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 if (*indexp != NULLIDX) {
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 dcaches 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);
1508 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1511 ReleaseReadLock(&tdc->tlock);
1515 ReleaseWriteLock(&afs_xdcache);
1517 /*printf("Missing %d chunks\n", totalChunks);*/
1519 return (totalChunks);
1526 * Given the cached info for a file and a byte offset into the
1527 * file, make sure the dcache entry for that file and containing
1528 * the given byte is available, returning it to our caller.
1531 * avc : Pointer to the (held) vcache entry to look in.
1532 * abyte : Which byte we want to get to.
1535 * Pointer to the dcache entry covering the file & desired byte,
1536 * or NULL if not found.
1539 * The vcache entry is held upon entry.
1543 afs_FindDCache(struct vcache *avc, afs_size_t abyte)
1547 struct dcache *tdc = NULL;
1549 AFS_STATCNT(afs_FindDCache);
1550 chunk = AFS_CHUNK(abyte);
1553 * Hash on the [fid, chunk] and get the corresponding dcache index
1554 * after write-locking the dcache.
1556 i = DCHash(&avc->f.fid, chunk);
1557 ObtainWriteLock(&afs_xdcache, 278);
1558 for (index = afs_dchashTbl[i]; index != NULLIDX; index = afs_dcnextTbl[index]) {
1559 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1560 tdc = afs_GetValidDSlot(index);
1562 /* afs_FindDCache is best-effort; we may not find the given
1563 * file/offset, so if we cannot find the given dcache due to
1564 * i/o errors, that is okay. */
1568 ReleaseReadLock(&tdc->tlock);
1569 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1570 break; /* leaving refCount high for caller */
1575 if (index != NULLIDX) {
1576 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1577 hadd32(afs_indexCounter, 1);
1578 ReleaseWriteLock(&afs_xdcache);
1581 ReleaseWriteLock(&afs_xdcache);
1583 } /*afs_FindDCache */
1585 /* only call these from afs_AllocDCache() */
1586 static struct dcache *
1587 afs_AllocFreeDSlot(void)
1591 tdc = afs_GetDSlotFromList(&afs_freeDCList);
1595 afs_indexFlags[tdc->index] &= ~IFFree;
1596 ObtainWriteLock(&tdc->lock, 604);
1601 static struct dcache *
1602 afs_AllocDiscardDSlot(afs_int32 lock)
1605 afs_uint32 size = 0;
1606 struct osi_file *file;
1608 tdc = afs_GetDSlotFromList(&afs_discardDCList);
1612 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1613 ObtainWriteLock(&tdc->lock, 605);
1614 afs_discardDCCount--;
1616 ((tdc->f.chunkBytes +
1617 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1618 tdc->f.states &= ~(DRO|DBackup|DRW);
1619 afs_DCMoveBucket(tdc, size, 0);
1620 afs_blocksDiscarded -= size;
1621 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1623 /* Truncate the chunk so zeroes get filled properly */
1624 file = afs_CFileOpen(&tdc->f.inode);
1626 afs_CFileTruncate(file, 0);
1627 afs_CFileClose(file);
1628 afs_AdjustSize(tdc, 0);
1635 * Get a fresh dcache from the free or discarded list.
1637 * \param avc Who's dcache is this going to be?
1638 * \param chunk The position where it will be placed in.
1639 * \param lock How are locks held.
1640 * \param ashFid If this dcache going to be used for a shadow dir,
1643 * \note Required locks:
1645 * - avc (R if (lock & 1) set and W otherwise)
1646 * \note It write locks the new dcache. The caller must unlock it.
1648 * \return The new dcache.
1651 afs_AllocDCache(struct vcache *avc, afs_int32 chunk, afs_int32 lock,
1652 struct VenusFid *ashFid)
1654 struct dcache *tdc = NULL;
1656 /* if (lock & 2), prefer 'free' dcaches; otherwise, prefer 'discard'
1657 * dcaches. In either case, try both if our first choice doesn't work. */
1659 tdc = afs_AllocFreeDSlot();
1661 tdc = afs_AllocDiscardDSlot(lock);
1664 tdc = afs_AllocDiscardDSlot(lock);
1666 tdc = afs_AllocFreeDSlot();
1675 * avc->lock(R) if setLocks
1676 * avc->lock(W) if !setLocks
1682 * Fill in the newly-allocated dcache record.
1684 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1686 /* Use shadow fid if provided. */
1687 tdc->f.fid = *ashFid;
1689 /* Use normal vcache's fid otherwise. */
1690 tdc->f.fid = avc->f.fid;
1691 if (avc->f.states & CRO)
1692 tdc->f.states = DRO;
1693 else if (avc->f.states & CBackup)
1694 tdc->f.states = DBackup;
1696 tdc->f.states = DRW;
1697 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1698 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1700 hones(tdc->f.versionNo); /* invalid value */
1701 tdc->f.chunk = chunk;
1702 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1704 if (tdc->lruq.prev == &tdc->lruq)
1705 osi_Panic("lruq 1");
1714 * This function is called to obtain a reference to data stored in
1715 * the disk cache, locating a chunk of data containing the desired
1716 * byte and returning a reference to the disk cache entry, with its
1717 * reference count incremented.
1721 * avc : Ptr to a vcache entry (unlocked)
1722 * abyte : Byte position in the file desired
1723 * areq : Request structure identifying the requesting user.
1724 * aflags : Settings as follows:
1726 * 2 : Return after creating entry.
1727 * 4 : called from afs_vnop_write.c
1728 * *alen contains length of data to be written.
1730 * aoffset : Set to the offset within the chunk where the resident
1732 * alen : Set to the number of bytes of data after the desired
1733 * byte (including the byte itself) which can be read
1737 * The vcache entry pointed to by avc is unlocked upon entry.
1741 * Update the vnode-to-dcache hint if we can get the vnode lock
1742 * right away. Assumes dcache entry is at least read-locked.
1745 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1747 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1748 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1751 ReleaseWriteLock(&v->lock);
1755 /* avc - Write-locked unless aflags & 1 */
1757 afs_GetDCache(struct vcache *avc, afs_size_t abyte,
1758 struct vrequest *areq, afs_size_t * aoffset,
1759 afs_size_t * alen, int aflags)
1761 afs_int32 i, code, shortcut;
1762 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1763 afs_int32 adjustsize = 0;
1769 afs_size_t Position = 0;
1770 afs_int32 size, tlen; /* size of segment to transfer */
1771 struct afs_FetchOutput *tsmall = 0;
1773 struct osi_file *file;
1774 struct afs_conn *tc;
1776 struct server *newCallback = NULL;
1777 char setNewCallback;
1778 char setVcacheStatus;
1779 char doVcacheUpdate;
1781 int doAdjustSize = 0;
1782 int doReallyAdjustSize = 0;
1783 int overWriteWholeChunk = 0;
1784 struct rx_connection *rxconn;
1787 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1788 int fromReplica; /*Are we reading from a replica? */
1789 int numFetchLoops; /*# times around the fetch/analyze loop */
1790 #endif /* AFS_NOSTATS */
1792 AFS_STATCNT(afs_GetDCache);
1796 setLocks = aflags & 1;
1799 * Determine the chunk number and offset within the chunk corresponding
1800 * to the desired byte.
1802 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1805 chunk = AFS_CHUNK(abyte);
1808 /* come back to here if we waited for the cache to drain. */
1811 setNewCallback = setVcacheStatus = 0;
1815 ObtainWriteLock(&avc->lock, 616);
1817 ObtainReadLock(&avc->lock);
1822 * avc->lock(R) if setLocks && !slowPass
1823 * avc->lock(W) if !setLocks || slowPass
1828 /* check hints first! (might could use bcmp or some such...) */
1829 if ((tdc = avc->dchint)) {
1833 * The locking order between afs_xdcache and dcache lock matters.
1834 * The hint dcache entry could be anywhere, even on the free list.
1835 * Locking afs_xdcache ensures that noone is trying to pull dcache
1836 * entries from the free list, and thereby assuming them to be not
1837 * referenced and not locked.
1839 ObtainReadLock(&afs_xdcache);
1840 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1842 if (dcLocked && (tdc->index != NULLIDX)
1843 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1844 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1845 /* got the right one. It might not be the right version, and it
1846 * might be fetching, but it's the right dcache entry.
1848 /* All this code should be integrated better with what follows:
1849 * I can save a good bit more time under a write lock if I do..
1851 ObtainWriteLock(&tdc->tlock, 603);
1853 ReleaseWriteLock(&tdc->tlock);
1855 ReleaseReadLock(&afs_xdcache);
1858 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1859 && !(tdc->dflags & DFFetching)) {
1861 afs_stats_cmperf.dcacheHits++;
1862 ObtainWriteLock(&afs_xdcache, 559);
1863 QRemove(&tdc->lruq);
1864 QAdd(&afs_DLRU, &tdc->lruq);
1865 ReleaseWriteLock(&afs_xdcache);
1868 * avc->lock(R) if setLocks && !slowPass
1869 * avc->lock(W) if !setLocks || slowPass
1876 ReleaseSharedLock(&tdc->lock);
1877 ReleaseReadLock(&afs_xdcache);
1885 * avc->lock(R) if setLocks && !slowPass
1886 * avc->lock(W) if !setLocks || slowPass
1887 * tdc->lock(S) if tdc
1890 if (!tdc) { /* If the hint wasn't the right dcache entry */
1891 int dslot_error = 0;
1893 * Hash on the [fid, chunk] and get the corresponding dcache index
1894 * after write-locking the dcache.
1899 * avc->lock(R) if setLocks && !slowPass
1900 * avc->lock(W) if !setLocks || slowPass
1903 i = DCHash(&avc->f.fid, chunk);
1904 /* check to make sure our space is fine */
1905 afs_MaybeWakeupTruncateDaemon();
1907 ObtainWriteLock(&afs_xdcache, 280);
1909 for (index = afs_dchashTbl[i]; index != NULLIDX; us = index, index = afs_dcnextTbl[index]) {
1910 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1911 tdc = afs_GetValidDSlot(index);
1913 /* we got an i/o error when trying to get the given dslot.
1914 * it's possible the dslot we're looking for is elsewhere,
1915 * but most likely the disk cache is currently unusable, so
1916 * all afs_GetValidDSlot calls will fail, so just bail out. */
1921 ReleaseReadLock(&tdc->tlock);
1924 * avc->lock(R) if setLocks && !slowPass
1925 * avc->lock(W) if !setLocks || slowPass
1928 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1929 /* Move it up in the beginning of the list */
1930 if (afs_dchashTbl[i] != index) {
1931 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1932 afs_dcnextTbl[index] = afs_dchashTbl[i];
1933 afs_dchashTbl[i] = index;
1935 ReleaseWriteLock(&afs_xdcache);
1936 ObtainSharedLock(&tdc->lock, 606);
1937 break; /* leaving refCount high for caller */
1945 * If we didn't find the entry, we'll create one.
1947 if (index == NULLIDX) {
1950 * avc->lock(R) if setLocks
1951 * avc->lock(W) if !setLocks
1954 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1955 avc, ICL_TYPE_INT32, chunk);
1958 /* We couldn't find the dcache we want, but we hit some i/o
1959 * errors when trying to find it, so we're not sure if the
1960 * dcache we want is in the cache or not. Error out, so we
1961 * don't try to possibly create 2 separate dcaches for the
1962 * same exact data. */
1963 ReleaseWriteLock(&afs_xdcache);
1967 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1969 avc->f.states |= CDCLock;
1970 /* just need slots */
1971 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1973 avc->f.states &= ~CDCLock;
1975 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1977 ReleaseWriteLock(&afs_xdcache);
1978 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1979 /* It looks like afs_AllocDCache failed because we don't
1980 * have any free dslots to use. Maybe if we wait a little
1981 * while, we'll be able to free up some slots, so try for 5
1982 * minutes, then bail out. */
1983 if (++downDCount > 300) {
1984 afs_warn("afs: Unable to get free cache space for file "
1985 "%u:%u.%u.%u for 5 minutes; failing with an i/o error\n",
1987 avc->f.fid.Fid.Volume,
1988 avc->f.fid.Fid.Vnode,
1989 avc->f.fid.Fid.Unique);
1992 afs_osi_Wait(1000, 0, 0);
1996 /* afs_AllocDCache failed, but not because we're out of free
1997 * dslots. Something must be screwy with the cache, so bail out
1998 * immediately without waiting. */
1999 afs_warn("afs: Error while alloc'ing cache slot for file "
2000 "%u:%u.%u.%u; failing with an i/o error\n",
2002 avc->f.fid.Fid.Volume,
2003 avc->f.fid.Fid.Vnode,
2004 avc->f.fid.Fid.Unique);
2010 * avc->lock(R) if setLocks
2011 * avc->lock(W) if !setLocks
2017 * Now add to the two hash chains - note that i is still set
2018 * from the above DCHash call.
2020 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
2021 afs_dchashTbl[i] = tdc->index;
2022 i = DVHash(&avc->f.fid);
2023 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
2024 afs_dvhashTbl[i] = tdc->index;
2025 tdc->dflags = DFEntryMod;
2027 afs_MaybeWakeupTruncateDaemon();
2028 ReleaseWriteLock(&afs_xdcache);
2029 ConvertWToSLock(&tdc->lock);
2034 /* vcache->dcache hint failed */
2037 * avc->lock(R) if setLocks && !slowPass
2038 * avc->lock(W) if !setLocks || slowPass
2041 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
2042 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2043 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
2044 hgetlo(avc->f.m.DataVersion));
2046 * Here we have the entry in tdc, with its refCount incremented.
2047 * Note: we don't use the S-lock on avc; it costs concurrency when
2048 * storing a file back to the server.
2052 * Not a newly created file so we need to check the file's length and
2053 * compare data versions since someone could have changed the data or we're
2054 * reading a file written elsewhere. We only want to bypass doing no-op
2055 * read rpcs on newly created files (dv of 0) since only then we guarantee
2056 * that this chunk's data hasn't been filled by another client.
2058 size = AFS_CHUNKSIZE(abyte);
2059 if (aflags & 4) /* called from write */
2061 else /* called from read */
2062 tlen = tdc->validPos - abyte;
2063 Position = AFS_CHUNKTOBASE(chunk);
2064 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
2065 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
2066 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
2067 ICL_HANDLE_OFFSET(Position));
2068 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
2070 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
2071 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
2072 overWriteWholeChunk = 1;
2073 if (doAdjustSize || overWriteWholeChunk) {
2074 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
2076 #ifdef AFS_SGI64_ENV
2079 #else /* AFS_SGI64_ENV */
2082 #endif /* AFS_SGI64_ENV */
2083 #else /* AFS_SGI_ENV */
2086 #endif /* AFS_SGI_ENV */
2087 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
2088 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2089 #if defined(AFS_SUN5_ENV)
2090 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
2092 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
2094 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2095 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
2096 doReallyAdjustSize = 1;
2098 if (doReallyAdjustSize || overWriteWholeChunk) {
2099 /* no data in file to read at this position */
2100 UpgradeSToWLock(&tdc->lock, 607);
2101 file = afs_CFileOpen(&tdc->f.inode);
2103 afs_CFileTruncate(file, 0);
2104 afs_CFileClose(file);
2105 afs_AdjustSize(tdc, 0);
2106 hset(tdc->f.versionNo, avc->f.m.DataVersion);
2107 tdc->dflags |= DFEntryMod;
2109 ConvertWToSLock(&tdc->lock);
2114 * We must read in the whole chunk if the version number doesn't
2118 /* don't need data, just a unique dcache entry */
2119 ObtainWriteLock(&afs_xdcache, 608);
2120 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2121 hadd32(afs_indexCounter, 1);
2122 ReleaseWriteLock(&afs_xdcache);
2124 updateV2DC(setLocks, avc, tdc, 553);
2125 if (vType(avc) == VDIR)
2128 *aoffset = AFS_CHUNKOFFSET(abyte);
2129 if (tdc->validPos < abyte)
2130 *alen = (afs_size_t) 0;
2132 *alen = tdc->validPos - abyte;
2133 ReleaseSharedLock(&tdc->lock);
2136 ReleaseWriteLock(&avc->lock);
2138 ReleaseReadLock(&avc->lock);
2140 return tdc; /* check if we're done */
2145 * avc->lock(R) if setLocks && !slowPass
2146 * avc->lock(W) if !setLocks || slowPass
2149 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2151 setNewCallback = setVcacheStatus = 0;
2155 * avc->lock(R) if setLocks && !slowPass
2156 * avc->lock(W) if !setLocks || slowPass
2159 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2161 * Version number mismatch.
2164 * If we are disconnected, then we can't do much of anything
2165 * because the data doesn't match the file.
2167 if (AFS_IS_DISCONNECTED) {
2168 ReleaseSharedLock(&tdc->lock);
2171 ReleaseWriteLock(&avc->lock);
2173 ReleaseReadLock(&avc->lock);
2175 /* Flush the Dcache */
2180 UpgradeSToWLock(&tdc->lock, 609);
2183 * If data ever existed for this vnode, and this is a text object,
2184 * do some clearing. Now, you'd think you need only do the flush
2185 * when VTEXT is on, but VTEXT is turned off when the text object
2186 * is freed, while pages are left lying around in memory marked
2187 * with this vnode. If we would reactivate (create a new text
2188 * object from) this vnode, we could easily stumble upon some of
2189 * these old pages in pagein. So, we always flush these guys.
2190 * Sun has a wonderful lack of useful invariants in this system.
2192 * avc->flushDV is the data version # of the file at the last text
2193 * flush. Clearly, at least, we don't have to flush the file more
2194 * often than it changes
2196 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2198 * By here, the cache entry is always write-locked. We can
2199 * deadlock if we call osi_Flush with the cache entry locked...
2200 * Unlock the dcache too.
2202 ReleaseWriteLock(&tdc->lock);
2203 if (setLocks && !slowPass)
2204 ReleaseReadLock(&avc->lock);
2206 ReleaseWriteLock(&avc->lock);
2210 * Call osi_FlushPages in open, read/write, and map, since it
2211 * is too hard here to figure out if we should lock the
2214 if (setLocks && !slowPass)
2215 ObtainReadLock(&avc->lock);
2217 ObtainWriteLock(&avc->lock, 66);
2218 ObtainWriteLock(&tdc->lock, 610);
2223 * avc->lock(R) if setLocks && !slowPass
2224 * avc->lock(W) if !setLocks || slowPass
2228 /* Watch for standard race condition around osi_FlushText */
2229 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2230 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2231 afs_stats_cmperf.dcacheHits++;
2232 ConvertWToSLock(&tdc->lock);
2236 /* Sleep here when cache needs to be drained. */
2237 if (setLocks && !slowPass
2238 && (afs_blocksUsed >
2239 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2240 /* Make sure truncate daemon is running */
2241 afs_MaybeWakeupTruncateDaemon();
2242 ObtainWriteLock(&tdc->tlock, 614);
2243 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2244 ReleaseWriteLock(&tdc->tlock);
2245 ReleaseWriteLock(&tdc->lock);
2246 ReleaseReadLock(&avc->lock);
2247 while ((afs_blocksUsed - afs_blocksDiscarded) >
2248 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2249 afs_WaitForCacheDrain = 1;
2250 afs_osi_Sleep(&afs_WaitForCacheDrain);
2252 afs_MaybeFreeDiscardedDCache();
2253 /* need to check if someone else got the chunk first. */
2254 goto RetryGetDCache;
2257 Position = AFS_CHUNKBASE(abyte);
2258 if (vType(avc) == VDIR) {
2259 size = avc->f.m.Length;
2260 if (size > tdc->f.chunkBytes) {
2261 /* pre-reserve space for file */
2262 afs_AdjustSize(tdc, size);
2264 size = 999999999; /* max size for transfer */
2266 afs_size_t maxGoodLength;
2268 /* estimate how much data we're expecting back from the server,
2269 * and reserve space in the dcache entry for it */
2271 maxGoodLength = avc->f.m.Length;
2272 if (avc->f.truncPos < maxGoodLength)
2273 maxGoodLength = avc->f.truncPos;
2275 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2276 if (Position > maxGoodLength) { /* If we're beyond EOF */
2278 } else if (Position + size > maxGoodLength) {
2279 size = maxGoodLength - Position;
2281 osi_Assert(size >= 0);
2283 if (size > tdc->f.chunkBytes) {
2284 /* pre-reserve estimated space for file */
2285 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2289 /* For the actual fetch, do not limit the request to the
2290 * length of the file. If this results in a read past EOF on
2291 * the server, the server will just reply with less data than
2292 * requested. If we limit ourselves to only requesting data up
2293 * to the avc file length, we open ourselves up to races if the
2294 * file is extended on the server at about the same time.
2296 * However, we must restrict ourselves to the avc->f.truncPos
2297 * length, since this represents an outstanding local
2298 * truncation of the file that will be committed to the
2299 * fileserver when we actually write the fileserver contents.
2300 * If we do not restrict the fetch length based on
2301 * avc->f.truncPos, a different truncate operation extending
2302 * the file length could cause the old data after
2303 * avc->f.truncPos to reappear, instead of extending the file
2304 * with NUL bytes. */
2305 size = AFS_CHUNKSIZE(abyte);
2306 if (Position > avc->f.truncPos) {
2308 } else if (Position + size > avc->f.truncPos) {
2309 size = avc->f.truncPos - Position;
2311 osi_Assert(size >= 0);
2314 if (afs_mariner && !tdc->f.chunk)
2315 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2317 * Right now, we only have one tool, and it's a hammer. So, we
2318 * fetch the whole file.
2320 DZap(tdc); /* pages in cache may be old */
2321 file = afs_CFileOpen(&tdc->f.inode);
2323 /* We can't access the file in the disk cache backing this dcache;
2325 ReleaseWriteLock(&tdc->lock);
2330 afs_RemoveVCB(&avc->f.fid);
2331 tdc->f.states |= DWriting;
2332 tdc->dflags |= DFFetching;
2333 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2334 if (tdc->mflags & DFFetchReq) {
2335 tdc->mflags &= ~DFFetchReq;
2336 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2337 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2338 __FILE__, ICL_TYPE_INT32, __LINE__,
2339 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2342 tsmall = osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2343 setVcacheStatus = 0;
2346 * Remember if we are doing the reading from a replicated volume,
2347 * and how many times we've zipped around the fetch/analyze loop.
2349 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2351 accP = &(afs_stats_cmfullperf.accessinf);
2353 (accP->replicatedRefs)++;
2355 (accP->unreplicatedRefs)++;
2356 #endif /* AFS_NOSTATS */
2357 /* this is a cache miss */
2358 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2359 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2360 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2363 afs_stats_cmperf.dcacheMisses++;
2366 * Dynamic root support: fetch data from local memory.
2368 if (afs_IsDynroot(avc)) {
2372 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2374 dynrootDir += Position;
2375 dynrootLen -= Position;
2376 if (size > dynrootLen)
2380 code = afs_CFileWrite(file, 0, dynrootDir, size);
2388 tdc->validPos = Position + size;
2389 afs_CFileTruncate(file, size); /* prune it */
2390 } else if (afs_IsDynrootMount(avc)) {
2394 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2396 dynrootDir += Position;
2397 dynrootLen -= Position;
2398 if (size > dynrootLen)
2402 code = afs_CFileWrite(file, 0, dynrootDir, size);
2410 tdc->validPos = Position + size;
2411 afs_CFileTruncate(file, size); /* prune it */
2414 * Not a dynamic vnode: do the real fetch.
2419 * avc->lock(R) if setLocks && !slowPass
2420 * avc->lock(W) if !setLocks || slowPass
2424 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
2429 (accP->numReplicasAccessed)++;
2431 #endif /* AFS_NOSTATS */
2432 if (!setLocks || slowPass) {
2433 avc->callback = tc->parent->srvr->server;
2435 newCallback = tc->parent->srvr->server;
2439 code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
2445 /* callback could have been broken (or expired) in a race here,
2446 * but we return the data anyway. It's as good as we knew about
2447 * when we started. */
2449 * validPos is updated by CacheFetchProc, and can only be
2450 * modifed under a dcache write lock, which we've blocked out
2452 size = tdc->validPos - Position; /* actual segment size */
2455 afs_CFileTruncate(file, size); /* prune it */
2457 if (!setLocks || slowPass) {
2458 afs_StaleVCacheFlags(avc, AFS_STALEVC_CLEARCB, CUnique);
2460 /* Something lost. Forget about performance, and go
2461 * back with a vcache write lock.
2463 afs_CFileTruncate(file, 0);
2464 afs_AdjustSize(tdc, 0);
2465 afs_CFileClose(file);
2466 osi_FreeLargeSpace(tsmall);
2468 ReleaseWriteLock(&tdc->lock);
2471 ReleaseReadLock(&avc->lock);
2474 /* If we have a connection, we must put it back,
2475 * since afs_Analyze will not be called here. */
2476 afs_PutConn(tc, rxconn, SHARED_LOCK);
2480 goto RetryGetDCache;
2484 } while (afs_Analyze
2485 (tc, rxconn, code, &avc->f.fid, areq,
2486 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2490 * avc->lock(R) if setLocks && !slowPass
2491 * avc->lock(W) if !setLocks || slowPass
2497 * In the case of replicated access, jot down info on the number of
2498 * attempts it took before we got through or gave up.
2501 if (numFetchLoops <= 1)
2502 (accP->refFirstReplicaOK)++;
2503 if (numFetchLoops > accP->maxReplicasPerRef)
2504 accP->maxReplicasPerRef = numFetchLoops;
2506 #endif /* AFS_NOSTATS */
2508 tdc->dflags &= ~DFFetching;
2509 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2510 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2511 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2512 tdc, ICL_TYPE_INT32, tdc->dflags);
2513 if (avc->execsOrWriters == 0)
2514 tdc->f.states &= ~DWriting;
2516 /* now, if code != 0, we have an error and should punt.
2517 * note that we have the vcache write lock, either because
2518 * !setLocks or slowPass.
2521 afs_CFileTruncate(file, 0);
2522 afs_AdjustSize(tdc, 0);
2523 afs_CFileClose(file);
2524 ZapDCE(tdc); /* sets DFEntryMod */
2525 if (vType(avc) == VDIR) {
2528 tdc->f.states &= ~(DRO|DBackup|DRW);
2529 afs_DCMoveBucket(tdc, 0, 0);
2530 ReleaseWriteLock(&tdc->lock);
2532 if (!afs_IsDynroot(avc)) {
2533 afs_StaleVCacheFlags(avc, 0, CUnique);
2536 * avc->lock(W); assert(!setLocks || slowPass)
2538 osi_Assert(!setLocks || slowPass);
2544 /* otherwise we copy in the just-fetched info */
2545 afs_CFileClose(file);
2546 afs_AdjustSize(tdc, size); /* new size */
2548 * Copy appropriate fields into vcache. Status is
2549 * copied later where we selectively acquire the
2550 * vcache write lock.
2553 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2555 setVcacheStatus = 1;
2556 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2557 tsmall->OutStatus.DataVersion);
2558 tdc->dflags |= DFEntryMod;
2559 afs_indexFlags[tdc->index] |= IFEverUsed;
2560 ConvertWToSLock(&tdc->lock);
2561 } /*Data version numbers don't match */
2564 * Data version numbers match.
2566 afs_stats_cmperf.dcacheHits++;
2567 } /*Data version numbers match */
2569 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2573 * avc->lock(R) if setLocks && !slowPass
2574 * avc->lock(W) if !setLocks || slowPass
2575 * tdc->lock(S) if tdc
2579 * See if this was a reference to a file in the local cell.
2581 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2582 afs_stats_cmperf.dlocalAccesses++;
2584 afs_stats_cmperf.dremoteAccesses++;
2586 /* Fix up LRU info */
2589 ObtainWriteLock(&afs_xdcache, 602);
2590 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2591 hadd32(afs_indexCounter, 1);
2592 ReleaseWriteLock(&afs_xdcache);
2594 /* return the data */
2595 if (vType(avc) == VDIR)
2598 *aoffset = AFS_CHUNKOFFSET(abyte);
2599 *alen = (tdc->f.chunkBytes - *aoffset);
2600 ReleaseSharedLock(&tdc->lock);
2605 * avc->lock(R) if setLocks && !slowPass
2606 * avc->lock(W) if !setLocks || slowPass
2609 /* Fix up the callback and status values in the vcache */
2611 if (setLocks && !slowPass) {
2614 * This is our dirty little secret to parallel fetches.
2615 * We don't write-lock the vcache while doing the fetch,
2616 * but potentially we'll need to update the vcache after
2617 * the fetch is done.
2619 * Drop the read lock and try to re-obtain the write
2620 * lock. If the vcache still has the same DV, it's
2621 * ok to go ahead and install the new data.
2623 afs_hyper_t currentDV, statusDV;
2625 hset(currentDV, avc->f.m.DataVersion);
2627 if (setNewCallback && avc->callback != newCallback)
2631 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2632 tsmall->OutStatus.DataVersion);
2634 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2636 if (setVcacheStatus && !hsame(currentDV, statusDV))
2640 ReleaseReadLock(&avc->lock);
2642 if (doVcacheUpdate) {
2643 ObtainWriteLock(&avc->lock, 615);
2644 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2645 /* We lose. Someone will beat us to it. */
2647 ReleaseWriteLock(&avc->lock);
2652 /* With slow pass, we've already done all the updates */
2654 ReleaseWriteLock(&avc->lock);
2657 /* Check if we need to perform any last-minute fixes with a write-lock */
2658 if (!setLocks || doVcacheUpdate) {
2660 avc->callback = newCallback;
2661 if (tsmall && setVcacheStatus)
2662 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2664 ReleaseWriteLock(&avc->lock);
2668 osi_FreeLargeSpace(tsmall);
2671 } /*afs_GetDCache */
2675 * afs_WriteThroughDSlots
2678 * Sweep through the dcache slots and write out any modified
2679 * in-memory data back on to our caching store.
2685 * The afs_xdcache is write-locked through this whole affair.
2688 afs_WriteThroughDSlots(void)
2691 afs_int32 i, touchedit = 0;
2694 struct afs_q DirtyQ, *tq;
2696 AFS_STATCNT(afs_WriteThroughDSlots);
2699 * Because of lock ordering, we can't grab dcache locks while
2700 * holding afs_xdcache. So we enter xdcache, get a reference
2701 * for every dcache entry, and exit xdcache.
2703 ObtainWriteLock(&afs_xdcache, 283);
2705 for (i = 0; i < afs_cacheFiles; i++) {
2706 tdc = afs_indexTable[i];
2708 /* Grab tlock in case the existing refcount isn't zero */
2709 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2710 ObtainWriteLock(&tdc->tlock, 623);
2712 ReleaseWriteLock(&tdc->tlock);
2714 QAdd(&DirtyQ, &tdc->dirty);
2717 ReleaseWriteLock(&afs_xdcache);
2720 * Now, for each dcache entry we found, check if it's dirty.
2721 * If so, get write-lock, get afs_xdcache, which protects
2722 * afs_cacheInodep, and flush it. Don't forget to put back
2726 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2728 for (tq = DirtyQ.prev; tq != &DirtyQ && code == 0; tq = QPrev(tq)) {
2730 if (tdc->dflags & DFEntryMod) {
2733 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2735 /* Now that we have the write lock, double-check */
2736 if (wrLock && (tdc->dflags & DFEntryMod)) {
2737 tdc->dflags &= ~DFEntryMod;
2738 ObtainWriteLock(&afs_xdcache, 620);
2739 code = afs_WriteDCache(tdc, 1);
2740 ReleaseWriteLock(&afs_xdcache);
2742 /* We didn't successfully write out the dslot; make sure we
2743 * try again later */
2744 tdc->dflags |= DFEntryMod;
2750 ReleaseWriteLock(&tdc->lock);
2760 ObtainWriteLock(&afs_xdcache, 617);
2761 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2762 /* Touch the file to make sure that the mtime on the file is kept
2763 * up-to-date to avoid losing cached files on cold starts because
2764 * their mtime seems old...
2766 struct afs_fheader theader;
2768 afs_InitFHeader(&theader);
2769 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2771 ReleaseWriteLock(&afs_xdcache);
2779 * Return a pointer to an freshly initialized dcache entry using
2780 * a memory-based cache. The tlock will be read-locked.
2783 * aslot : Dcache slot to look at.
2784 * type : What 'type' of dslot to get; see the dslot_state enum
2787 * Must be called with afs_xdcache write-locked.
2791 afs_MemGetDSlot(afs_int32 aslot, dslot_state type)
2796 AFS_STATCNT(afs_MemGetDSlot);
2797 if (CheckLock(&afs_xdcache) != -1)
2798 osi_Panic("getdslot nolock");
2799 if (aslot < 0 || aslot >= afs_cacheFiles)
2800 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2801 tdc = afs_indexTable[aslot];
2803 QRemove(&tdc->lruq); /* move to queue head */
2804 QAdd(&afs_DLRU, &tdc->lruq);
2805 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2806 ObtainWriteLock(&tdc->tlock, 624);
2808 ConvertWToRLock(&tdc->tlock);
2812 /* if we got here, the given slot is not in memory in our list of known
2813 * slots. for memcache, the only place a dslot can exist is in memory, so
2814 * if the caller is expecting to get back a known dslot, and we've reached
2815 * here, something is very wrong. DSLOT_NEW is the only type of dslot that
2816 * may not exist; for all others, the caller assumes the given dslot
2817 * already exists. so, 'type' had better be DSLOT_NEW here, or something is
2819 osi_Assert(type == DSLOT_NEW);
2821 if (!afs_freeDSList)
2822 afs_GetDownDSlot(4);
2823 if (!afs_freeDSList) {
2824 /* none free, making one is better than a panic */
2825 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2826 tdc = afs_osi_Alloc(sizeof(struct dcache));
2827 osi_Assert(tdc != NULL);
2828 #ifdef KERNEL_HAVE_PIN
2829 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2832 tdc = afs_freeDSList;
2833 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2836 tdc->dflags = 0; /* up-to-date, not in free q */
2838 QAdd(&afs_DLRU, &tdc->lruq);
2839 if (tdc->lruq.prev == &tdc->lruq)
2840 osi_Panic("lruq 3");
2842 /* initialize entry */
2843 tdc->f.fid.Cell = 0;
2844 tdc->f.fid.Fid.Volume = 0;
2846 hones(tdc->f.versionNo);
2847 tdc->f.inode.mem = aslot;
2848 tdc->dflags |= DFEntryMod;
2851 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2854 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2855 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2856 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2859 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2860 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2861 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2862 ObtainReadLock(&tdc->tlock);
2864 afs_indexTable[aslot] = tdc;
2867 } /*afs_MemGetDSlot */
2869 unsigned int last_error = 0, lasterrtime = 0;
2875 * Return a pointer to an freshly initialized dcache entry using
2876 * a UFS-based disk cache. The dcache tlock will be read-locked.
2879 * aslot : Dcache slot to look at.
2880 * type : What 'type' of dslot to get; see the dslot_state enum
2883 * afs_xdcache lock write-locked.
2886 afs_UFSGetDSlot(afs_int32 aslot, dslot_state type)
2894 AFS_STATCNT(afs_UFSGetDSlot);
2895 if (CheckLock(&afs_xdcache) != -1)
2896 osi_Panic("getdslot nolock");
2897 if (aslot < 0 || aslot >= afs_cacheFiles)
2898 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2899 tdc = afs_indexTable[aslot];
2901 QRemove(&tdc->lruq); /* move to queue head */
2902 QAdd(&afs_DLRU, &tdc->lruq);
2903 /* Grab tlock in case refCount != 0 */
2904 ObtainWriteLock(&tdc->tlock, 625);
2906 ConvertWToRLock(&tdc->tlock);
2910 /* otherwise we should read it in from the cache file */
2911 if (!afs_freeDSList)
2912 afs_GetDownDSlot(4);
2913 if (!afs_freeDSList) {
2914 /* none free, making one is better than a panic */
2915 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2916 tdc = afs_osi_Alloc(sizeof(struct dcache));
2917 osi_Assert(tdc != NULL);
2918 #ifdef KERNEL_HAVE_PIN
2919 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2922 tdc = afs_freeDSList;
2923 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2926 tdc->dflags = 0; /* up-to-date, not in free q */
2928 QAdd(&afs_DLRU, &tdc->lruq);
2929 if (tdc->lruq.prev == &tdc->lruq)
2930 osi_Panic("lruq 3");
2933 * Seek to the aslot'th entry and read it in.
2935 off = sizeof(struct fcache)*aslot + sizeof(struct afs_fheader);
2937 afs_osi_Read(afs_cacheInodep,
2938 off, (char *)(&tdc->f),
2939 sizeof(struct fcache));
2941 if (code != sizeof(struct fcache)) {
2943 #if defined(KERNEL_HAVE_UERROR)
2944 last_error = getuerror();
2948 lasterrtime = osi_Time();
2949 if (type != DSLOT_NEW) {
2950 /* If we are requesting a non-DSLOT_NEW slot, this is an error.
2951 * non-DSLOT_NEW slots are supposed to already exist, so if we
2952 * failed to read in the slot, something is wrong. */
2953 struct osi_stat tstat;
2954 if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
2957 afs_warn("afs: disk cache read error in CacheItems slot %d "
2958 "off %d/%d code %d/%d\n",
2960 off, (int)tstat.size,
2961 (int)code, (int)sizeof(struct fcache));
2962 /* put tdc back on the free dslot list */
2963 QRemove(&tdc->lruq);
2964 tdc->index = NULLIDX;
2965 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
2966 afs_freeDSList = tdc;
2970 if (!afs_CellNumValid(tdc->f.fid.Cell)) {
2972 if (type == DSLOT_VALID) {
2973 osi_Panic("afs: needed valid dcache but index %d off %d has "
2974 "invalid cell num %d\n",
2975 (int)aslot, off, (int)tdc->f.fid.Cell);
2979 if (type == DSLOT_VALID && tdc->f.fid.Fid.Volume == 0) {
2980 osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
2984 if (type == DSLOT_UNUSED) {
2985 /* the requested dslot is known to exist, but contain invalid data
2986 * (this happens when we're using a dslot from the free or discard
2987 * list). be sure not to re-use the data in it, so force invalidation.
2993 tdc->f.fid.Cell = 0;
2994 tdc->f.fid.Fid.Volume = 0;
2996 hones(tdc->f.versionNo);
2997 tdc->dflags |= DFEntryMod;
2998 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2999 tdc->f.states &= ~(DRO|DBackup|DRW);
3000 afs_DCMoveBucket(tdc, 0, 0);
3002 if (tdc->f.states & DRO) {
3003 afs_DCMoveBucket(tdc, 0, 2);
3004 } else if (tdc->f.states & DBackup) {
3005 afs_DCMoveBucket(tdc, 0, 1);
3007 afs_DCMoveBucket(tdc, 0, 1);
3012 if (tdc->f.chunk >= 0)
3013 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
3018 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
3019 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
3020 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
3023 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
3024 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
3025 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
3026 ObtainReadLock(&tdc->tlock);
3029 * If we didn't read into a temporary dcache region, update the
3030 * slot pointer table.
3032 afs_indexTable[aslot] = tdc;
3035 } /*afs_UFSGetDSlot */
3040 * Write a particular dcache entry back to its home in the
3043 * \param adc Pointer to the dcache entry to write.
3044 * \param atime If true, set the modtime on the file to the current time.
3046 * \note Environment:
3047 * Must be called with the afs_xdcache lock at least read-locked,
3048 * and dcache entry at least read-locked.
3049 * The reference count is not changed.
3053 afs_WriteDCache(struct dcache *adc, int atime)
3057 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
3059 AFS_STATCNT(afs_WriteDCache);
3060 osi_Assert(WriteLocked(&afs_xdcache));
3062 adc->f.modTime = osi_Time();
3064 if ((afs_indexFlags[adc->index] & (IFFree | IFDiscarded)) == 0 &&
3065 adc->f.fid.Fid.Volume == 0) {
3066 /* If a dcache slot is not on the free or discard list, it must be
3067 * in the hash table. Thus, the volume must be non-zero, since that
3068 * is how we determine whether or not to unhash the entry when kicking
3069 * it out of the cache. Do this check now, since otherwise this can
3070 * cause hash table corruption and a panic later on after we read the
3072 osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
3073 adc->index, (unsigned)afs_indexFlags[adc->index]);
3077 * Seek to the right dcache slot and write the in-memory image out to disk.
3079 afs_cellname_write();
3081 afs_osi_Write(afs_cacheInodep,
3082 sizeof(struct fcache) * adc->index +
3083 sizeof(struct afs_fheader), (char *)(&adc->f),
3084 sizeof(struct fcache));
3085 if (code != sizeof(struct fcache)) {
3086 afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
3087 (long)(sizeof(struct fcache) * adc->index + sizeof(struct afs_fheader)),
3088 (int)code, (int)sizeof(struct fcache));
3097 * Wake up users of a particular file waiting for stores to take
3100 * \param avc Ptr to related vcache entry.
3102 * \note Environment:
3103 * Nothing interesting.
3106 afs_wakeup(struct vcache *avc)
3109 struct brequest *tb;
3111 AFS_STATCNT(afs_wakeup);
3112 for (i = 0; i < NBRS; i++, tb++) {
3113 /* if request is valid and for this file, we've found it */
3114 if (tb->refCount > 0 && avc == tb->vc) {
3117 * If CSafeStore is on, then we don't awaken the guy
3118 * waiting for the store until the whole store has finished.
3119 * Otherwise, we do it now. Note that if CSafeStore is on,
3120 * the BStore routine actually wakes up the user, instead
3122 * I think this is redundant now because this sort of thing
3123 * is already being handled by the higher-level code.
3125 if ((avc->f.states & CSafeStore) == 0) {
3126 tb->code_raw = tb->code_checkcode = 0;
3127 tb->flags |= BUVALID;
3128 if (tb->flags & BUWAIT) {
3129 tb->flags &= ~BUWAIT;
3140 * Given a file name and inode, set up that file to be an
3141 * active member in the AFS cache. This also involves checking
3142 * the usability of its data.
3144 * \param afile Name of the cache file to initialize.
3145 * \param ainode Inode of the file.
3147 * \note Environment:
3148 * This function is called only during initialization.
3151 afs_InitCacheFile(char *afile, ino_t ainode)
3156 struct osi_file *tfile;
3157 struct osi_stat tstat;
3160 AFS_STATCNT(afs_InitCacheFile);
3161 index = afs_stats_cmperf.cacheNumEntries;
3162 if (index >= afs_cacheFiles)
3165 ObtainWriteLock(&afs_xdcache, 282);
3166 tdc = afs_GetNewDSlot(index);
3167 ReleaseReadLock(&tdc->tlock);
3168 ReleaseWriteLock(&afs_xdcache);
3170 ObtainWriteLock(&tdc->lock, 621);
3171 ObtainWriteLock(&afs_xdcache, 622);
3172 if (!afile && !ainode) {
3177 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
3179 ReleaseWriteLock(&afs_xdcache);
3180 ReleaseWriteLock(&tdc->lock);
3185 /* Add any other 'complex' inode types here ... */
3186 #if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
3187 tdc->f.inode.ufs = ainode;
3189 osi_Panic("Can't init cache with inode numbers when complex inodes are "
3194 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3196 tfile = osi_UFSOpen(&tdc->f.inode);
3198 ReleaseWriteLock(&afs_xdcache);
3199 ReleaseWriteLock(&tdc->lock);
3204 code = afs_osi_Stat(tfile, &tstat);
3206 osi_Panic("initcachefile stat");
3209 * If file size doesn't match the cache info file, it's probably bad.
3211 if (tdc->f.chunkBytes != tstat.size)
3214 * If file changed within T (120?) seconds of cache info file, it's
3215 * probably bad. In addition, if slot changed within last T seconds,
3216 * the cache info file may be incorrectly identified, and so slot
3219 if (cacheInfoModTime < tstat.mtime + 120)
3221 if (cacheInfoModTime < tdc->f.modTime + 120)
3223 /* In case write through is behind, make sure cache items entry is
3224 * at least as new as the chunk.
3226 if (tdc->f.modTime < tstat.mtime)
3229 tdc->f.chunkBytes = 0;
3232 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3233 if (tfile && tstat.size != 0)
3234 osi_UFSTruncate(tfile, 0);
3235 tdc->f.states &= ~(DRO|DBackup|DRW);
3236 afs_DCMoveBucket(tdc, 0, 0);
3237 /* put entry in free cache slot list */
3238 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3239 afs_freeDCList = index;
3241 afs_indexFlags[index] |= IFFree;
3242 afs_indexUnique[index] = 0;
3245 * We must put this entry in the appropriate hash tables.
3246 * Note that i is still set from the above DCHash call
3248 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3249 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3250 afs_dchashTbl[code] = tdc->index;
3251 code = DVHash(&tdc->f.fid);
3252 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3253 afs_dvhashTbl[code] = tdc->index;
3254 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3256 /* has nontrivial amt of data */
3257 afs_indexFlags[index] |= IFEverUsed;
3258 afs_stats_cmperf.cacheFilesReused++;
3260 * Initialize index times to file's mod times; init indexCounter
3263 hset32(afs_indexTimes[index], tstat.atime);
3264 if (hgetlo(afs_indexCounter) < tstat.atime) {
3265 hset32(afs_indexCounter, tstat.atime);
3267 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3268 } /*File is not bad */
3271 osi_UFSClose(tfile);
3272 tdc->f.states &= ~DWriting;
3273 tdc->dflags &= ~DFEntryMod;
3274 /* don't set f.modTime; we're just cleaning up */
3275 osi_Assert(afs_WriteDCache(tdc, 0) == 0);
3276 ReleaseWriteLock(&afs_xdcache);
3277 ReleaseWriteLock(&tdc->lock);
3279 afs_stats_cmperf.cacheNumEntries++;
3284 /*Max # of struct dcache's resident at any time*/
3286 * If 'dchint' is enabled then in-memory dcache min is increased because of
3292 * Initialize dcache related variables.
3302 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3309 afs_freeDCList = NULLIDX;
3310 afs_discardDCList = NULLIDX;
3311 afs_freeDCCount = 0;
3312 afs_freeDSList = NULL;
3313 hzero(afs_indexCounter);
3315 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3321 if (achunk < 0 || achunk > 30)
3322 achunk = 13; /* Use default */
3323 AFS_SETCHUNKSIZE(achunk);
3329 /* afs_dhashsize defaults to 1024 */
3330 if (aDentries > 512)
3331 afs_dhashsize = 2048;
3332 /* Try to keep the average chain length around two unless the table
3333 * would be ridiculously big. */
3334 if (aDentries > 4096) {
3335 afs_dhashbits = opr_fls(aDentries) - 3;
3336 /* Cap the hash tables to 32k entries. */
3337 if (afs_dhashbits > 15)
3339 afs_dhashsize = opr_jhash_size(afs_dhashbits);
3341 /* initialize hash tables */
3342 afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3343 osi_Assert(afs_dvhashTbl != NULL);
3344 afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3345 osi_Assert(afs_dchashTbl != NULL);
3346 for (i = 0; i < afs_dhashsize; i++) {
3347 afs_dvhashTbl[i] = NULLIDX;
3348 afs_dchashTbl[i] = NULLIDX;
3350 afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3351 osi_Assert(afs_dvnextTbl != NULL);
3352 afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3353 osi_Assert(afs_dcnextTbl != NULL);
3354 for (i = 0; i < afiles; i++) {
3355 afs_dvnextTbl[i] = NULLIDX;
3356 afs_dcnextTbl[i] = NULLIDX;
3359 /* Allocate and zero the pointer array to the dcache entries */
3360 afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3361 osi_Assert(afs_indexTable != NULL);
3362 memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3363 afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3364 osi_Assert(afs_indexTimes != NULL);
3365 memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3366 afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
3367 osi_Assert(afs_indexUnique != NULL);
3368 memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3369 afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
3370 osi_Assert(afs_indexFlags != NULL);
3371 memset(afs_indexFlags, 0, afiles * sizeof(char));
3373 /* Allocate and thread the struct dcache entries themselves */
3374 tdp = afs_Initial_freeDSList =
3375 afs_osi_Alloc(aDentries * sizeof(struct dcache));
3376 osi_Assert(tdp != NULL);
3377 memset(tdp, 0, aDentries * sizeof(struct dcache));
3378 #ifdef KERNEL_HAVE_PIN
3379 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3380 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3381 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3382 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3383 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3384 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3385 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3386 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3387 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3390 afs_freeDSList = &tdp[0];
3391 for (i = 0; i < aDentries - 1; i++) {
3392 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3393 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3394 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3395 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3397 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3398 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3399 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3400 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3402 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3403 afs_cacheBlocks = ablocks;
3404 afs_ComputeCacheParms(); /* compute parms based on cache size */
3406 afs_dcentries = aDentries;
3408 afs_stats_cmperf.cacheBucket0_Discarded =
3409 afs_stats_cmperf.cacheBucket1_Discarded =
3410 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3414 if (aflags & AFSCALL_INIT_MEMCACHE) {
3416 * Use a memory cache instead of a disk cache
3418 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3419 afs_cacheType = &afs_MemCacheOps;
3420 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3421 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3422 /* ablocks is reported in 1K blocks */
3423 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3425 afs_warn("afsd: memory cache too large for available memory.\n");
3426 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3429 afs_warn("Memory cache: Allocating %d dcache entries...",
3432 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3433 afs_cacheType = &afs_UfsCacheOps;
3438 * Shuts down the cache.
3442 shutdown_dcache(void)
3446 #ifdef AFS_CACHE_VNODE_PATH
3447 if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
3449 for (i = 0; i < afs_cacheFiles; i++) {
3450 tdc = afs_indexTable[i];
3452 afs_osi_FreeStr(tdc->f.inode.ufs);
3458 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3459 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3460 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3461 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3462 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3463 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3464 afs_osi_Free(afs_Initial_freeDSList,
3465 afs_dcentries * sizeof(struct dcache));
3466 #ifdef KERNEL_HAVE_PIN
3467 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3468 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3469 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3470 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3471 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3472 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3473 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3477 for (i = 0; i < afs_dhashsize; i++) {
3478 afs_dvhashTbl[i] = NULLIDX;
3479 afs_dchashTbl[i] = NULLIDX;
3482 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3483 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3485 afs_blocksUsed = afs_dcentries = 0;
3486 afs_stats_cmperf.cacheBucket0_Discarded =
3487 afs_stats_cmperf.cacheBucket1_Discarded =
3488 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3489 hzero(afs_indexCounter);
3491 afs_freeDCCount = 0;
3492 afs_freeDCList = NULLIDX;
3493 afs_discardDCList = NULLIDX;
3494 afs_freeDSList = afs_Initial_freeDSList = 0;
3496 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3502 * Get a dcache ready for writing, respecting the current cache size limits
3504 * len is required because afs_GetDCache with flag == 4 expects the length
3505 * field to be filled. It decides from this whether it's necessary to fetch
3506 * data into the chunk before writing or not (when the whole chunk is
3509 * \param avc The vcache to fetch a dcache for
3510 * \param filePos The start of the section to be written
3511 * \param len The length of the section to be written
3515 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3516 * must be released and afs_PutDCache() called to free dcache.
3519 * \note avc->lock must be held on entry. Function may release and reobtain
3520 * avc->lock and GLOCK.
3524 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3525 afs_size_t len, struct vrequest *areq,
3528 struct dcache *tdc = NULL;
3531 /* read the cached info */
3533 tdc = afs_FindDCache(avc, filePos);
3535 ObtainWriteLock(&tdc->lock, 657);
3536 } else if (afs_blocksUsed >
3537 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3538 tdc = afs_FindDCache(avc, filePos);
3540 ObtainWriteLock(&tdc->lock, 658);
3541 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3542 || (tdc->dflags & DFFetching)) {
3543 ReleaseWriteLock(&tdc->lock);
3549 afs_MaybeWakeupTruncateDaemon();
3550 while (afs_blocksUsed >
3551 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3552 ReleaseWriteLock(&avc->lock);
3553 if (afs_blocksUsed - afs_blocksDiscarded >
3554 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3555 afs_WaitForCacheDrain = 1;
3556 afs_osi_Sleep(&afs_WaitForCacheDrain);
3558 afs_MaybeFreeDiscardedDCache();
3559 afs_MaybeWakeupTruncateDaemon();
3560 ObtainWriteLock(&avc->lock, 509);
3562 avc->f.states |= CDirty;
3563 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3565 ObtainWriteLock(&tdc->lock, 659);
3568 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3570 ObtainWriteLock(&tdc->lock, 660);
3573 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3574 afs_stats_cmperf.cacheCurrDirtyChunks++;
3575 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3577 if (!(tdc->f.states & DWriting)) {
3578 /* don't mark entry as mod if we don't have to */
3579 tdc->f.states |= DWriting;
3580 tdc->dflags |= DFEntryMod;
3587 * Make a shadow copy of a dir's dcache. It's used for disconnected
3588 * operations like remove/create/rename to keep the original directory data.
3589 * On reconnection, we can diff the original data with the server and get the
3590 * server changes and with the local data to get the local changes.
3592 * \param avc The dir vnode.
3593 * \param adc The dir dcache.
3595 * \return 0 for success.
3597 * \note The vcache entry must be write locked.
3598 * \note The dcache entry must be read locked.
3601 afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3603 int i, code, ret_code = 0, written, trans_size;
3604 struct dcache *new_dc = NULL;
3605 struct osi_file *tfile_src, *tfile_dst;
3606 struct VenusFid shadow_fid;
3609 /* Is this a dir? */
3610 if (vType(avc) != VDIR)
3613 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3616 /* Generate a fid for the shadow dir. */
3617 shadow_fid.Cell = avc->f.fid.Cell;
3618 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3619 afs_GenShadowFid(&shadow_fid);
3621 ObtainWriteLock(&afs_xdcache, 716);
3623 /* Get a fresh dcache. */
3624 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3627 ObtainReadLock(&adc->mflock);
3629 /* Set up the new fid. */
3630 /* Copy interesting data from original dir dcache. */
3631 new_dc->mflags = adc->mflags;
3632 new_dc->dflags = adc->dflags;
3633 new_dc->f.modTime = adc->f.modTime;
3634 new_dc->f.versionNo = adc->f.versionNo;
3635 new_dc->f.states = adc->f.states;
3636 new_dc->f.chunk= adc->f.chunk;
3637 new_dc->f.chunkBytes = adc->f.chunkBytes;
3639 ReleaseReadLock(&adc->mflock);
3641 /* Now add to the two hash chains */
3642 i = DCHash(&shadow_fid, 0);
3643 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3644 afs_dchashTbl[i] = new_dc->index;
3646 i = DVHash(&shadow_fid);
3647 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3648 afs_dvhashTbl[i] = new_dc->index;
3650 ReleaseWriteLock(&afs_xdcache);
3652 /* Alloc a 4k block. */
3653 data = afs_osi_Alloc(4096);
3655 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3660 /* Open the files. */
3661 tfile_src = afs_CFileOpen(&adc->f.inode);
3662 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3663 osi_Assert(tfile_src);
3664 osi_Assert(tfile_dst);
3666 /* And now copy dir dcache data into this dcache,
3670 while (written < adc->f.chunkBytes) {
3671 trans_size = adc->f.chunkBytes - written;
3672 if (trans_size > 4096)
3675 /* Read a chunk from the dcache. */
3676 code = afs_CFileRead(tfile_src, written, data, trans_size);
3677 if (code < trans_size) {
3682 /* Write it to the new dcache. */
3683 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3684 if (code < trans_size) {
3689 written+=trans_size;
3692 afs_CFileClose(tfile_dst);
3693 afs_CFileClose(tfile_src);
3695 afs_osi_Free(data, 4096);
3697 ReleaseWriteLock(&new_dc->lock);
3698 afs_PutDCache(new_dc);
3701 ObtainWriteLock(&afs_xvcache, 763);
3702 ObtainWriteLock(&afs_disconDirtyLock, 765);
3703 QAdd(&afs_disconShadow, &avc->shadowq);
3704 osi_Assert((afs_RefVCache(avc) == 0));
3705 ReleaseWriteLock(&afs_disconDirtyLock);
3706 ReleaseWriteLock(&afs_xvcache);
3708 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3709 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3717 * Delete the dcaches of a shadow dir.
3719 * \param avc The vcache containing the shadow fid.
3721 * \note avc must be write locked.
3724 afs_DeleteShadowDir(struct vcache *avc)
3727 struct VenusFid shadow_fid;
3729 shadow_fid.Cell = avc->f.fid.Cell;
3730 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3731 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3732 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3734 tdc = afs_FindDCacheByFid(&shadow_fid);
3736 afs_HashOutDCache(tdc, 1);
3737 afs_DiscardDCache(tdc);
3740 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3741 ObtainWriteLock(&afs_disconDirtyLock, 708);
3742 QRemove(&avc->shadowq);
3743 ReleaseWriteLock(&afs_disconDirtyLock);
3744 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3748 * Populate a dcache with empty chunks up to a given file size,
3749 * used before extending a file in order to avoid 'holes' which
3750 * we can't access in disconnected mode.
3752 * \param avc The vcache which is being extended (locked)
3753 * \param alen The new length of the file
3757 afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
3760 afs_size_t len, offset;
3761 afs_int32 start, end;
3763 /* We're doing this to deal with the situation where we extend
3764 * by writing after lseek()ing past the end of the file . If that
3765 * extension skips chunks, then those chunks won't be created, and
3766 * GetDCache will assume that they have to be fetched from the server.
3767 * So, for each chunk between the current file position, and the new
3768 * length we GetDCache for that chunk.
3771 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3774 if (avc->f.m.Length == 0)
3777 start = AFS_CHUNK(avc->f.m.Length)+1;
3779 end = AFS_CHUNK(apos);
3782 len = AFS_CHUNKTOSIZE(start);
3783 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);