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"
25 /* Forward declarations. */
26 static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
27 static int 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);
38 * --------------------- Exported definitions ---------------------
41 afs_int32 afs_blocksUsed_0; /*!< 1K blocks in cache - in theory is zero */
42 afs_int32 afs_blocksUsed_1; /*!< 1K blocks in cache */
43 afs_int32 afs_blocksUsed_2; /*!< 1K blocks in cache */
44 afs_int32 afs_pct1 = -1;
45 afs_int32 afs_pct2 = -1;
46 afs_uint32 afs_tpct1 = 0;
47 afs_uint32 afs_tpct2 = 0;
48 afs_uint32 splitdcache = 0;
50 afs_lock_t afs_xdcache; /*!< Lock: alloc new disk cache entries */
51 afs_int32 afs_freeDCList; /*!< Free list for disk cache entries */
52 afs_int32 afs_freeDCCount; /*!< Count of elts in freeDCList */
53 afs_int32 afs_discardDCList; /*!< Discarded disk cache entries */
54 afs_int32 afs_discardDCCount; /*!< Count of elts in discardDCList */
55 struct dcache *afs_freeDSList; /*!< Free list for disk slots */
56 struct dcache *afs_Initial_freeDSList; /*!< Initial list for above */
57 afs_dcache_id_t cacheInode; /*!< Inode for CacheItems file */
58 struct osi_file *afs_cacheInodep = 0; /*!< file for CacheItems inode */
59 struct afs_q afs_DLRU; /*!< dcache LRU */
60 afs_int32 afs_dhashsize = 1024;
61 afs_int32 *afs_dvhashTbl; /*!< Data cache hash table: hashed by FID + chunk number. */
62 afs_int32 *afs_dchashTbl; /*!< Data cache hash table: hashed by FID. */
63 afs_int32 *afs_dvnextTbl; /*!< Dcache hash table links */
64 afs_int32 *afs_dcnextTbl; /*!< Dcache hash table links */
65 struct dcache **afs_indexTable; /*!< Pointers to dcache entries */
66 afs_hyper_t *afs_indexTimes; /*!< Dcache entry Access times */
67 afs_int32 *afs_indexUnique; /*!< dcache entry Fid.Unique */
68 unsigned char *afs_indexFlags; /*!< (only one) Is there data there? */
69 afs_hyper_t afs_indexCounter; /*!< Fake time for marking index
71 afs_int32 afs_cacheFiles = 0; /*!< Size of afs_indexTable */
72 afs_int32 afs_cacheBlocks; /*!< 1K blocks in cache */
73 afs_int32 afs_cacheStats; /*!< Stat entries in cache */
74 afs_int32 afs_blocksUsed; /*!< Number of blocks in use */
75 afs_int32 afs_blocksDiscarded; /*!<Blocks freed but not truncated */
76 afs_int32 afs_fsfragsize = AFS_MIN_FRAGSIZE; /*!< Underlying Filesystem minimum unit
77 *of disk allocation usually 1K
78 *this value is (truefrag -1 ) to
79 *save a bunch of subtracts... */
80 #ifdef AFS_64BIT_CLIENT
81 #ifdef AFS_VM_RDWR_ENV
82 afs_size_t afs_vmMappingEnd; /* !< For large files (>= 2GB) the VM
83 * mapping an 32bit addressing machines
84 * can only be used below the 2 GB
85 * line. From this point upwards we
86 * must do direct I/O into the cache
87 * files. The value should be on a
89 #endif /* AFS_VM_RDWR_ENV */
90 #endif /* AFS_64BIT_CLIENT */
92 /* The following is used to ensure that new dcache's aren't obtained when
93 * the cache is nearly full.
95 int afs_WaitForCacheDrain = 0;
96 int afs_TruncateDaemonRunning = 0;
97 int afs_CacheTooFull = 0;
99 afs_int32 afs_dcentries; /*!< In-memory dcache entries */
102 int dcacheDisabled = 0;
104 struct afs_cacheOps afs_UfsCacheOps = {
105 #ifndef HAVE_STRUCT_LABEL_SUPPORT
118 .truncate = osi_UFSTruncate,
119 .fread = afs_osi_Read,
120 .fwrite = afs_osi_Write,
121 .close = osi_UFSClose,
122 .vreadUIO = afs_UFSReadUIO,
123 .vwriteUIO = afs_UFSWriteUIO,
124 .GetDSlot = afs_UFSGetDSlot,
125 .GetVolSlot = afs_UFSGetVolSlot,
126 .HandleLink = afs_UFSHandleLink,
130 struct afs_cacheOps afs_MemCacheOps = {
131 #ifndef HAVE_STRUCT_LABEL_SUPPORT
133 afs_MemCacheTruncate,
143 .open = afs_MemCacheOpen,
144 .truncate = afs_MemCacheTruncate,
145 .fread = afs_MemReadBlk,
146 .fwrite = afs_MemWriteBlk,
147 .close = afs_MemCacheClose,
148 .vreadUIO = afs_MemReadUIO,
149 .vwriteUIO = afs_MemWriteUIO,
150 .GetDSlot = afs_MemGetDSlot,
151 .GetVolSlot = afs_MemGetVolSlot,
152 .HandleLink = afs_MemHandleLink,
156 int cacheDiskType; /*Type of backing disk for cache */
157 struct afs_cacheOps *afs_cacheType;
161 * The PFlush algorithm makes use of the fact that Fid.Unique is not used in
162 * below hash algorithms. Change it if need be so that flushing algorithm
163 * doesn't move things from one hash chain to another.
165 /*Vnode, Chunk -> Hash table index */
166 int DCHash(struct VenusFid *fid, afs_int32 chunk)
170 buf[0] = fid->Fid.Volume;
171 buf[1] = fid->Fid.Vnode;
173 return opr_jhash(buf, 3, 0) & (afs_dhashsize - 1);
175 /*Vnode -> Other hash table index */
176 int DVHash(struct VenusFid *fid)
178 return opr_jhash_int2(fid->Fid.Volume, fid->Fid.Vnode, 0) &
183 * Where is this vcache's entry associated dcache located/
184 * \param avc The vcache entry.
185 * \return Bucket index:
190 afs_DCGetBucket(struct vcache *avc)
195 /* This should be replaced with some sort of user configurable function */
196 if (avc->f.states & CRO) {
198 } else if (avc->f.states & CBackup) {
208 * Readjust a dcache's size.
210 * \param adc The dcache to be adjusted.
211 * \param oldSize Old size for the dcache.
212 * \param newSize The new size to be adjusted to.
216 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
218 afs_int32 adjustSize = newSize - oldSize;
226 afs_blocksUsed_0 += adjustSize;
227 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
230 afs_blocksUsed_1 += adjustSize;
231 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
234 afs_blocksUsed_2 += adjustSize;
235 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
243 * Move a dcache from one bucket to another.
245 * \param adc Operate on this dcache.
246 * \param size Size in bucket (?).
247 * \param newBucket Destination bucket.
251 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
256 /* Substract size from old bucket. */
260 afs_blocksUsed_0 -= size;
263 afs_blocksUsed_1 -= size;
266 afs_blocksUsed_2 -= size;
270 /* Set new bucket and increase destination bucket size. */
271 adc->bucket = newBucket;
276 afs_blocksUsed_0 += size;
279 afs_blocksUsed_1 += size;
282 afs_blocksUsed_2 += size;
290 * Init split caches size.
295 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
304 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
309 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
310 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
312 /* Short cut: if we don't know about it, try to kill it */
313 if (phase < 2 && afs_blocksUsed_0)
316 if (afs_pct1 > afs_tpct1)
318 if (afs_pct2 > afs_tpct2)
320 return 0; /* unlikely */
325 * Warn about failing to store a file.
327 * \param acode Associated error code.
328 * \param avolume Volume involved.
329 * \param aflags How to handle the output:
330 * aflags & 1: Print out on console
331 * aflags & 2: Print out on controlling tty
333 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
337 afs_StoreWarn(afs_int32 acode, afs_int32 avolume,
340 static char problem_fmt[] =
341 "afs: failed to store file in volume %d (%s)\n";
342 static char problem_fmt_w_error[] =
343 "afs: failed to store file in volume %d (error %d)\n";
344 static char netproblems[] = "network problems";
345 static char partfull[] = "partition full";
346 static char overquota[] = "over quota";
348 AFS_STATCNT(afs_StoreWarn);
354 afs_warn(problem_fmt, avolume, netproblems);
356 afs_warnuser(problem_fmt, avolume, netproblems);
357 } else if (acode == ENOSPC) {
362 afs_warn(problem_fmt, avolume, partfull);
364 afs_warnuser(problem_fmt, avolume, partfull);
367 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
368 * Instead ENOSPC will be sent...
370 if (acode == EDQUOT) {
375 afs_warn(problem_fmt, avolume, overquota);
377 afs_warnuser(problem_fmt, avolume, overquota);
385 afs_warn(problem_fmt_w_error, avolume, acode);
387 afs_warnuser(problem_fmt_w_error, avolume, acode);
392 * Try waking up truncation daemon, if it's worth it.
395 afs_MaybeWakeupTruncateDaemon(void)
397 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
398 afs_CacheTooFull = 1;
399 if (!afs_TruncateDaemonRunning)
400 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
401 } else if (!afs_TruncateDaemonRunning
402 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
403 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
410 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
411 * struct so we need only export one symbol for AIX.
413 static struct CTD_stats {
414 osi_timeval_t CTD_beforeSleep;
415 osi_timeval_t CTD_afterSleep;
416 osi_timeval_t CTD_sleepTime;
417 osi_timeval_t CTD_runTime;
421 u_int afs_min_cache = 0;
424 * If there are waiters for the cache to drain, wake them if
425 * the number of free or discarded cache blocks reaches the
426 * CM_CACHESIZEDDRAINEDPCT limit.
429 * This routine must be called with the afs_xdcache lock held
433 afs_WakeCacheWaitersIfDrained(void)
435 if (afs_WaitForCacheDrain) {
436 if ((afs_blocksUsed - afs_blocksDiscarded) <=
437 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
438 afs_WaitForCacheDrain = 0;
439 afs_osi_Wakeup(&afs_WaitForCacheDrain);
445 * Keeps the cache clean and free by truncating uneeded files, when used.
450 afs_CacheTruncateDaemon(void)
452 osi_timeval_t CTD_tmpTime;
456 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
458 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
460 osi_GetuTime(&CTD_stats.CTD_afterSleep);
461 afs_TruncateDaemonRunning = 1;
463 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
464 ObtainWriteLock(&afs_xdcache, 266);
465 if (afs_CacheTooFull || afs_WaitForCacheDrain) {
466 int space_needed, slots_needed;
467 /* if we get woken up, we should try to clean something out */
468 for (counter = 0; counter < 10; counter++) {
470 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
471 if (space_needed < 0)
474 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
475 if (slots_needed < 0)
477 if (slots_needed || space_needed)
478 afs_GetDownD(slots_needed, &space_needed, 0);
479 if ((space_needed <= 0) && (slots_needed <= 0)) {
482 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
485 if (!afs_CacheIsTooFull()) {
486 afs_CacheTooFull = 0;
487 afs_WakeCacheWaitersIfDrained();
489 } /* end of cache cleanup */
490 ReleaseWriteLock(&afs_xdcache);
493 * This is a defensive check to try to avoid starving threads
494 * that may need the global lock so thay can help free some
495 * cache space. If this thread won't be sleeping or truncating
496 * any cache files then give up the global lock so other
497 * threads get a chance to run.
499 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
500 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
501 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
505 * This is where we free the discarded cache elements.
507 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
508 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
509 int code = afs_FreeDiscardedDCache();
511 /* If we can't free any discarded dcache entries, that's okay.
512 * We're just doing this in the background; if someone needs
513 * discarded entries freed, they will try it themselves and/or
514 * signal us that the cache is too full. In any case, we'll
515 * try doing this again the next time we run through the loop.
521 /* See if we need to continue to run. Someone may have
522 * signalled us while we were executing.
524 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
525 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
526 /* Collect statistics on truncate daemon. */
527 CTD_stats.CTD_nSleeps++;
528 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
529 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
530 CTD_stats.CTD_beforeSleep);
531 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
533 afs_TruncateDaemonRunning = 0;
534 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
535 afs_TruncateDaemonRunning = 1;
537 osi_GetuTime(&CTD_stats.CTD_afterSleep);
538 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
539 CTD_stats.CTD_afterSleep);
540 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
542 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
543 afs_termState = AFSOP_STOP_AFSDB;
544 afs_osi_Wakeup(&afs_termState);
552 * Make adjustment for the new size in the disk cache entry
554 * \note Major Assumptions Here:
555 * Assumes that frag size is an integral power of two, less one,
556 * and that this is a two's complement machine. I don't
557 * know of any filesystems which violate this assumption...
559 * \param adc Ptr to dcache entry.
560 * \param anewsize New size desired.
565 afs_AdjustSize(struct dcache *adc, afs_int32 newSize)
569 AFS_STATCNT(afs_AdjustSize);
571 if (newSize > afs_OtherCSize && !(adc->f.fid.Fid.Vnode & 1)) {
572 /* No non-dir cache files should be larger than the chunk size.
573 * (Directory blobs are fetched in a single chunk file, so directories
574 * can be larger.) If someone is requesting that a chunk is larger than
575 * the chunk size, something strange is happening. Log a message about
576 * it, to give a hint to subsequent strange behavior, if any occurs. */
580 afs_warn("afs: Warning: dcache %d is very large (%d > %d). This "
581 "should not happen, but trying to continue regardless. If "
582 "AFS starts hanging or behaving strangely, this might be "
584 adc->index, newSize, afs_OtherCSize);
588 adc->dflags |= DFEntryMod;
589 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
590 adc->f.chunkBytes = newSize;
593 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
594 afs_DCAdjustSize(adc, oldSize, newSize);
595 if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {
597 /* We're growing the file, wakeup the daemon */
598 afs_MaybeWakeupTruncateDaemon();
600 afs_blocksUsed += (newSize - oldSize);
601 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
606 * This routine is responsible for moving at least one entry (but up
607 * to some number of them) from the LRU queue to the free queue.
609 * \param anumber Number of entries that should ideally be moved.
610 * \param aneedSpace How much space we need (1K blocks);
613 * The anumber parameter is just a hint; at least one entry MUST be
614 * moved, or we'll panic. We must be called with afs_xdcache
615 * write-locked. We should try to satisfy both anumber and aneedspace,
616 * whichever is more demanding - need to do several things:
617 * 1. only grab up to anumber victims if aneedSpace <= 0, not
618 * the whole set of MAXATONCE.
619 * 2. dynamically choose MAXATONCE to reflect severity of
620 * demand: something like (*aneedSpace >> (logChunk - 9))
622 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
623 * indicates that the cache is not properly configured/tuned or
624 * something. We should be able to automatically correct that problem.
627 #define MAXATONCE 16 /* max we can obtain at once */
629 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
633 struct VenusFid *afid;
638 afs_uint32 victims[MAXATONCE];
639 struct dcache *victimDCs[MAXATONCE];
640 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
641 afs_uint32 victimPtr; /* next free item in victim arrays */
642 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
643 afs_uint32 maxVictimPtr; /* where it is */
647 AFS_STATCNT(afs_GetDownD);
649 if (CheckLock(&afs_xdcache) != -1)
650 osi_Panic("getdownd nolock");
651 /* decrement anumber first for all dudes in free list */
652 /* SHOULD always decrement anumber first, even if aneedSpace >0,
653 * because we should try to free space even if anumber <=0 */
654 if (!aneedSpace || *aneedSpace <= 0) {
655 anumber -= afs_freeDCCount;
657 return; /* enough already free */
661 /* bounds check parameter */
662 if (anumber > MAXATONCE)
663 anumber = MAXATONCE; /* all we can do */
665 /* rewrite so phases include a better eligiblity for gc test*/
667 * The phase variable manages reclaims. Set to 0, the first pass,
668 * we don't reclaim active entries, or other than target bucket.
669 * Set to 1, we reclaim even active ones in target bucket.
670 * Set to 2, we reclaim any inactive one.
671 * Set to 3, we reclaim even active ones. On Solaris, we also reclaim
672 * entries whose corresponding vcache has a nonempty multiPage list, when
681 for (i = 0; i < afs_cacheFiles; i++)
682 /* turn off all flags */
683 afs_indexFlags[i] &= ~IFFlag;
685 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
686 /* find oldest entries for reclamation */
687 maxVictimPtr = victimPtr = 0;
688 hzero(maxVictimTime);
689 curbucket = afs_DCWhichBucket(phase, buckethint);
690 /* select victims from access time array */
691 for (i = 0; i < afs_cacheFiles; i++) {
692 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
693 /* skip if dirty or already free */
696 tdc = afs_indexTable[i];
697 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
699 /* Wrong bucket; can't use it! */
702 if (tdc && (tdc->refCount != 0)) {
703 /* Referenced; can't use it! */
706 hset(vtime, afs_indexTimes[i]);
708 /* if we've already looked at this one, skip it */
709 if (afs_indexFlags[i] & IFFlag)
712 if (victimPtr < MAXATONCE) {
713 /* if there's at least one free victim slot left */
714 victims[victimPtr] = i;
715 hset(victimTimes[victimPtr], vtime);
716 if (hcmp(vtime, maxVictimTime) > 0) {
717 hset(maxVictimTime, vtime);
718 maxVictimPtr = victimPtr;
721 } else if (hcmp(vtime, maxVictimTime) < 0) {
723 * We're older than youngest victim, so we replace at
726 /* find youngest (largest LRU) victim */
729 osi_Panic("getdownd local");
731 hset(victimTimes[j], vtime);
732 /* recompute maxVictimTime */
733 hset(maxVictimTime, vtime);
734 for (j = 0; j < victimPtr; j++)
735 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
736 hset(maxVictimTime, victimTimes[j]);
742 /* now really reclaim the victims */
743 j = 0; /* flag to track if we actually got any of the victims */
744 /* first, hold all the victims, since we're going to release the lock
745 * during the truncate operation.
747 for (i = 0; i < victimPtr; i++) {
748 tdc = afs_GetValidDSlot(victims[i]);
749 /* We got tdc->tlock(R) here */
750 if (tdc && tdc->refCount == 1)
755 ReleaseReadLock(&tdc->tlock);
760 for (i = 0; i < victimPtr; i++) {
761 /* q is first elt in dcache entry */
763 /* now, since we're dropping the afs_xdcache lock below, we
764 * have to verify, before proceeding, that there are no other
765 * references to this dcache entry, even now. Note that we
766 * compare with 1, since we bumped it above when we called
767 * afs_GetValidDSlot to preserve the entry's identity.
769 if (tdc && tdc->refCount == 1) {
770 unsigned char chunkFlags;
771 afs_size_t tchunkoffset = 0;
773 /* xdcache is lower than the xvcache lock */
774 ReleaseWriteLock(&afs_xdcache);
775 ObtainReadLock(&afs_xvcache);
776 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
777 ReleaseReadLock(&afs_xvcache);
778 ObtainWriteLock(&afs_xdcache, 527);
780 if (tdc->refCount > 1)
783 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
784 chunkFlags = afs_indexFlags[tdc->index];
785 if (((phase & 1) == 0) && osi_Active(tvc))
787 if (((phase & 1) == 1) && osi_Active(tvc)
788 && (tvc->f.states & CDCLock)
789 && (chunkFlags & IFAnyPages))
791 if (chunkFlags & IFDataMod)
793 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
794 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
795 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
796 ICL_HANDLE_OFFSET(tchunkoffset));
798 #if defined(AFS_SUN5_ENV)
800 * Now we try to invalidate pages. We do this only for
801 * Solaris. For other platforms, it's OK to recycle a
802 * dcache entry out from under a page, because the strategy
803 * function can call afs_GetDCache().
805 if (!skip && (chunkFlags & IFAnyPages)) {
808 ReleaseWriteLock(&afs_xdcache);
809 ObtainWriteLock(&tvc->vlock, 543);
810 if (!QEmpty(&tvc->multiPage)) {
811 if (phase < 3 || osi_VM_MultiPageConflict(tvc, tdc)) {
816 /* block locking pages */
817 tvc->vstates |= VPageCleaning;
818 /* block getting new pages */
820 ReleaseWriteLock(&tvc->vlock);
821 /* One last recheck */
822 ObtainWriteLock(&afs_xdcache, 333);
823 chunkFlags = afs_indexFlags[tdc->index];
824 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
825 || (osi_Active(tvc) && (tvc->f.states & CDCLock)
826 && (chunkFlags & IFAnyPages))) {
828 ReleaseWriteLock(&afs_xdcache);
831 ReleaseWriteLock(&afs_xdcache);
833 code = osi_VM_GetDownD(tvc, tdc);
835 ObtainWriteLock(&afs_xdcache, 269);
836 /* we actually removed all pages, clean and dirty */
838 afs_indexFlags[tdc->index] &=
839 ~(IFDirtyPages | IFAnyPages);
842 ReleaseWriteLock(&afs_xdcache);
844 ObtainWriteLock(&tvc->vlock, 544);
845 if (--tvc->activeV == 0
846 && (tvc->vstates & VRevokeWait)) {
847 tvc->vstates &= ~VRevokeWait;
848 afs_osi_Wakeup((char *)&tvc->vstates);
851 if (tvc->vstates & VPageCleaning) {
852 tvc->vstates &= ~VPageCleaning;
853 afs_osi_Wakeup((char *)&tvc->vstates);
856 ReleaseWriteLock(&tvc->vlock);
858 #endif /* AFS_SUN5_ENV */
860 ReleaseWriteLock(&afs_xdcache);
863 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
864 ObtainWriteLock(&afs_xdcache, 528);
865 if (afs_indexFlags[tdc->index] &
866 (IFDataMod | IFDirtyPages | IFAnyPages))
868 if (tdc->refCount > 1)
871 #if defined(AFS_SUN5_ENV)
873 /* no vnode, so IFDirtyPages is spurious (we don't
874 * sweep dcaches on vnode recycling, so we can have
875 * DIRTYPAGES set even when all pages are gone). Just
877 * Hold vcache lock to prevent vnode from being
878 * created while we're clearing IFDirtyPages.
880 afs_indexFlags[tdc->index] &=
881 ~(IFDirtyPages | IFAnyPages);
885 /* skip this guy and mark him as recently used */
886 afs_indexFlags[tdc->index] |= IFFlag;
887 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
888 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
889 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
890 ICL_HANDLE_OFFSET(tchunkoffset));
892 /* flush this dude from the data cache and reclaim;
893 * first, make sure no one will care that we damage
894 * it, by removing it from all hash tables. Then,
895 * melt it down for parts. Note that any concurrent
896 * (new possibility!) calls to GetDownD won't touch
897 * this guy because his reference count is > 0. */
898 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
899 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
900 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
901 ICL_HANDLE_OFFSET(tchunkoffset));
902 AFS_STATCNT(afs_gget);
903 afs_HashOutDCache(tdc, 1);
904 if (tdc->f.chunkBytes != 0) {
908 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
913 afs_DiscardDCache(tdc);
918 j = 1; /* we reclaimed at least one victim */
923 } /* end of for victims loop */
926 /* Phase is 0 and no one was found, so try phase 1 (ignore
927 * osi_Active flag) */
930 for (i = 0; i < afs_cacheFiles; i++)
931 /* turn off all flags */
932 afs_indexFlags[i] &= ~IFFlag;
935 /* found no one in phases 0-5, we're hosed */
939 } /* big while loop */
947 * Remove adc from any hash tables that would allow it to be located
948 * again by afs_FindDCache or afs_GetDCache.
950 * \param adc Pointer to dcache entry to remove from hash tables.
952 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
956 afs_HashOutDCache(struct dcache *adc, int zap)
960 AFS_STATCNT(afs_glink);
962 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
964 /* if this guy is in the hash table, pull him out */
965 if (adc->f.fid.Fid.Volume != 0) {
966 /* remove entry from first hash chains */
967 i = DCHash(&adc->f.fid, adc->f.chunk);
968 us = afs_dchashTbl[i];
969 if (us == adc->index) {
970 /* first dude in the list */
971 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
973 /* somewhere on the chain */
974 while (us != NULLIDX) {
975 if (afs_dcnextTbl[us] == adc->index) {
976 /* found item pointing at the one to delete */
977 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
980 us = afs_dcnextTbl[us];
983 osi_Panic("dcache hc");
985 /* remove entry from *other* hash chain */
986 i = DVHash(&adc->f.fid);
987 us = afs_dvhashTbl[i];
988 if (us == adc->index) {
989 /* first dude in the list */
990 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
992 /* somewhere on the chain */
993 while (us != NULLIDX) {
994 if (afs_dvnextTbl[us] == adc->index) {
995 /* found item pointing at the one to delete */
996 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
999 us = afs_dvnextTbl[us];
1002 osi_Panic("dcache hv");
1007 /* prevent entry from being found on a reboot (it is already out of
1008 * the hash table, but after a crash, we just look at fid fields of
1009 * stable (old) entries).
1011 adc->f.fid.Fid.Volume = 0; /* invalid */
1013 /* mark entry as modified */
1014 adc->dflags |= DFEntryMod;
1019 } /*afs_HashOutDCache */
1022 * Flush the given dcache entry, pulling it from hash chains
1023 * and truncating the associated cache file.
1025 * \param adc Ptr to dcache entry to flush.
1027 * \note Environment:
1028 * This routine must be called with the afs_xdcache lock held
1032 afs_FlushDCache(struct dcache *adc)
1034 AFS_STATCNT(afs_FlushDCache);
1036 * Bump the number of cache files flushed.
1038 afs_stats_cmperf.cacheFlushes++;
1040 /* remove from all hash tables */
1041 afs_HashOutDCache(adc, 1);
1043 /* Free its space; special case null operation, since truncate operation
1044 * in UFS is slow even in this case, and this allows us to pre-truncate
1045 * these files at more convenient times with fewer locks set
1046 * (see afs_GetDownD).
1048 if (adc->f.chunkBytes != 0) {
1049 afs_DiscardDCache(adc);
1050 afs_MaybeWakeupTruncateDaemon();
1052 afs_FreeDCache(adc);
1054 } /*afs_FlushDCache */
1058 * Put a dcache entry on the free dcache entry list.
1060 * \param adc dcache entry to free.
1062 * \note Environment: called with afs_xdcache lock write-locked.
1065 afs_FreeDCache(struct dcache *adc)
1067 /* Thread on free list, update free list count and mark entry as
1068 * freed in its indexFlags element. Also, ensure DCache entry gets
1069 * written out (set DFEntryMod).
1072 afs_dvnextTbl[adc->index] = afs_freeDCList;
1073 afs_freeDCList = adc->index;
1075 afs_indexFlags[adc->index] |= IFFree;
1076 adc->dflags |= DFEntryMod;
1078 afs_WakeCacheWaitersIfDrained();
1079 } /* afs_FreeDCache */
1082 * Discard the cache element by moving it to the discardDCList.
1083 * This puts the cache element into a quasi-freed state, where
1084 * the space may be reused, but the file has not been truncated.
1086 * \note Major Assumptions Here:
1087 * Assumes that frag size is an integral power of two, less one,
1088 * and that this is a two's complement machine. I don't
1089 * know of any filesystems which violate this assumption...
1091 * \param adr Ptr to dcache entry.
1093 * \note Environment:
1094 * Must be called with afs_xdcache write-locked.
1098 afs_DiscardDCache(struct dcache *adc)
1102 AFS_STATCNT(afs_DiscardDCache);
1104 osi_Assert(adc->refCount == 1);
1106 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1107 afs_blocksDiscarded += size;
1108 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1110 afs_dvnextTbl[adc->index] = afs_discardDCList;
1111 afs_discardDCList = adc->index;
1112 afs_discardDCCount++;
1114 adc->f.fid.Fid.Volume = 0;
1115 adc->dflags |= DFEntryMod;
1116 afs_indexFlags[adc->index] |= IFDiscarded;
1118 afs_WakeCacheWaitersIfDrained();
1119 } /*afs_DiscardDCache */
1122 * Get a dcache entry from the discard or free list
1124 * @param[in] indexp A pointer to the head of the dcache free list or discard
1125 * list (afs_freeDCList, or afs_discardDCList)
1127 * @return A dcache from that list, or NULL if none could be retrieved.
1129 * @pre afs_xdcache is write-locked
1131 static struct dcache *
1132 afs_GetDSlotFromList(afs_int32 *indexp)
1136 for ( ; *indexp != NULLIDX; indexp = &afs_dvnextTbl[*indexp]) {
1137 tdc = afs_GetUnusedDSlot(*indexp);
1139 osi_Assert(tdc->refCount == 1);
1140 ReleaseReadLock(&tdc->tlock);
1141 *indexp = afs_dvnextTbl[tdc->index];
1142 afs_dvnextTbl[tdc->index] = NULLIDX;
1150 * Free the next element on the list of discarded cache elements.
1152 * Returns -1 if we encountered an error preventing us from freeing a
1153 * discarded dcache, or 0 on success.
1156 afs_FreeDiscardedDCache(void)
1159 struct osi_file *tfile;
1162 AFS_STATCNT(afs_FreeDiscardedDCache);
1164 ObtainWriteLock(&afs_xdcache, 510);
1165 if (!afs_blocksDiscarded) {
1166 ReleaseWriteLock(&afs_xdcache);
1171 * Get an entry from the list of discarded cache elements
1173 tdc = afs_GetDSlotFromList(&afs_discardDCList);
1175 ReleaseWriteLock(&afs_xdcache);
1179 afs_discardDCCount--;
1180 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1181 afs_blocksDiscarded -= size;
1182 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1183 /* We can lock because we just took it off the free list */
1184 ObtainWriteLock(&tdc->lock, 626);
1185 ReleaseWriteLock(&afs_xdcache);
1188 * Truncate the element to reclaim its space
1190 tfile = afs_CFileOpen(&tdc->f.inode);
1192 afs_CFileTruncate(tfile, 0);
1193 afs_CFileClose(tfile);
1194 afs_AdjustSize(tdc, 0);
1195 afs_DCMoveBucket(tdc, 0, 0);
1198 * Free the element we just truncated
1200 ObtainWriteLock(&afs_xdcache, 511);
1201 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1202 afs_FreeDCache(tdc);
1203 tdc->f.states &= ~(DRO|DBackup|DRW);
1204 ReleaseWriteLock(&tdc->lock);
1206 ReleaseWriteLock(&afs_xdcache);
1212 * Free as many entries from the list of discarded cache elements
1213 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1218 afs_MaybeFreeDiscardedDCache(void)
1221 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1223 while (afs_blocksDiscarded
1224 && (afs_blocksUsed >
1225 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1226 int code = afs_FreeDiscardedDCache();
1228 /* Callers depend on us to get the afs_blocksDiscarded count down.
1229 * If we cannot do that, the callers can spin by calling us over
1230 * and over. Panic for now until we can figure out something
1232 osi_Panic("Error freeing discarded dcache");
1239 * Try to free up a certain number of disk slots.
1241 * \param anumber Targeted number of disk slots to free up.
1243 * \note Environment:
1244 * Must be called with afs_xdcache write-locked.
1248 afs_GetDownDSlot(int anumber)
1250 struct afs_q *tq, *nq;
1255 AFS_STATCNT(afs_GetDownDSlot);
1256 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1257 osi_Panic("diskless getdowndslot");
1259 if (CheckLock(&afs_xdcache) != -1)
1260 osi_Panic("getdowndslot nolock");
1262 /* decrement anumber first for all dudes in free list */
1263 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1266 return; /* enough already free */
1268 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1270 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1271 nq = QPrev(tq); /* in case we remove it */
1272 if (tdc->refCount == 0) {
1273 if ((ix = tdc->index) == NULLIDX)
1274 osi_Panic("getdowndslot");
1275 /* pull the entry out of the lruq and put it on the free list */
1276 QRemove(&tdc->lruq);
1278 /* write-through if modified */
1279 if (tdc->dflags & DFEntryMod) {
1280 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1282 * ask proxy to do this for us - we don't have the stack space
1284 while (tdc->dflags & DFEntryMod) {
1287 s = SPLOCK(afs_sgibklock);
1288 if (afs_sgibklist == NULL) {
1289 /* if slot is free, grab it. */
1290 afs_sgibklist = tdc;
1291 SV_SIGNAL(&afs_sgibksync);
1293 /* wait for daemon to (start, then) finish. */
1294 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1298 tdc->dflags &= ~DFEntryMod;
1299 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
1303 /* finally put the entry in the free list */
1304 afs_indexTable[ix] = NULL;
1305 afs_indexFlags[ix] &= ~IFEverUsed;
1306 tdc->index = NULLIDX;
1307 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1308 afs_freeDSList = tdc;
1312 } /*afs_GetDownDSlot */
1319 * Increment the reference count on a disk cache entry,
1320 * which already has a non-zero refcount. In order to
1321 * increment the refcount of a zero-reference entry, you
1322 * have to hold afs_xdcache.
1325 * adc : Pointer to the dcache entry to increment.
1328 * Nothing interesting.
1331 afs_RefDCache(struct dcache *adc)
1333 ObtainWriteLock(&adc->tlock, 627);
1334 if (adc->refCount < 0)
1335 osi_Panic("RefDCache: negative refcount");
1337 ReleaseWriteLock(&adc->tlock);
1346 * Decrement the reference count on a disk cache entry.
1349 * ad : Ptr to the dcache entry to decrement.
1352 * Nothing interesting.
1355 afs_PutDCache(struct dcache *adc)
1357 AFS_STATCNT(afs_PutDCache);
1358 ObtainWriteLock(&adc->tlock, 276);
1359 if (adc->refCount <= 0)
1360 osi_Panic("putdcache");
1362 ReleaseWriteLock(&adc->tlock);
1371 * Try to discard all data associated with this file from the
1375 * avc : Pointer to the cache info for the file.
1378 * Both pvnLock and lock are write held.
1381 afs_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
1386 AFS_STATCNT(afs_TryToSmush);
1387 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1388 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
1389 sync = 1; /* XX Temp testing XX */
1391 #if defined(AFS_SUN5_ENV)
1392 ObtainWriteLock(&avc->vlock, 573);
1393 avc->activeV++; /* block new getpages */
1394 ReleaseWriteLock(&avc->vlock);
1397 /* Flush VM pages */
1398 osi_VM_TryToSmush(avc, acred, sync);
1401 * Get the hash chain containing all dce's for this fid
1403 i = DVHash(&avc->f.fid);
1404 ObtainWriteLock(&afs_xdcache, 277);
1405 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1406 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1407 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1408 int releaseTlock = 1;
1409 tdc = afs_GetValidDSlot(index);
1411 /* afs_TryToSmush is best-effort; we may not actually discard
1412 * everything, so failure to discard a dcache due to an i/o
1416 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1418 if ((afs_indexFlags[index] & IFDataMod) == 0
1419 && tdc->refCount == 1) {
1420 ReleaseReadLock(&tdc->tlock);
1422 afs_FlushDCache(tdc);
1425 afs_indexTable[index] = 0;
1428 ReleaseReadLock(&tdc->tlock);
1432 #if defined(AFS_SUN5_ENV)
1433 ObtainWriteLock(&avc->vlock, 545);
1434 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1435 avc->vstates &= ~VRevokeWait;
1436 afs_osi_Wakeup((char *)&avc->vstates);
1438 ReleaseWriteLock(&avc->vlock);
1440 ReleaseWriteLock(&afs_xdcache);
1442 * It's treated like a callback so that when we do lookups we'll
1443 * invalidate the unique bit if any
1444 * trytoSmush occured during the lookup call
1450 * afs_DCacheMissingChunks
1453 * Given the cached info for a file, return the number of chunks that
1454 * are not available from the dcache.
1457 * avc: Pointer to the (held) vcache entry to look in.
1460 * The number of chunks which are not currently cached.
1463 * The vcache entry is held upon entry.
1467 afs_DCacheMissingChunks(struct vcache *avc)
1470 afs_size_t totalLength = 0;
1471 afs_uint32 totalChunks = 0;
1474 totalLength = avc->f.m.Length;
1475 if (avc->f.truncPos < totalLength)
1476 totalLength = avc->f.truncPos;
1478 /* Length is 0, no chunk missing. */
1479 if (totalLength == 0)
1482 /* If totalLength is a multiple of chunksize, the last byte appears
1483 * as being part of the next chunk, which does not exist.
1484 * Decrementing totalLength by one fixes that.
1487 totalChunks = (AFS_CHUNK(totalLength) + 1);
1489 /* If we're a directory, we only ever have one chunk, regardless of
1490 * the size of the dir.
1492 if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
1496 printf("Should have %d chunks for %u bytes\n",
1497 totalChunks, (totalLength + 1));
1499 i = DVHash(&avc->f.fid);
1500 ObtainWriteLock(&afs_xdcache, 1001);
1501 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1502 i = afs_dvnextTbl[index];
1503 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1504 tdc = afs_GetValidDSlot(index);
1506 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1509 ReleaseReadLock(&tdc->tlock);
1514 ReleaseWriteLock(&afs_xdcache);
1516 /*printf("Missing %d chunks\n", totalChunks);*/
1518 return (totalChunks);
1525 * Given the cached info for a file and a byte offset into the
1526 * file, make sure the dcache entry for that file and containing
1527 * the given byte is available, returning it to our caller.
1530 * avc : Pointer to the (held) vcache entry to look in.
1531 * abyte : Which byte we want to get to.
1534 * Pointer to the dcache entry covering the file & desired byte,
1535 * or NULL if not found.
1538 * The vcache entry is held upon entry.
1542 afs_FindDCache(struct vcache *avc, afs_size_t abyte)
1546 struct dcache *tdc = NULL;
1548 AFS_STATCNT(afs_FindDCache);
1549 chunk = AFS_CHUNK(abyte);
1552 * Hash on the [fid, chunk] and get the corresponding dcache index
1553 * after write-locking the dcache.
1555 i = DCHash(&avc->f.fid, chunk);
1556 ObtainWriteLock(&afs_xdcache, 278);
1557 for (index = afs_dchashTbl[i]; index != NULLIDX; index = afs_dcnextTbl[index]) {
1558 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1559 tdc = afs_GetValidDSlot(index);
1561 /* afs_FindDCache is best-effort; we may not find the given
1562 * file/offset, so if we cannot find the given dcache due to
1563 * i/o errors, that is okay. */
1566 ReleaseReadLock(&tdc->tlock);
1567 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1568 break; /* leaving refCount high for caller */
1573 if (index != NULLIDX) {
1574 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1575 hadd32(afs_indexCounter, 1);
1576 ReleaseWriteLock(&afs_xdcache);
1579 ReleaseWriteLock(&afs_xdcache);
1581 } /*afs_FindDCache */
1583 /* only call these from afs_AllocDCache() */
1584 static struct dcache *
1585 afs_AllocFreeDSlot(void)
1589 tdc = afs_GetDSlotFromList(&afs_freeDCList);
1593 afs_indexFlags[tdc->index] &= ~IFFree;
1594 ObtainWriteLock(&tdc->lock, 604);
1599 static struct dcache *
1600 afs_AllocDiscardDSlot(afs_int32 lock)
1603 afs_uint32 size = 0;
1604 struct osi_file *file;
1606 tdc = afs_GetDSlotFromList(&afs_discardDCList);
1610 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1611 ObtainWriteLock(&tdc->lock, 605);
1612 afs_discardDCCount--;
1614 ((tdc->f.chunkBytes +
1615 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1616 tdc->f.states &= ~(DRO|DBackup|DRW);
1617 afs_DCMoveBucket(tdc, size, 0);
1618 afs_blocksDiscarded -= size;
1619 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1621 /* Truncate the chunk so zeroes get filled properly */
1622 file = afs_CFileOpen(&tdc->f.inode);
1624 afs_CFileTruncate(file, 0);
1625 afs_CFileClose(file);
1626 afs_AdjustSize(tdc, 0);
1633 * Get a fresh dcache from the free or discarded list.
1635 * \param avc Who's dcache is this going to be?
1636 * \param chunk The position where it will be placed in.
1637 * \param lock How are locks held.
1638 * \param ashFid If this dcache going to be used for a shadow dir,
1641 * \note Required locks:
1643 * - avc (R if (lock & 1) set and W otherwise)
1644 * \note It write locks the new dcache. The caller must unlock it.
1646 * \return The new dcache.
1649 afs_AllocDCache(struct vcache *avc, afs_int32 chunk, afs_int32 lock,
1650 struct VenusFid *ashFid)
1652 struct dcache *tdc = NULL;
1654 /* if (lock & 2), prefer 'free' dcaches; otherwise, prefer 'discard'
1655 * dcaches. In either case, try both if our first choice doesn't work. */
1657 tdc = afs_AllocFreeDSlot();
1659 tdc = afs_AllocDiscardDSlot(lock);
1662 tdc = afs_AllocDiscardDSlot(lock);
1664 tdc = afs_AllocFreeDSlot();
1673 * avc->lock(R) if setLocks
1674 * avc->lock(W) if !setLocks
1680 * Fill in the newly-allocated dcache record.
1682 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1684 /* Use shadow fid if provided. */
1685 tdc->f.fid = *ashFid;
1687 /* Use normal vcache's fid otherwise. */
1688 tdc->f.fid = avc->f.fid;
1689 if (avc->f.states & CRO)
1690 tdc->f.states = DRO;
1691 else if (avc->f.states & CBackup)
1692 tdc->f.states = DBackup;
1694 tdc->f.states = DRW;
1695 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1696 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1698 hones(tdc->f.versionNo); /* invalid value */
1699 tdc->f.chunk = chunk;
1700 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1702 if (tdc->lruq.prev == &tdc->lruq)
1703 osi_Panic("lruq 1");
1712 * This function is called to obtain a reference to data stored in
1713 * the disk cache, locating a chunk of data containing the desired
1714 * byte and returning a reference to the disk cache entry, with its
1715 * reference count incremented.
1719 * avc : Ptr to a vcache entry (unlocked)
1720 * abyte : Byte position in the file desired
1721 * areq : Request structure identifying the requesting user.
1722 * aflags : Settings as follows:
1724 * 2 : Return after creating entry.
1725 * 4 : called from afs_vnop_write.c
1726 * *alen contains length of data to be written.
1728 * aoffset : Set to the offset within the chunk where the resident
1730 * alen : Set to the number of bytes of data after the desired
1731 * byte (including the byte itself) which can be read
1735 * The vcache entry pointed to by avc is unlocked upon entry.
1739 * Update the vnode-to-dcache hint if we can get the vnode lock
1740 * right away. Assumes dcache entry is at least read-locked.
1743 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1745 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1746 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1749 ReleaseWriteLock(&v->lock);
1753 /* avc - Write-locked unless aflags & 1 */
1755 afs_GetDCache(struct vcache *avc, afs_size_t abyte,
1756 struct vrequest *areq, afs_size_t * aoffset,
1757 afs_size_t * alen, int aflags)
1759 afs_int32 i, code, shortcut;
1760 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1761 afs_int32 adjustsize = 0;
1767 afs_size_t Position = 0;
1768 afs_int32 size, tlen; /* size of segment to transfer */
1769 struct afs_FetchOutput *tsmall = 0;
1771 struct osi_file *file;
1772 struct afs_conn *tc;
1774 struct server *newCallback = NULL;
1775 char setNewCallback;
1776 char setVcacheStatus;
1777 char doVcacheUpdate;
1779 int doAdjustSize = 0;
1780 int doReallyAdjustSize = 0;
1781 int overWriteWholeChunk = 0;
1782 struct rx_connection *rxconn;
1785 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1786 int fromReplica; /*Are we reading from a replica? */
1787 int numFetchLoops; /*# times around the fetch/analyze loop */
1788 #endif /* AFS_NOSTATS */
1790 AFS_STATCNT(afs_GetDCache);
1794 setLocks = aflags & 1;
1797 * Determine the chunk number and offset within the chunk corresponding
1798 * to the desired byte.
1800 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1803 chunk = AFS_CHUNK(abyte);
1806 /* come back to here if we waited for the cache to drain. */
1809 setNewCallback = setVcacheStatus = 0;
1813 ObtainWriteLock(&avc->lock, 616);
1815 ObtainReadLock(&avc->lock);
1820 * avc->lock(R) if setLocks && !slowPass
1821 * avc->lock(W) if !setLocks || slowPass
1826 /* check hints first! (might could use bcmp or some such...) */
1827 if ((tdc = avc->dchint)) {
1831 * The locking order between afs_xdcache and dcache lock matters.
1832 * The hint dcache entry could be anywhere, even on the free list.
1833 * Locking afs_xdcache ensures that noone is trying to pull dcache
1834 * entries from the free list, and thereby assuming them to be not
1835 * referenced and not locked.
1837 ObtainReadLock(&afs_xdcache);
1838 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1840 if (dcLocked && (tdc->index != NULLIDX)
1841 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1842 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1843 /* got the right one. It might not be the right version, and it
1844 * might be fetching, but it's the right dcache entry.
1846 /* All this code should be integrated better with what follows:
1847 * I can save a good bit more time under a write lock if I do..
1849 ObtainWriteLock(&tdc->tlock, 603);
1851 ReleaseWriteLock(&tdc->tlock);
1853 ReleaseReadLock(&afs_xdcache);
1856 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1857 && !(tdc->dflags & DFFetching)) {
1859 afs_stats_cmperf.dcacheHits++;
1860 ObtainWriteLock(&afs_xdcache, 559);
1861 QRemove(&tdc->lruq);
1862 QAdd(&afs_DLRU, &tdc->lruq);
1863 ReleaseWriteLock(&afs_xdcache);
1866 * avc->lock(R) if setLocks && !slowPass
1867 * avc->lock(W) if !setLocks || slowPass
1874 ReleaseSharedLock(&tdc->lock);
1875 ReleaseReadLock(&afs_xdcache);
1883 * avc->lock(R) if setLocks && !slowPass
1884 * avc->lock(W) if !setLocks || slowPass
1885 * tdc->lock(S) if tdc
1888 if (!tdc) { /* If the hint wasn't the right dcache entry */
1889 int dslot_error = 0;
1891 * Hash on the [fid, chunk] and get the corresponding dcache index
1892 * after write-locking the dcache.
1897 * avc->lock(R) if setLocks && !slowPass
1898 * avc->lock(W) if !setLocks || slowPass
1901 i = DCHash(&avc->f.fid, chunk);
1902 /* check to make sure our space is fine */
1903 afs_MaybeWakeupTruncateDaemon();
1905 ObtainWriteLock(&afs_xdcache, 280);
1907 for (index = afs_dchashTbl[i]; index != NULLIDX; us = index, index = afs_dcnextTbl[index]) {
1908 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1909 tdc = afs_GetValidDSlot(index);
1911 /* we got an i/o error when trying to get the given dslot,
1912 * but do not bail out just yet; it is possible the dcache
1913 * we're looking for is elsewhere, so it doesn't matter if
1914 * we can't load this one. */
1918 ReleaseReadLock(&tdc->tlock);
1921 * avc->lock(R) if setLocks && !slowPass
1922 * avc->lock(W) if !setLocks || slowPass
1925 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1926 /* Move it up in the beginning of the list */
1927 if (afs_dchashTbl[i] != index) {
1928 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1929 afs_dcnextTbl[index] = afs_dchashTbl[i];
1930 afs_dchashTbl[i] = index;
1932 ReleaseWriteLock(&afs_xdcache);
1933 ObtainSharedLock(&tdc->lock, 606);
1934 break; /* leaving refCount high for caller */
1942 * If we didn't find the entry, we'll create one.
1944 if (index == NULLIDX) {
1947 * avc->lock(R) if setLocks
1948 * avc->lock(W) if !setLocks
1951 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1952 avc, ICL_TYPE_INT32, chunk);
1955 /* We couldn't find the dcache we want, but we hit some i/o
1956 * errors when trying to find it, so we're not sure if the
1957 * dcache we want is in the cache or not. Error out, so we
1958 * don't try to possibly create 2 separate dcaches for the
1959 * same exact data. */
1960 ReleaseWriteLock(&afs_xdcache);
1964 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1966 avc->f.states |= CDCLock;
1967 /* just need slots */
1968 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1970 avc->f.states &= ~CDCLock;
1972 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1974 /* If we can't get space for 5 mins we give up and panic */
1975 if (++downDCount > 300)
1976 osi_Panic("getdcache");
1977 ReleaseWriteLock(&afs_xdcache);
1980 * avc->lock(R) if setLocks
1981 * avc->lock(W) if !setLocks
1983 afs_osi_Wait(1000, 0, 0);
1989 * avc->lock(R) if setLocks
1990 * avc->lock(W) if !setLocks
1996 * Now add to the two hash chains - note that i is still set
1997 * from the above DCHash call.
1999 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
2000 afs_dchashTbl[i] = tdc->index;
2001 i = DVHash(&avc->f.fid);
2002 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
2003 afs_dvhashTbl[i] = tdc->index;
2004 tdc->dflags = DFEntryMod;
2006 afs_MaybeWakeupTruncateDaemon();
2007 ReleaseWriteLock(&afs_xdcache);
2008 ConvertWToSLock(&tdc->lock);
2013 /* vcache->dcache hint failed */
2016 * avc->lock(R) if setLocks && !slowPass
2017 * avc->lock(W) if !setLocks || slowPass
2020 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
2021 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2022 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
2023 hgetlo(avc->f.m.DataVersion));
2025 * Here we have the entry in tdc, with its refCount incremented.
2026 * Note: we don't use the S-lock on avc; it costs concurrency when
2027 * storing a file back to the server.
2031 * Not a newly created file so we need to check the file's length and
2032 * compare data versions since someone could have changed the data or we're
2033 * reading a file written elsewhere. We only want to bypass doing no-op
2034 * read rpcs on newly created files (dv of 0) since only then we guarantee
2035 * that this chunk's data hasn't been filled by another client.
2037 size = AFS_CHUNKSIZE(abyte);
2038 if (aflags & 4) /* called from write */
2040 else /* called from read */
2041 tlen = tdc->validPos - abyte;
2042 Position = AFS_CHUNKTOBASE(chunk);
2043 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
2044 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
2045 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
2046 ICL_HANDLE_OFFSET(Position));
2047 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
2049 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
2050 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
2051 overWriteWholeChunk = 1;
2052 if (doAdjustSize || overWriteWholeChunk) {
2053 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
2055 #ifdef AFS_SGI64_ENV
2058 #else /* AFS_SGI64_ENV */
2061 #endif /* AFS_SGI64_ENV */
2062 #else /* AFS_SGI_ENV */
2065 #endif /* AFS_SGI_ENV */
2066 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
2067 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2068 #if defined(AFS_SUN5_ENV)
2069 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
2071 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
2073 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2074 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
2075 doReallyAdjustSize = 1;
2077 if (doReallyAdjustSize || overWriteWholeChunk) {
2078 /* no data in file to read at this position */
2079 UpgradeSToWLock(&tdc->lock, 607);
2080 file = afs_CFileOpen(&tdc->f.inode);
2082 afs_CFileTruncate(file, 0);
2083 afs_CFileClose(file);
2084 afs_AdjustSize(tdc, 0);
2085 hset(tdc->f.versionNo, avc->f.m.DataVersion);
2086 tdc->dflags |= DFEntryMod;
2088 ConvertWToSLock(&tdc->lock);
2093 * We must read in the whole chunk if the version number doesn't
2097 /* don't need data, just a unique dcache entry */
2098 ObtainWriteLock(&afs_xdcache, 608);
2099 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2100 hadd32(afs_indexCounter, 1);
2101 ReleaseWriteLock(&afs_xdcache);
2103 updateV2DC(setLocks, avc, tdc, 553);
2104 if (vType(avc) == VDIR)
2107 *aoffset = AFS_CHUNKOFFSET(abyte);
2108 if (tdc->validPos < abyte)
2109 *alen = (afs_size_t) 0;
2111 *alen = tdc->validPos - abyte;
2112 ReleaseSharedLock(&tdc->lock);
2115 ReleaseWriteLock(&avc->lock);
2117 ReleaseReadLock(&avc->lock);
2119 return tdc; /* check if we're done */
2124 * avc->lock(R) if setLocks && !slowPass
2125 * avc->lock(W) if !setLocks || slowPass
2128 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2130 setNewCallback = setVcacheStatus = 0;
2134 * avc->lock(R) if setLocks && !slowPass
2135 * avc->lock(W) if !setLocks || slowPass
2138 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2140 * Version number mismatch.
2143 * If we are disconnected, then we can't do much of anything
2144 * because the data doesn't match the file.
2146 if (AFS_IS_DISCONNECTED) {
2147 ReleaseSharedLock(&tdc->lock);
2150 ReleaseWriteLock(&avc->lock);
2152 ReleaseReadLock(&avc->lock);
2154 /* Flush the Dcache */
2159 UpgradeSToWLock(&tdc->lock, 609);
2162 * If data ever existed for this vnode, and this is a text object,
2163 * do some clearing. Now, you'd think you need only do the flush
2164 * when VTEXT is on, but VTEXT is turned off when the text object
2165 * is freed, while pages are left lying around in memory marked
2166 * with this vnode. If we would reactivate (create a new text
2167 * object from) this vnode, we could easily stumble upon some of
2168 * these old pages in pagein. So, we always flush these guys.
2169 * Sun has a wonderful lack of useful invariants in this system.
2171 * avc->flushDV is the data version # of the file at the last text
2172 * flush. Clearly, at least, we don't have to flush the file more
2173 * often than it changes
2175 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2177 * By here, the cache entry is always write-locked. We can
2178 * deadlock if we call osi_Flush with the cache entry locked...
2179 * Unlock the dcache too.
2181 ReleaseWriteLock(&tdc->lock);
2182 if (setLocks && !slowPass)
2183 ReleaseReadLock(&avc->lock);
2185 ReleaseWriteLock(&avc->lock);
2189 * Call osi_FlushPages in open, read/write, and map, since it
2190 * is too hard here to figure out if we should lock the
2193 if (setLocks && !slowPass)
2194 ObtainReadLock(&avc->lock);
2196 ObtainWriteLock(&avc->lock, 66);
2197 ObtainWriteLock(&tdc->lock, 610);
2202 * avc->lock(R) if setLocks && !slowPass
2203 * avc->lock(W) if !setLocks || slowPass
2207 /* Watch for standard race condition around osi_FlushText */
2208 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2209 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2210 afs_stats_cmperf.dcacheHits++;
2211 ConvertWToSLock(&tdc->lock);
2215 /* Sleep here when cache needs to be drained. */
2216 if (setLocks && !slowPass
2217 && (afs_blocksUsed >
2218 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2219 /* Make sure truncate daemon is running */
2220 afs_MaybeWakeupTruncateDaemon();
2221 ObtainWriteLock(&tdc->tlock, 614);
2222 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2223 ReleaseWriteLock(&tdc->tlock);
2224 ReleaseWriteLock(&tdc->lock);
2225 ReleaseReadLock(&avc->lock);
2226 while ((afs_blocksUsed - afs_blocksDiscarded) >
2227 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2228 afs_WaitForCacheDrain = 1;
2229 afs_osi_Sleep(&afs_WaitForCacheDrain);
2231 afs_MaybeFreeDiscardedDCache();
2232 /* need to check if someone else got the chunk first. */
2233 goto RetryGetDCache;
2236 Position = AFS_CHUNKBASE(abyte);
2237 if (vType(avc) == VDIR) {
2238 size = avc->f.m.Length;
2239 if (size > tdc->f.chunkBytes) {
2240 /* pre-reserve space for file */
2241 afs_AdjustSize(tdc, size);
2243 size = 999999999; /* max size for transfer */
2245 afs_size_t maxGoodLength;
2247 /* estimate how much data we're expecting back from the server,
2248 * and reserve space in the dcache entry for it */
2250 maxGoodLength = avc->f.m.Length;
2251 if (avc->f.truncPos < maxGoodLength)
2252 maxGoodLength = avc->f.truncPos;
2254 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2255 if (Position > maxGoodLength) { /* If we're beyond EOF */
2257 } else if (Position + size > maxGoodLength) {
2258 size = maxGoodLength - Position;
2260 osi_Assert(size >= 0);
2262 if (size > tdc->f.chunkBytes) {
2263 /* pre-reserve estimated space for file */
2264 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2268 /* For the actual fetch, do not limit the request to the
2269 * length of the file. If this results in a read past EOF on
2270 * the server, the server will just reply with less data than
2271 * requested. If we limit ourselves to only requesting data up
2272 * to the avc file length, we open ourselves up to races if the
2273 * file is extended on the server at about the same time.
2275 * However, we must restrict ourselves to the avc->f.truncPos
2276 * length, since this represents an outstanding local
2277 * truncation of the file that will be committed to the
2278 * fileserver when we actually write the fileserver contents.
2279 * If we do not restrict the fetch length based on
2280 * avc->f.truncPos, a different truncate operation extending
2281 * the file length could cause the old data after
2282 * avc->f.truncPos to reappear, instead of extending the file
2283 * with NUL bytes. */
2284 size = AFS_CHUNKSIZE(abyte);
2285 if (Position > avc->f.truncPos) {
2287 } else if (Position + size > avc->f.truncPos) {
2288 size = avc->f.truncPos - Position;
2290 osi_Assert(size >= 0);
2293 if (afs_mariner && !tdc->f.chunk)
2294 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2296 * Right now, we only have one tool, and it's a hammer. So, we
2297 * fetch the whole file.
2299 DZap(tdc); /* pages in cache may be old */
2300 file = afs_CFileOpen(&tdc->f.inode);
2302 afs_RemoveVCB(&avc->f.fid);
2303 tdc->f.states |= DWriting;
2304 tdc->dflags |= DFFetching;
2305 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2306 if (tdc->mflags & DFFetchReq) {
2307 tdc->mflags &= ~DFFetchReq;
2308 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2309 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2310 __FILE__, ICL_TYPE_INT32, __LINE__,
2311 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2314 tsmall = osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2315 setVcacheStatus = 0;
2318 * Remember if we are doing the reading from a replicated volume,
2319 * and how many times we've zipped around the fetch/analyze loop.
2321 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2323 accP = &(afs_stats_cmfullperf.accessinf);
2325 (accP->replicatedRefs)++;
2327 (accP->unreplicatedRefs)++;
2328 #endif /* AFS_NOSTATS */
2329 /* this is a cache miss */
2330 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2331 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2332 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2335 afs_stats_cmperf.dcacheMisses++;
2338 * Dynamic root support: fetch data from local memory.
2340 if (afs_IsDynroot(avc)) {
2344 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2346 dynrootDir += Position;
2347 dynrootLen -= Position;
2348 if (size > dynrootLen)
2352 code = afs_CFileWrite(file, 0, dynrootDir, size);
2360 tdc->validPos = Position + size;
2361 afs_CFileTruncate(file, size); /* prune it */
2362 } else if (afs_IsDynrootMount(avc)) {
2366 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2368 dynrootDir += Position;
2369 dynrootLen -= Position;
2370 if (size > dynrootLen)
2374 code = afs_CFileWrite(file, 0, dynrootDir, size);
2382 tdc->validPos = Position + size;
2383 afs_CFileTruncate(file, size); /* prune it */
2386 * Not a dynamic vnode: do the real fetch.
2391 * avc->lock(R) if setLocks && !slowPass
2392 * avc->lock(W) if !setLocks || slowPass
2396 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
2401 (accP->numReplicasAccessed)++;
2403 #endif /* AFS_NOSTATS */
2404 if (!setLocks || slowPass) {
2405 avc->callback = tc->parent->srvr->server;
2407 newCallback = tc->parent->srvr->server;
2411 code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
2417 /* callback could have been broken (or expired) in a race here,
2418 * but we return the data anyway. It's as good as we knew about
2419 * when we started. */
2421 * validPos is updated by CacheFetchProc, and can only be
2422 * modifed under a dcache write lock, which we've blocked out
2424 size = tdc->validPos - Position; /* actual segment size */
2427 afs_CFileTruncate(file, size); /* prune it */
2429 if (!setLocks || slowPass) {
2430 afs_StaleVCacheFlags(avc, AFS_STALEVC_CLEARCB, CUnique);
2432 /* Something lost. Forget about performance, and go
2433 * back with a vcache write lock.
2435 afs_CFileTruncate(file, 0);
2436 afs_AdjustSize(tdc, 0);
2437 afs_CFileClose(file);
2438 osi_FreeLargeSpace(tsmall);
2440 ReleaseWriteLock(&tdc->lock);
2443 ReleaseReadLock(&avc->lock);
2446 /* If we have a connection, we must put it back,
2447 * since afs_Analyze will not be called here. */
2448 afs_PutConn(tc, rxconn, SHARED_LOCK);
2452 goto RetryGetDCache;
2456 } while (afs_Analyze
2457 (tc, rxconn, code, &avc->f.fid, areq,
2458 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2462 * avc->lock(R) if setLocks && !slowPass
2463 * avc->lock(W) if !setLocks || slowPass
2469 * In the case of replicated access, jot down info on the number of
2470 * attempts it took before we got through or gave up.
2473 if (numFetchLoops <= 1)
2474 (accP->refFirstReplicaOK)++;
2475 if (numFetchLoops > accP->maxReplicasPerRef)
2476 accP->maxReplicasPerRef = numFetchLoops;
2478 #endif /* AFS_NOSTATS */
2480 tdc->dflags &= ~DFFetching;
2481 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2482 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2483 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2484 tdc, ICL_TYPE_INT32, tdc->dflags);
2485 if (avc->execsOrWriters == 0)
2486 tdc->f.states &= ~DWriting;
2488 /* now, if code != 0, we have an error and should punt.
2489 * note that we have the vcache write lock, either because
2490 * !setLocks or slowPass.
2493 afs_CFileTruncate(file, 0);
2494 afs_AdjustSize(tdc, 0);
2495 afs_CFileClose(file);
2496 ZapDCE(tdc); /* sets DFEntryMod */
2497 if (vType(avc) == VDIR) {
2500 tdc->f.states &= ~(DRO|DBackup|DRW);
2501 afs_DCMoveBucket(tdc, 0, 0);
2502 ReleaseWriteLock(&tdc->lock);
2504 if (!afs_IsDynroot(avc)) {
2505 afs_StaleVCacheFlags(avc, 0, CUnique);
2508 * avc->lock(W); assert(!setLocks || slowPass)
2510 osi_Assert(!setLocks || slowPass);
2516 /* otherwise we copy in the just-fetched info */
2517 afs_CFileClose(file);
2518 afs_AdjustSize(tdc, size); /* new size */
2520 * Copy appropriate fields into vcache. Status is
2521 * copied later where we selectively acquire the
2522 * vcache write lock.
2525 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2527 setVcacheStatus = 1;
2528 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2529 tsmall->OutStatus.DataVersion);
2530 tdc->dflags |= DFEntryMod;
2531 afs_indexFlags[tdc->index] |= IFEverUsed;
2532 ConvertWToSLock(&tdc->lock);
2533 } /*Data version numbers don't match */
2536 * Data version numbers match.
2538 afs_stats_cmperf.dcacheHits++;
2539 } /*Data version numbers match */
2541 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2545 * avc->lock(R) if setLocks && !slowPass
2546 * avc->lock(W) if !setLocks || slowPass
2547 * tdc->lock(S) if tdc
2551 * See if this was a reference to a file in the local cell.
2553 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2554 afs_stats_cmperf.dlocalAccesses++;
2556 afs_stats_cmperf.dremoteAccesses++;
2558 /* Fix up LRU info */
2561 ObtainWriteLock(&afs_xdcache, 602);
2562 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2563 hadd32(afs_indexCounter, 1);
2564 ReleaseWriteLock(&afs_xdcache);
2566 /* return the data */
2567 if (vType(avc) == VDIR)
2570 *aoffset = AFS_CHUNKOFFSET(abyte);
2571 *alen = (tdc->f.chunkBytes - *aoffset);
2572 ReleaseSharedLock(&tdc->lock);
2577 * avc->lock(R) if setLocks && !slowPass
2578 * avc->lock(W) if !setLocks || slowPass
2581 /* Fix up the callback and status values in the vcache */
2583 if (setLocks && !slowPass) {
2586 * This is our dirty little secret to parallel fetches.
2587 * We don't write-lock the vcache while doing the fetch,
2588 * but potentially we'll need to update the vcache after
2589 * the fetch is done.
2591 * Drop the read lock and try to re-obtain the write
2592 * lock. If the vcache still has the same DV, it's
2593 * ok to go ahead and install the new data.
2595 afs_hyper_t currentDV, statusDV;
2597 hset(currentDV, avc->f.m.DataVersion);
2599 if (setNewCallback && avc->callback != newCallback)
2603 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2604 tsmall->OutStatus.DataVersion);
2606 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2608 if (setVcacheStatus && !hsame(currentDV, statusDV))
2612 ReleaseReadLock(&avc->lock);
2614 if (doVcacheUpdate) {
2615 ObtainWriteLock(&avc->lock, 615);
2616 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2617 /* We lose. Someone will beat us to it. */
2619 ReleaseWriteLock(&avc->lock);
2624 /* With slow pass, we've already done all the updates */
2626 ReleaseWriteLock(&avc->lock);
2629 /* Check if we need to perform any last-minute fixes with a write-lock */
2630 if (!setLocks || doVcacheUpdate) {
2632 avc->callback = newCallback;
2633 if (tsmall && setVcacheStatus)
2634 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2636 ReleaseWriteLock(&avc->lock);
2640 osi_FreeLargeSpace(tsmall);
2643 } /*afs_GetDCache */
2647 * afs_WriteThroughDSlots
2650 * Sweep through the dcache slots and write out any modified
2651 * in-memory data back on to our caching store.
2657 * The afs_xdcache is write-locked through this whole affair.
2660 afs_WriteThroughDSlots(void)
2663 afs_int32 i, touchedit = 0;
2665 struct afs_q DirtyQ, *tq;
2667 AFS_STATCNT(afs_WriteThroughDSlots);
2670 * Because of lock ordering, we can't grab dcache locks while
2671 * holding afs_xdcache. So we enter xdcache, get a reference
2672 * for every dcache entry, and exit xdcache.
2674 ObtainWriteLock(&afs_xdcache, 283);
2676 for (i = 0; i < afs_cacheFiles; i++) {
2677 tdc = afs_indexTable[i];
2679 /* Grab tlock in case the existing refcount isn't zero */
2680 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2681 ObtainWriteLock(&tdc->tlock, 623);
2683 ReleaseWriteLock(&tdc->tlock);
2685 QAdd(&DirtyQ, &tdc->dirty);
2688 ReleaseWriteLock(&afs_xdcache);
2691 * Now, for each dcache entry we found, check if it's dirty.
2692 * If so, get write-lock, get afs_xdcache, which protects
2693 * afs_cacheInodep, and flush it. Don't forget to put back
2697 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2699 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2701 if (tdc->dflags & DFEntryMod) {
2704 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2706 /* Now that we have the write lock, double-check */
2707 if (wrLock && (tdc->dflags & DFEntryMod)) {
2708 tdc->dflags &= ~DFEntryMod;
2709 ObtainWriteLock(&afs_xdcache, 620);
2710 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
2711 ReleaseWriteLock(&afs_xdcache);
2715 ReleaseWriteLock(&tdc->lock);
2721 ObtainWriteLock(&afs_xdcache, 617);
2722 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2723 /* Touch the file to make sure that the mtime on the file is kept
2724 * up-to-date to avoid losing cached files on cold starts because
2725 * their mtime seems old...
2727 struct afs_fheader theader;
2729 afs_InitFHeader(&theader);
2730 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2732 ReleaseWriteLock(&afs_xdcache);
2739 * Return a pointer to an freshly initialized dcache entry using
2740 * a memory-based cache. The tlock will be read-locked.
2743 * aslot : Dcache slot to look at.
2744 * type : What 'type' of dslot to get; see the dslot_state enum
2747 * Must be called with afs_xdcache write-locked.
2751 afs_MemGetDSlot(afs_int32 aslot, dslot_state type)
2756 AFS_STATCNT(afs_MemGetDSlot);
2757 if (CheckLock(&afs_xdcache) != -1)
2758 osi_Panic("getdslot nolock");
2759 if (aslot < 0 || aslot >= afs_cacheFiles)
2760 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2761 tdc = afs_indexTable[aslot];
2763 QRemove(&tdc->lruq); /* move to queue head */
2764 QAdd(&afs_DLRU, &tdc->lruq);
2765 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2766 ObtainWriteLock(&tdc->tlock, 624);
2768 ConvertWToRLock(&tdc->tlock);
2772 /* if we got here, the given slot is not in memory in our list of known
2773 * slots. for memcache, the only place a dslot can exist is in memory, so
2774 * if the caller is expecting to get back a known dslot, and we've reached
2775 * here, something is very wrong. DSLOT_NEW is the only type of dslot that
2776 * may not exist; for all others, the caller assumes the given dslot
2777 * already exists. so, 'type' had better be DSLOT_NEW here, or something is
2779 osi_Assert(type == DSLOT_NEW);
2781 if (!afs_freeDSList)
2782 afs_GetDownDSlot(4);
2783 if (!afs_freeDSList) {
2784 /* none free, making one is better than a panic */
2785 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2786 tdc = afs_osi_Alloc(sizeof(struct dcache));
2787 osi_Assert(tdc != NULL);
2788 #ifdef KERNEL_HAVE_PIN
2789 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2792 tdc = afs_freeDSList;
2793 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2796 tdc->dflags = 0; /* up-to-date, not in free q */
2798 QAdd(&afs_DLRU, &tdc->lruq);
2799 if (tdc->lruq.prev == &tdc->lruq)
2800 osi_Panic("lruq 3");
2802 /* initialize entry */
2803 tdc->f.fid.Cell = 0;
2804 tdc->f.fid.Fid.Volume = 0;
2806 hones(tdc->f.versionNo);
2807 tdc->f.inode.mem = aslot;
2808 tdc->dflags |= DFEntryMod;
2811 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2814 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2815 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2816 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2819 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2820 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2821 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2822 ObtainReadLock(&tdc->tlock);
2824 afs_indexTable[aslot] = tdc;
2827 } /*afs_MemGetDSlot */
2829 unsigned int last_error = 0, lasterrtime = 0;
2835 * Return a pointer to an freshly initialized dcache entry using
2836 * a UFS-based disk cache. The dcache tlock will be read-locked.
2839 * aslot : Dcache slot to look at.
2840 * type : What 'type' of dslot to get; see the dslot_state enum
2843 * afs_xdcache lock write-locked.
2846 afs_UFSGetDSlot(afs_int32 aslot, dslot_state type)
2854 AFS_STATCNT(afs_UFSGetDSlot);
2855 if (CheckLock(&afs_xdcache) != -1)
2856 osi_Panic("getdslot nolock");
2857 if (aslot < 0 || aslot >= afs_cacheFiles)
2858 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2859 tdc = afs_indexTable[aslot];
2861 QRemove(&tdc->lruq); /* move to queue head */
2862 QAdd(&afs_DLRU, &tdc->lruq);
2863 /* Grab tlock in case refCount != 0 */
2864 ObtainWriteLock(&tdc->tlock, 625);
2866 ConvertWToRLock(&tdc->tlock);
2870 /* otherwise we should read it in from the cache file */
2871 if (!afs_freeDSList)
2872 afs_GetDownDSlot(4);
2873 if (!afs_freeDSList) {
2874 /* none free, making one is better than a panic */
2875 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2876 tdc = afs_osi_Alloc(sizeof(struct dcache));
2877 osi_Assert(tdc != NULL);
2878 #ifdef KERNEL_HAVE_PIN
2879 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2882 tdc = afs_freeDSList;
2883 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2886 tdc->dflags = 0; /* up-to-date, not in free q */
2888 QAdd(&afs_DLRU, &tdc->lruq);
2889 if (tdc->lruq.prev == &tdc->lruq)
2890 osi_Panic("lruq 3");
2893 * Seek to the aslot'th entry and read it in.
2895 off = sizeof(struct fcache)*aslot + sizeof(struct afs_fheader);
2897 afs_osi_Read(afs_cacheInodep,
2898 off, (char *)(&tdc->f),
2899 sizeof(struct fcache));
2901 if (code != sizeof(struct fcache)) {
2903 #if defined(KERNEL_HAVE_UERROR)
2904 last_error = getuerror();
2908 lasterrtime = osi_Time();
2909 if (type != DSLOT_NEW) {
2910 /* If we are requesting a non-DSLOT_NEW slot, this is an error.
2911 * non-DSLOT_NEW slots are supposed to already exist, so if we
2912 * failed to read in the slot, something is wrong. */
2913 struct osi_stat tstat;
2914 if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
2917 afs_warn("afs: disk cache read error in CacheItems slot %d "
2918 "off %d/%d code %d/%d\n",
2920 off, (int)tstat.size,
2921 (int)code, (int)sizeof(struct fcache));
2922 /* put tdc back on the free dslot list */
2923 QRemove(&tdc->lruq);
2924 tdc->index = NULLIDX;
2925 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
2926 afs_freeDSList = tdc;
2930 if (!afs_CellNumValid(tdc->f.fid.Cell)) {
2932 if (type == DSLOT_VALID) {
2933 osi_Panic("afs: needed valid dcache but index %d off %d has "
2934 "invalid cell num %d\n",
2935 (int)aslot, off, (int)tdc->f.fid.Cell);
2939 if (type == DSLOT_VALID && tdc->f.fid.Fid.Volume == 0) {
2940 osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
2944 if (type == DSLOT_UNUSED) {
2945 /* the requested dslot is known to exist, but contain invalid data
2946 * (this happens when we're using a dslot from the free or discard
2947 * list). be sure not to re-use the data in it, so force invalidation.
2953 tdc->f.fid.Cell = 0;
2954 tdc->f.fid.Fid.Volume = 0;
2956 hones(tdc->f.versionNo);
2957 tdc->dflags |= DFEntryMod;
2958 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2959 tdc->f.states &= ~(DRO|DBackup|DRW);
2960 afs_DCMoveBucket(tdc, 0, 0);
2962 if (tdc->f.states & DRO) {
2963 afs_DCMoveBucket(tdc, 0, 2);
2964 } else if (tdc->f.states & DBackup) {
2965 afs_DCMoveBucket(tdc, 0, 1);
2967 afs_DCMoveBucket(tdc, 0, 1);
2972 if (tdc->f.chunk >= 0)
2973 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2978 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2979 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2980 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2983 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2984 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2985 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2986 ObtainReadLock(&tdc->tlock);
2989 * If we didn't read into a temporary dcache region, update the
2990 * slot pointer table.
2992 afs_indexTable[aslot] = tdc;
2995 } /*afs_UFSGetDSlot */
3000 * Write a particular dcache entry back to its home in the
3003 * \param adc Pointer to the dcache entry to write.
3004 * \param atime If true, set the modtime on the file to the current time.
3006 * \note Environment:
3007 * Must be called with the afs_xdcache lock at least read-locked,
3008 * and dcache entry at least read-locked.
3009 * The reference count is not changed.
3013 afs_WriteDCache(struct dcache *adc, int atime)
3017 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
3019 AFS_STATCNT(afs_WriteDCache);
3020 osi_Assert(WriteLocked(&afs_xdcache));
3022 adc->f.modTime = osi_Time();
3024 if ((afs_indexFlags[adc->index] & (IFFree | IFDiscarded)) == 0 &&
3025 adc->f.fid.Fid.Volume == 0) {
3026 /* If a dcache slot is not on the free or discard list, it must be
3027 * in the hash table. Thus, the volume must be non-zero, since that
3028 * is how we determine whether or not to unhash the entry when kicking
3029 * it out of the cache. Do this check now, since otherwise this can
3030 * cause hash table corruption and a panic later on after we read the
3032 osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
3033 adc->index, (unsigned)afs_indexFlags[adc->index]);
3037 * Seek to the right dcache slot and write the in-memory image out to disk.
3039 afs_cellname_write();
3041 afs_osi_Write(afs_cacheInodep,
3042 sizeof(struct fcache) * adc->index +
3043 sizeof(struct afs_fheader), (char *)(&adc->f),
3044 sizeof(struct fcache));
3045 if (code != sizeof(struct fcache)) {
3046 afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
3047 (long)(sizeof(struct fcache) * adc->index + sizeof(struct afs_fheader)),
3048 (int)code, (int)sizeof(struct fcache));
3057 * Wake up users of a particular file waiting for stores to take
3060 * \param avc Ptr to related vcache entry.
3062 * \note Environment:
3063 * Nothing interesting.
3066 afs_wakeup(struct vcache *avc)
3069 struct brequest *tb;
3071 AFS_STATCNT(afs_wakeup);
3072 for (i = 0; i < NBRS; i++, tb++) {
3073 /* if request is valid and for this file, we've found it */
3074 if (tb->refCount > 0 && avc == tb->vc) {
3077 * If CSafeStore is on, then we don't awaken the guy
3078 * waiting for the store until the whole store has finished.
3079 * Otherwise, we do it now. Note that if CSafeStore is on,
3080 * the BStore routine actually wakes up the user, instead
3082 * I think this is redundant now because this sort of thing
3083 * is already being handled by the higher-level code.
3085 if ((avc->f.states & CSafeStore) == 0) {
3086 tb->code_raw = tb->code_checkcode = 0;
3087 tb->flags |= BUVALID;
3088 if (tb->flags & BUWAIT) {
3089 tb->flags &= ~BUWAIT;
3100 * Given a file name and inode, set up that file to be an
3101 * active member in the AFS cache. This also involves checking
3102 * the usability of its data.
3104 * \param afile Name of the cache file to initialize.
3105 * \param ainode Inode of the file.
3107 * \note Environment:
3108 * This function is called only during initialization.
3111 afs_InitCacheFile(char *afile, ino_t ainode)
3116 struct osi_file *tfile;
3117 struct osi_stat tstat;
3120 AFS_STATCNT(afs_InitCacheFile);
3121 index = afs_stats_cmperf.cacheNumEntries;
3122 if (index >= afs_cacheFiles)
3125 ObtainWriteLock(&afs_xdcache, 282);
3126 tdc = afs_GetNewDSlot(index);
3127 ReleaseReadLock(&tdc->tlock);
3128 ReleaseWriteLock(&afs_xdcache);
3130 ObtainWriteLock(&tdc->lock, 621);
3131 ObtainWriteLock(&afs_xdcache, 622);
3132 if (!afile && !ainode) {
3137 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
3139 ReleaseWriteLock(&afs_xdcache);
3140 ReleaseWriteLock(&tdc->lock);
3145 /* Add any other 'complex' inode types here ... */
3146 #if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
3147 tdc->f.inode.ufs = ainode;
3149 osi_Panic("Can't init cache with inode numbers when complex inodes are "
3154 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3156 tfile = osi_UFSOpen(&tdc->f.inode);
3158 ReleaseWriteLock(&afs_xdcache);
3159 ReleaseWriteLock(&tdc->lock);
3164 code = afs_osi_Stat(tfile, &tstat);
3166 osi_Panic("initcachefile stat");
3169 * If file size doesn't match the cache info file, it's probably bad.
3171 if (tdc->f.chunkBytes != tstat.size)
3174 * If file changed within T (120?) seconds of cache info file, it's
3175 * probably bad. In addition, if slot changed within last T seconds,
3176 * the cache info file may be incorrectly identified, and so slot
3179 if (cacheInfoModTime < tstat.mtime + 120)
3181 if (cacheInfoModTime < tdc->f.modTime + 120)
3183 /* In case write through is behind, make sure cache items entry is
3184 * at least as new as the chunk.
3186 if (tdc->f.modTime < tstat.mtime)
3189 tdc->f.chunkBytes = 0;
3192 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3193 if (tfile && tstat.size != 0)
3194 osi_UFSTruncate(tfile, 0);
3195 tdc->f.states &= ~(DRO|DBackup|DRW);
3196 afs_DCMoveBucket(tdc, 0, 0);
3197 /* put entry in free cache slot list */
3198 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3199 afs_freeDCList = index;
3201 afs_indexFlags[index] |= IFFree;
3202 afs_indexUnique[index] = 0;
3205 * We must put this entry in the appropriate hash tables.
3206 * Note that i is still set from the above DCHash call
3208 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3209 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3210 afs_dchashTbl[code] = tdc->index;
3211 code = DVHash(&tdc->f.fid);
3212 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3213 afs_dvhashTbl[code] = tdc->index;
3214 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3216 /* has nontrivial amt of data */
3217 afs_indexFlags[index] |= IFEverUsed;
3218 afs_stats_cmperf.cacheFilesReused++;
3220 * Initialize index times to file's mod times; init indexCounter
3223 hset32(afs_indexTimes[index], tstat.atime);
3224 if (hgetlo(afs_indexCounter) < tstat.atime) {
3225 hset32(afs_indexCounter, tstat.atime);
3227 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3228 } /*File is not bad */
3231 osi_UFSClose(tfile);
3232 tdc->f.states &= ~DWriting;
3233 tdc->dflags &= ~DFEntryMod;
3234 /* don't set f.modTime; we're just cleaning up */
3235 osi_Assert(afs_WriteDCache(tdc, 0) == 0);
3236 ReleaseWriteLock(&afs_xdcache);
3237 ReleaseWriteLock(&tdc->lock);
3239 afs_stats_cmperf.cacheNumEntries++;
3244 /*Max # of struct dcache's resident at any time*/
3246 * If 'dchint' is enabled then in-memory dcache min is increased because of
3252 * Initialize dcache related variables.
3262 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3269 afs_freeDCList = NULLIDX;
3270 afs_discardDCList = NULLIDX;
3271 afs_freeDCCount = 0;
3272 afs_freeDSList = NULL;
3273 hzero(afs_indexCounter);
3275 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3281 if (achunk < 0 || achunk > 30)
3282 achunk = 13; /* Use default */
3283 AFS_SETCHUNKSIZE(achunk);
3289 /* afs_dhashsize defaults to 1024 */
3290 if (aDentries > 512)
3291 afs_dhashsize = 2048;
3292 /* Try to keep the average chain length around two unless the table
3293 * would be ridiculously big. */
3294 if (aDentries > 4096) {
3295 afs_dhashbits = opr_fls(aDentries) - 3;
3296 /* Cap the hash tables to 32k entries. */
3297 if (afs_dhashbits > 15)
3299 afs_dhashsize = opr_jhash_size(afs_dhashbits);
3301 /* initialize hash tables */
3302 afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3303 osi_Assert(afs_dvhashTbl != NULL);
3304 afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3305 osi_Assert(afs_dchashTbl != NULL);
3306 for (i = 0; i < afs_dhashsize; i++) {
3307 afs_dvhashTbl[i] = NULLIDX;
3308 afs_dchashTbl[i] = NULLIDX;
3310 afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3311 osi_Assert(afs_dvnextTbl != NULL);
3312 afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3313 osi_Assert(afs_dcnextTbl != NULL);
3314 for (i = 0; i < afiles; i++) {
3315 afs_dvnextTbl[i] = NULLIDX;
3316 afs_dcnextTbl[i] = NULLIDX;
3319 /* Allocate and zero the pointer array to the dcache entries */
3320 afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3321 osi_Assert(afs_indexTable != NULL);
3322 memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3323 afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3324 osi_Assert(afs_indexTimes != NULL);
3325 memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3326 afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
3327 osi_Assert(afs_indexUnique != NULL);
3328 memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3329 afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
3330 osi_Assert(afs_indexFlags != NULL);
3331 memset(afs_indexFlags, 0, afiles * sizeof(char));
3333 /* Allocate and thread the struct dcache entries themselves */
3334 tdp = afs_Initial_freeDSList =
3335 afs_osi_Alloc(aDentries * sizeof(struct dcache));
3336 osi_Assert(tdp != NULL);
3337 memset(tdp, 0, aDentries * sizeof(struct dcache));
3338 #ifdef KERNEL_HAVE_PIN
3339 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3340 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3341 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3342 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3343 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3344 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3345 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3346 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3347 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3350 afs_freeDSList = &tdp[0];
3351 for (i = 0; i < aDentries - 1; i++) {
3352 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3353 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3354 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3355 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3357 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3358 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3359 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3360 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3362 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3363 afs_cacheBlocks = ablocks;
3364 afs_ComputeCacheParms(); /* compute parms based on cache size */
3366 afs_dcentries = aDentries;
3368 afs_stats_cmperf.cacheBucket0_Discarded =
3369 afs_stats_cmperf.cacheBucket1_Discarded =
3370 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3374 if (aflags & AFSCALL_INIT_MEMCACHE) {
3376 * Use a memory cache instead of a disk cache
3378 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3379 afs_cacheType = &afs_MemCacheOps;
3380 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3381 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3382 /* ablocks is reported in 1K blocks */
3383 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3385 afs_warn("afsd: memory cache too large for available memory.\n");
3386 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3389 afs_warn("Memory cache: Allocating %d dcache entries...",
3392 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3393 afs_cacheType = &afs_UfsCacheOps;
3398 * Shuts down the cache.
3402 shutdown_dcache(void)
3406 #ifdef AFS_CACHE_VNODE_PATH
3407 if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
3409 for (i = 0; i < afs_cacheFiles; i++) {
3410 tdc = afs_indexTable[i];
3412 afs_osi_FreeStr(tdc->f.inode.ufs);
3418 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3419 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3420 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3421 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3422 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3423 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3424 afs_osi_Free(afs_Initial_freeDSList,
3425 afs_dcentries * sizeof(struct dcache));
3426 #ifdef KERNEL_HAVE_PIN
3427 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3428 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3429 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3430 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3431 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3432 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3433 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3437 for (i = 0; i < afs_dhashsize; i++) {
3438 afs_dvhashTbl[i] = NULLIDX;
3439 afs_dchashTbl[i] = NULLIDX;
3442 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3443 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3445 afs_blocksUsed = afs_dcentries = 0;
3446 afs_stats_cmperf.cacheBucket0_Discarded =
3447 afs_stats_cmperf.cacheBucket1_Discarded =
3448 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3449 hzero(afs_indexCounter);
3451 afs_freeDCCount = 0;
3452 afs_freeDCList = NULLIDX;
3453 afs_discardDCList = NULLIDX;
3454 afs_freeDSList = afs_Initial_freeDSList = 0;
3456 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3462 * Get a dcache ready for writing, respecting the current cache size limits
3464 * len is required because afs_GetDCache with flag == 4 expects the length
3465 * field to be filled. It decides from this whether it's necessary to fetch
3466 * data into the chunk before writing or not (when the whole chunk is
3469 * \param avc The vcache to fetch a dcache for
3470 * \param filePos The start of the section to be written
3471 * \param len The length of the section to be written
3475 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3476 * must be released and afs_PutDCache() called to free dcache.
3479 * \note avc->lock must be held on entry. Function may release and reobtain
3480 * avc->lock and GLOCK.
3484 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3485 afs_size_t len, struct vrequest *areq,
3488 struct dcache *tdc = NULL;
3491 /* read the cached info */
3493 tdc = afs_FindDCache(avc, filePos);
3495 ObtainWriteLock(&tdc->lock, 657);
3496 } else if (afs_blocksUsed >
3497 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3498 tdc = afs_FindDCache(avc, filePos);
3500 ObtainWriteLock(&tdc->lock, 658);
3501 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3502 || (tdc->dflags & DFFetching)) {
3503 ReleaseWriteLock(&tdc->lock);
3509 afs_MaybeWakeupTruncateDaemon();
3510 while (afs_blocksUsed >
3511 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3512 ReleaseWriteLock(&avc->lock);
3513 if (afs_blocksUsed - afs_blocksDiscarded >
3514 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3515 afs_WaitForCacheDrain = 1;
3516 afs_osi_Sleep(&afs_WaitForCacheDrain);
3518 afs_MaybeFreeDiscardedDCache();
3519 afs_MaybeWakeupTruncateDaemon();
3520 ObtainWriteLock(&avc->lock, 509);
3522 avc->f.states |= CDirty;
3523 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3525 ObtainWriteLock(&tdc->lock, 659);
3528 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3530 ObtainWriteLock(&tdc->lock, 660);
3533 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3534 afs_stats_cmperf.cacheCurrDirtyChunks++;
3535 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3537 if (!(tdc->f.states & DWriting)) {
3538 /* don't mark entry as mod if we don't have to */
3539 tdc->f.states |= DWriting;
3540 tdc->dflags |= DFEntryMod;
3547 * Make a shadow copy of a dir's dcache. It's used for disconnected
3548 * operations like remove/create/rename to keep the original directory data.
3549 * On reconnection, we can diff the original data with the server and get the
3550 * server changes and with the local data to get the local changes.
3552 * \param avc The dir vnode.
3553 * \param adc The dir dcache.
3555 * \return 0 for success.
3557 * \note The vcache entry must be write locked.
3558 * \note The dcache entry must be read locked.
3561 afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3563 int i, code, ret_code = 0, written, trans_size;
3564 struct dcache *new_dc = NULL;
3565 struct osi_file *tfile_src, *tfile_dst;
3566 struct VenusFid shadow_fid;
3569 /* Is this a dir? */
3570 if (vType(avc) != VDIR)
3573 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3576 /* Generate a fid for the shadow dir. */
3577 shadow_fid.Cell = avc->f.fid.Cell;
3578 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3579 afs_GenShadowFid(&shadow_fid);
3581 ObtainWriteLock(&afs_xdcache, 716);
3583 /* Get a fresh dcache. */
3584 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3587 ObtainReadLock(&adc->mflock);
3589 /* Set up the new fid. */
3590 /* Copy interesting data from original dir dcache. */
3591 new_dc->mflags = adc->mflags;
3592 new_dc->dflags = adc->dflags;
3593 new_dc->f.modTime = adc->f.modTime;
3594 new_dc->f.versionNo = adc->f.versionNo;
3595 new_dc->f.states = adc->f.states;
3596 new_dc->f.chunk= adc->f.chunk;
3597 new_dc->f.chunkBytes = adc->f.chunkBytes;
3599 ReleaseReadLock(&adc->mflock);
3601 /* Now add to the two hash chains */
3602 i = DCHash(&shadow_fid, 0);
3603 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3604 afs_dchashTbl[i] = new_dc->index;
3606 i = DVHash(&shadow_fid);
3607 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3608 afs_dvhashTbl[i] = new_dc->index;
3610 ReleaseWriteLock(&afs_xdcache);
3612 /* Alloc a 4k block. */
3613 data = afs_osi_Alloc(4096);
3615 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3620 /* Open the files. */
3621 tfile_src = afs_CFileOpen(&adc->f.inode);
3622 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3623 osi_Assert(tfile_src);
3624 osi_Assert(tfile_dst);
3626 /* And now copy dir dcache data into this dcache,
3630 while (written < adc->f.chunkBytes) {
3631 trans_size = adc->f.chunkBytes - written;
3632 if (trans_size > 4096)
3635 /* Read a chunk from the dcache. */
3636 code = afs_CFileRead(tfile_src, written, data, trans_size);
3637 if (code < trans_size) {
3642 /* Write it to the new dcache. */
3643 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3644 if (code < trans_size) {
3649 written+=trans_size;
3652 afs_CFileClose(tfile_dst);
3653 afs_CFileClose(tfile_src);
3655 afs_osi_Free(data, 4096);
3657 ReleaseWriteLock(&new_dc->lock);
3658 afs_PutDCache(new_dc);
3661 ObtainWriteLock(&afs_xvcache, 763);
3662 ObtainWriteLock(&afs_disconDirtyLock, 765);
3663 QAdd(&afs_disconShadow, &avc->shadowq);
3664 osi_Assert((afs_RefVCache(avc) == 0));
3665 ReleaseWriteLock(&afs_disconDirtyLock);
3666 ReleaseWriteLock(&afs_xvcache);
3668 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3669 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3677 * Delete the dcaches of a shadow dir.
3679 * \param avc The vcache containing the shadow fid.
3681 * \note avc must be write locked.
3684 afs_DeleteShadowDir(struct vcache *avc)
3687 struct VenusFid shadow_fid;
3689 shadow_fid.Cell = avc->f.fid.Cell;
3690 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3691 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3692 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3694 tdc = afs_FindDCacheByFid(&shadow_fid);
3696 afs_HashOutDCache(tdc, 1);
3697 afs_DiscardDCache(tdc);
3700 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3701 ObtainWriteLock(&afs_disconDirtyLock, 708);
3702 QRemove(&avc->shadowq);
3703 ReleaseWriteLock(&afs_disconDirtyLock);
3704 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3708 * Populate a dcache with empty chunks up to a given file size,
3709 * used before extending a file in order to avoid 'holes' which
3710 * we can't access in disconnected mode.
3712 * \param avc The vcache which is being extended (locked)
3713 * \param alen The new length of the file
3717 afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
3720 afs_size_t len, offset;
3721 afs_int32 start, end;
3723 /* We're doing this to deal with the situation where we extend
3724 * by writing after lseek()ing past the end of the file . If that
3725 * extension skips chunks, then those chunks won't be created, and
3726 * GetDCache will assume that they have to be fetched from the server.
3727 * So, for each chunk between the current file position, and the new
3728 * length we GetDCache for that chunk.
3731 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3734 if (avc->f.m.Length == 0)
3737 start = AFS_CHUNK(avc->f.m.Length)+1;
3739 end = AFS_CHUNK(apos);
3742 len = AFS_CHUNKTOSIZE(start);
3743 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);