseWriteLock(&afs_xdcache);
2985 ReleaseWriteLock(&tdc->lock);
2987 afs_stats_cmperf.cacheNumEntries++;
2992 /*Max # of struct dcache's resident at any time*/
2994 * If 'dchint' is enabled then in-memory dcache min is increased because of
3000 * Initialize dcache related variables.
3010 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3016 afs_freeDCList = NULLIDX;
3017 afs_discardDCList = NULLIDX;
3018 afs_freeDCCount = 0;
3019 afs_freeDSList = NULL;
3020 hzero(afs_indexCounter);
3022 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3028 if (achunk < 0 || achunk > 30)
3029 achunk = 13; /* Use default */
3030 AFS_SETCHUNKSIZE(achunk);
3036 if (aflags & AFSCALL_INIT_MEMCACHE) {
3038 * Use a memory cache instead of a disk cache
3040 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3041 afs_cacheType = &afs_MemCacheOps;
3042 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3043 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3044 /* ablocks is reported in 1K blocks */
3045 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3047 afs_warn("afsd: memory cache too large for available memory.\n");
3048 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3050 afiles = ablocks = 0;
3052 afs_warn("Memory cache: Allocating %d dcache entries...",
3055 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3056 afs_cacheType = &afs_UfsCacheOps;
3059 if (aDentries > 512)
3060 afs_dhashsize = 2048;
3061 /* initialize hash tables */
3062 afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3063 osi_Assert(afs_dvhashTbl != NULL);
3064 afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3065 osi_Assert(afs_dchashTbl != NULL);
3066 for (i = 0; i < afs_dhashsize; i++) {
3067 afs_dvhashTbl[i] = NULLIDX;
3068 afs_dchashTbl[i] = NULLIDX;
3070 afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3071 osi_Assert(afs_dvnextTbl != NULL);
3072 afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3073 osi_Assert(afs_dcnextTbl != NULL);
3074 for (i = 0; i < afiles; i++) {
3075 afs_dvnextTbl[i] = NULLIDX;
3076 afs_dcnextTbl[i] = NULLIDX;
3079 /* Allocate and zero the pointer array to the dcache entries */
3080 afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3081 osi_Assert(afs_indexTable != NULL);
3082 memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3083 afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3084 osi_Assert(afs_indexTimes != NULL);
3085 memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3086 afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
3087 osi_Assert(afs_indexUnique != NULL);
3088 memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3089 afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
3090 osi_Assert(afs_indexFlags != NULL);
3091 memset(afs_indexFlags, 0, afiles * sizeof(char));
3093 /* Allocate and thread the struct dcache entries themselves */
3094 tdp = afs_Initial_freeDSList =
3095 afs_osi_Alloc(aDentries * sizeof(struct dcache));
3096 osi_Assert(tdp != NULL);
3097 memset(tdp, 0, aDentries * sizeof(struct dcache));
3098 #ifdef KERNEL_HAVE_PIN
3099 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3100 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3101 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3102 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3103 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3104 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3105 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3106 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3107 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3110 afs_freeDSList = &tdp[0];
3111 for (i = 0; i < aDentries - 1; i++) {
3112 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3113 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3114 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3115 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3117 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3118 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3119 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3120 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3122 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3123 afs_cacheBlocks = ablocks;
3124 afs_ComputeCacheParms(); /* compute parms based on cache size */
3126 afs_dcentries = aDentries;
3128 afs_stats_cmperf.cacheBucket0_Discarded =
3129 afs_stats_cmperf.cacheBucket1_Discarded =
3130 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3136 * Shuts down the cache.
3140 shutdown_dcache(void)
3144 #ifdef AFS_CACHE_VNODE_PATH
3145 if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
3147 for (i = 0; i < afs_cacheFiles; i++) {
3148 tdc = afs_indexTable[i];
3150 afs_osi_FreeStr(tdc->f.inode.ufs);
3156 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3157 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3158 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3159 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3160 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3161 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3162 afs_osi_Free(afs_Initial_freeDSList,
3163 afs_dcentries * sizeof(struct dcache));
3164 #ifdef KERNEL_HAVE_PIN
3165 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3166 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3167 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3168 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3169 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3170 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3171 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3175 for (i = 0; i < afs_dhashsize; i++) {
3176 afs_dvhashTbl[i] = NULLIDX;
3177 afs_dchashTbl[i] = NULLIDX;
3180 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3181 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3183 afs_blocksUsed = afs_dcentries = 0;
3184 afs_stats_cmperf.cacheBucket0_Discarded =
3185 afs_stats_cmperf.cacheBucket1_Discarded =
3186 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3187 hzero(afs_indexCounter);
3189 afs_freeDCCount = 0;
3190 afs_freeDCList = NULLIDX;
3191 afs_discardDCList = NULLIDX;
3192 afs_freeDSList = afs_Initial_freeDSList = 0;
3194 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3200 * Get a dcache ready for writing, respecting the current cache size limits
3202 * len is required because afs_GetDCache with flag == 4 expects the length
3203 * field to be filled. It decides from this whether it's necessary to fetch
3204 * data into the chunk before writing or not (when the whole chunk is
3207 * \param avc The vcache to fetch a dcache for
3208 * \param filePos The start of the section to be written
3209 * \param len The length of the section to be written
3213 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3214 * must be released and afs_PutDCache() called to free dcache.
3217 * \note avc->lock must be held on entry. Function may release and reobtain
3218 * avc->lock and GLOCK.
3222 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3223 afs_size_t len, struct vrequest *areq,
3226 struct dcache *tdc = NULL;
3229 /* read the cached info */
3231 tdc = afs_FindDCache(avc, filePos);
3233 ObtainWriteLock(&tdc->lock, 657);
3234 } else if (afs_blocksUsed >
3235 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3236 tdc = afs_FindDCache(avc, filePos);
3238 ObtainWriteLock(&tdc->lock, 658);
3239 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3240 || (tdc->dflags & DFFetching)) {
3241 ReleaseWriteLock(&tdc->lock);
3247 afs_MaybeWakeupTruncateDaemon();
3248 while (afs_blocksUsed >
3249 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3250 ReleaseWriteLock(&avc->lock);
3251 if (afs_blocksUsed - afs_blocksDiscarded >
3252 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3253 afs_WaitForCacheDrain = 1;
3254 afs_osi_Sleep(&afs_WaitForCacheDrain);
3256 afs_MaybeFreeDiscardedDCache();
3257 afs_MaybeWakeupTruncateDaemon();
3258 ObtainWriteLock(&avc->lock, 509);
3260 avc->f.states |= CDirty;
3261 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3263 ObtainWriteLock(&tdc->lock, 659);
3266 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3268 ObtainWriteLock(&tdc->lock, 660);
3271 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3272 afs_stats_cmperf.cacheCurrDirtyChunks++;
3273 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3275 if (!(tdc->f.states & DWriting)) {
3276 /* don't mark entry as mod if we don't have to */
3277 tdc->f.states |= DWriting;
3278 tdc->dflags |= DFEntryMod;
3285 * Make a shadow copy of a dir's dcache. It's used for disconnected
3286 * operations like remove/create/rename to keep the original directory data.
3287 * On reconnection, we can diff the original data with the server and get the
3288 * server changes and with the local data to get the local changes.
3290 * \param avc The dir vnode.
3291 * \param adc The dir dcache.
3293 * \return 0 for success.
3295 * \note The vcache entry must be write locked.
3296 * \note The dcache entry must be read locked.
3299 afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3301 int i, code, ret_code = 0, written, trans_size;
3302 struct dcache *new_dc = NULL;
3303 struct osi_file *tfile_src, *tfile_dst;
3304 struct VenusFid shadow_fid;
3307 /* Is this a dir? */
3308 if (vType(avc) != VDIR)
3311 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3314 /* Generate a fid for the shadow dir. */
3315 shadow_fid.Cell = avc->f.fid.Cell;
3316 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3317 afs_GenShadowFid(&shadow_fid);
3319 ObtainWriteLock(&afs_xdcache, 716);
3321 /* Get a fresh dcache. */
3322 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3324 ObtainReadLock(&adc->mflock);
3326 /* Set up the new fid. */
3327 /* Copy interesting data from original dir dcache. */
3328 new_dc->mflags = adc->mflags;
3329 new_dc->dflags = adc->dflags;
3330 new_dc->f.modTime = adc->f.modTime;
3331 new_dc->f.versionNo = adc->f.versionNo;
3332 new_dc->f.states = adc->f.states;
3333 new_dc->f.chunk= adc->f.chunk;
3334 new_dc->f.chunkBytes = adc->f.chunkBytes;
3336 ReleaseReadLock(&adc->mflock);
3338 /* Now add to the two hash chains */
3339 i = DCHash(&shadow_fid, 0);
3340 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3341 afs_dchashTbl[i] = new_dc->index;
3343 i = DVHash(&shadow_fid);
3344 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3345 afs_dvhashTbl[i] = new_dc->index;
3347 ReleaseWriteLock(&afs_xdcache);
3349 /* Alloc a 4k block. */
3350 data = afs_osi_Alloc(4096);
3352 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3357 /* Open the files. */
3358 tfile_src = afs_CFileOpen(&adc->f.inode);
3359 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3361 /* And now copy dir dcache data into this dcache,
3365 while (written < adc->f.chunkBytes) {
3366 trans_size = adc->f.chunkBytes - written;
3367 if (trans_size > 4096)
3370 /* Read a chunk from the dcache. */
3371 code = afs_CFileRead(tfile_src, written, data, trans_size);
3372 if (code < trans_size) {
3377 /* Write it to the new dcache. */
3378 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3379 if (code < trans_size) {
3384 written+=trans_size;
3387 afs_CFileClose(tfile_dst);
3388 afs_CFileClose(tfile_src);
3390 afs_osi_Free(data, 4096);
3392 ReleaseWriteLock(&new_dc->lock);
3393 afs_PutDCache(new_dc);
3396 ObtainWriteLock(&afs_xvcache, 763);
3397 ObtainWriteLock(&afs_disconDirtyLock, 765);
3398 QAdd(&afs_disconShadow, &avc->shadowq);
3399 osi_Assert((afs_RefVCache(avc) == 0));
3400 ReleaseWriteLock(&afs_disconDirtyLock);
3401 ReleaseWriteLock(&afs_xvcache);
3403 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3404 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3412 * Delete the dcaches of a shadow dir.
3414 * \param avc The vcache containing the shadow fid.
3416 * \note avc must be write locked.
3419 afs_DeleteShadowDir(struct vcache *avc)
3422 struct VenusFid shadow_fid;
3424 shadow_fid.Cell = avc->f.fid.Cell;
3425 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3426 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3427 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3429 tdc = afs_FindDCacheByFid(&shadow_fid);
3431 afs_HashOutDCache(tdc, 1);
3432 afs_DiscardDCache(tdc);
3435 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3436 ObtainWriteLock(&afs_disconDirtyLock, 708);
3437 QRemove(&avc->shadowq);
3438 ReleaseWriteLock(&afs_disconDirtyLock);
3439 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3443 * Populate a dcache with empty chunks up to a given file size,
3444 * used before extending a file in order to avoid 'holes' which
3445 * we can't access in disconnected mode.
3447 * \param avc The vcache which is being extended (locked)
3448 * \param alen The new length of the file
3452 afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
3455 afs_size_t len, offset;
3456 afs_int32 start, end;
3458 /* We're doing this to deal with the situation where we extend
3459 * by writing after lseek()ing past the end of the file . If that
3460 * extension skips chunks, then those chunks won't be created, and
3461 * GetDCache will assume that they have to be fetched from the server.
3462 * So, for each chunk between the current file position, and the new
3463 * length we GetDCache for that chunk.
3466 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3469 if (avc->f.m.Length == 0)
3472 start = AFS_CHUNK(avc->f.m.Length)+1;
3474 end = AFS_CHUNK(apos);
3477 len = AFS_CHUNKTOSIZE(start);
3478 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);