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;
98 afs_uint32 afs_CacheTooFullCount
= 0;
99 afs_uint32 afs_WaitForCacheDrainCount
= 0;
101 afs_int32 afs_dcentries
; /*!< In-memory dcache entries */
104 int dcacheDisabled
= 0;
106 struct afs_cacheOps afs_UfsCacheOps
= {
107 #ifndef HAVE_STRUCT_LABEL_SUPPORT
120 .truncate
= osi_UFSTruncate
,
121 .fread
= afs_osi_Read
,
122 .fwrite
= afs_osi_Write
,
123 .close
= osi_UFSClose
,
124 .vreadUIO
= afs_UFSReadUIO
,
125 .vwriteUIO
= afs_UFSWriteUIO
,
126 .GetDSlot
= afs_UFSGetDSlot
,
127 .GetVolSlot
= afs_UFSGetVolSlot
,
128 .HandleLink
= afs_UFSHandleLink
,
132 struct afs_cacheOps afs_MemCacheOps
= {
133 #ifndef HAVE_STRUCT_LABEL_SUPPORT
135 afs_MemCacheTruncate
,
145 .open
= afs_MemCacheOpen
,
146 .truncate
= afs_MemCacheTruncate
,
147 .fread
= afs_MemReadBlk
,
148 .fwrite
= afs_MemWriteBlk
,
149 .close
= afs_MemCacheClose
,
150 .vreadUIO
= afs_MemReadUIO
,
151 .vwriteUIO
= afs_MemWriteUIO
,
152 .GetDSlot
= afs_MemGetDSlot
,
153 .GetVolSlot
= afs_MemGetVolSlot
,
154 .HandleLink
= afs_MemHandleLink
,
158 int cacheDiskType
; /*Type of backing disk for cache */
159 struct afs_cacheOps
*afs_cacheType
;
161 /* calculate number of 1k blocks needed, rounded up to nearest afs_fsfragsize */
162 static_inline afs_int32
163 afs_round_to_fsfragsize(afs_int32 bytes
)
165 return (((bytes
+ afs_fsfragsize
) ^ afs_fsfragsize
) >> 10);
169 * The PFlush algorithm makes use of the fact that Fid.Unique is not used in
170 * below hash algorithms. Change it if need be so that flushing algorithm
171 * doesn't move things from one hash chain to another.
173 /*Vnode, Chunk -> Hash table index */
174 int DCHash(struct VenusFid
*fid
, afs_int32 chunk
)
178 buf
[0] = fid
->Fid
.Volume
;
179 buf
[1] = fid
->Fid
.Vnode
;
181 return opr_jhash(buf
, 3, 0) & (afs_dhashsize
- 1);
183 /*Vnode -> Other hash table index */
184 int DVHash(struct VenusFid
*fid
)
186 return opr_jhash_int2(fid
->Fid
.Volume
, fid
->Fid
.Vnode
, 0) &
191 * Where is this vcache's entry associated dcache located/
192 * \param avc The vcache entry.
193 * \return Bucket index:
198 afs_DCGetBucket(struct vcache
*avc
)
203 /* This should be replaced with some sort of user configurable function */
204 if (avc
->f
.states
& CRO
) {
206 } else if (avc
->f
.states
& CBackup
) {
216 * Readjust a dcache's size.
218 * \param adc The dcache to be adjusted.
219 * \param oldSize Old size for the dcache.
220 * \param newSize The new size to be adjusted to.
224 afs_DCAdjustSize(struct dcache
*adc
, afs_int32 oldSize
, afs_int32 newSize
)
226 afs_int32 adjustSize
= newSize
- oldSize
;
234 afs_blocksUsed_0
+= adjustSize
;
235 afs_stats_cmperf
.cacheBucket0_Discarded
+= oldSize
;
238 afs_blocksUsed_1
+= adjustSize
;
239 afs_stats_cmperf
.cacheBucket1_Discarded
+= oldSize
;
242 afs_blocksUsed_2
+= adjustSize
;
243 afs_stats_cmperf
.cacheBucket2_Discarded
+= oldSize
;
251 * Move a dcache from one bucket to another.
253 * \param adc Operate on this dcache.
254 * \param size Size in bucket (?).
255 * \param newBucket Destination bucket.
259 afs_DCMoveBucket(struct dcache
*adc
, afs_int32 size
, afs_int32 newBucket
)
264 /* Substract size from old bucket. */
268 afs_blocksUsed_0
-= size
;
271 afs_blocksUsed_1
-= size
;
274 afs_blocksUsed_2
-= size
;
278 /* Set new bucket and increase destination bucket size. */
279 adc
->bucket
= newBucket
;
284 afs_blocksUsed_0
+= size
;
287 afs_blocksUsed_1
+= size
;
290 afs_blocksUsed_2
+= size
;
298 * Init split caches size.
303 afs_blocksUsed_0
= afs_blocksUsed_1
= afs_blocksUsed_2
= 0;
312 afs_DCWhichBucket(afs_int32 phase
, afs_int32 bucket
)
317 afs_pct1
= afs_blocksUsed_1
/ (afs_cacheBlocks
/ 100);
318 afs_pct2
= afs_blocksUsed_2
/ (afs_cacheBlocks
/ 100);
320 /* Short cut: if we don't know about it, try to kill it */
321 if (phase
< 2 && afs_blocksUsed_0
)
324 if (afs_pct1
> afs_tpct1
)
326 if (afs_pct2
> afs_tpct2
)
328 return 0; /* unlikely */
333 * Warn about failing to store a file.
335 * \param acode Associated error code.
336 * \param avolume Volume involved.
337 * \param aflags How to handle the output:
338 * aflags & 1: Print out on console
339 * aflags & 2: Print out on controlling tty
341 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
345 afs_StoreWarn(afs_int32 acode
, afs_int32 avolume
,
348 static char problem_fmt
[] =
349 "afs: failed to store file in volume %d (%s)\n";
350 static char problem_fmt_w_error
[] =
351 "afs: failed to store file in volume %d (error %d)\n";
352 static char netproblems
[] = "network problems";
353 static char partfull
[] = "partition full";
354 static char overquota
[] = "over quota";
356 AFS_STATCNT(afs_StoreWarn
);
362 afs_warn(problem_fmt
, avolume
, netproblems
);
364 afs_warnuser(problem_fmt
, avolume
, netproblems
);
365 } else if (acode
== ENOSPC
) {
370 afs_warn(problem_fmt
, avolume
, partfull
);
372 afs_warnuser(problem_fmt
, avolume
, partfull
);
375 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
376 * Instead ENOSPC will be sent...
378 if (acode
== EDQUOT
) {
383 afs_warn(problem_fmt
, avolume
, overquota
);
385 afs_warnuser(problem_fmt
, avolume
, overquota
);
393 afs_warn(problem_fmt_w_error
, avolume
, acode
);
395 afs_warnuser(problem_fmt_w_error
, avolume
, acode
);
400 * Try waking up truncation daemon, if it's worth it.
403 afs_MaybeWakeupTruncateDaemon(void)
405 if (!afs_CacheTooFull
&& afs_CacheIsTooFull()) {
406 afs_CacheTooFullCount
++;
407 afs_CacheTooFull
= 1;
408 if (!afs_TruncateDaemonRunning
)
409 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon
);
410 } else if (!afs_TruncateDaemonRunning
411 && afs_blocksDiscarded
> CM_MAXDISCARDEDCHUNKS
) {
412 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon
);
417 * Wait for cache drain if conditions warrant.
420 afs_MaybeWaitForCacheDrain(void)
422 if (afs_blocksUsed
- afs_blocksDiscarded
>
423 PERCENT(CM_WAITFORDRAINPCT
, afs_cacheBlocks
)) {
424 if (afs_WaitForCacheDrain
== 0)
425 afs_WaitForCacheDrainCount
++;
426 afs_WaitForCacheDrain
= 1;
427 afs_osi_Sleep(&afs_WaitForCacheDrain
);
432 * Keep statistics on run time for afs_CacheTruncateDaemon.
434 struct afs_CTD_stats CTD_stats
;
436 u_int afs_min_cache
= 0;
439 * If there are waiters for the cache to drain, wake them if
440 * the number of free or discarded cache blocks reaches the
441 * CM_CACHESIZEDDRAINEDPCT limit.
444 * This routine must be called with the afs_xdcache lock held
448 afs_WakeCacheWaitersIfDrained(void)
450 if (afs_WaitForCacheDrain
) {
451 if ((afs_blocksUsed
- afs_blocksDiscarded
) <=
452 PERCENT(CM_CACHESIZEDRAINEDPCT
, afs_cacheBlocks
)) {
453 afs_WaitForCacheDrain
= 0;
454 afs_osi_Wakeup(&afs_WaitForCacheDrain
);
460 * Keeps the cache clean and free by truncating uneeded files, when used.
465 afs_CacheTruncateDaemon(void)
467 osi_timeval32_t CTD_tmpTime
;
468 struct afs_CTD_stats
*ctd_stats
= &CTD_stats
;
472 PERCENT((100 - CM_DCACHECOUNTFREEPCT
+ CM_DCACHEEXTRAPCT
), afs_cacheFiles
);
475 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize
) & ~afs_fsfragsize
) >> 10;
477 osi_GetTime(&ctd_stats
->CTD_afterSleep
);
478 afs_TruncateDaemonRunning
= 1;
480 cb_lowat
= PERCENT((CM_DCACHESPACEFREEPCT
- CM_DCACHEEXTRAPCT
), afs_cacheBlocks
);
481 ObtainWriteLock(&afs_xdcache
, 266);
482 if (afs_CacheTooFull
|| afs_WaitForCacheDrain
) {
483 int space_needed
, slots_needed
;
484 /* if we get woken up, we should try to clean something out */
485 for (counter
= 0; counter
< 10; counter
++) {
487 afs_blocksUsed
- afs_blocksDiscarded
- cb_lowat
;
488 if (space_needed
< 0)
491 dc_hiwat
- afs_freeDCCount
- afs_discardDCCount
;
492 if (slots_needed
< 0)
494 if (slots_needed
|| space_needed
)
495 afs_GetDownD(slots_needed
, &space_needed
, 0);
496 if ((space_needed
<= 0) && (slots_needed
<= 0)) {
499 if (afs_termState
== AFSOP_STOP_TRUNCDAEMON
)
502 if (!afs_CacheIsTooFull()) {
503 afs_CacheTooFull
= 0;
504 afs_WakeCacheWaitersIfDrained();
506 } /* end of cache cleanup */
507 ReleaseWriteLock(&afs_xdcache
);
510 * This is a defensive check to try to avoid starving threads
511 * that may need the global lock so thay can help free some
512 * cache space. If this thread won't be sleeping or truncating
513 * any cache files then give up the global lock so other
514 * threads get a chance to run.
516 if ((afs_termState
!= AFSOP_STOP_TRUNCDAEMON
) && afs_CacheTooFull
517 && (!afs_blocksDiscarded
|| afs_WaitForCacheDrain
)) {
518 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
522 * This is where we free the discarded cache elements.
524 while (afs_blocksDiscarded
&& !afs_WaitForCacheDrain
525 && (afs_termState
!= AFSOP_STOP_TRUNCDAEMON
)) {
526 int code
= afs_FreeDiscardedDCache();
528 /* If we can't free any discarded dcache entries, that's okay.
529 * We're just doing this in the background; if someone needs
530 * discarded entries freed, they will try it themselves and/or
531 * signal us that the cache is too full. In any case, we'll
532 * try doing this again the next time we run through the loop.
538 /* See if we need to continue to run. Someone may have
539 * signalled us while we were executing.
541 if (!afs_WaitForCacheDrain
&& !afs_CacheTooFull
542 && (afs_termState
!= AFSOP_STOP_TRUNCDAEMON
)) {
543 /* Collect statistics on truncate daemon. */
544 ctd_stats
->CTD_nSleeps
++;
545 osi_GetTime(&ctd_stats
->CTD_beforeSleep
);
546 afs_stats_GetDiff(CTD_tmpTime
, ctd_stats
->CTD_afterSleep
,
547 ctd_stats
->CTD_beforeSleep
);
548 afs_stats_AddTo(ctd_stats
->CTD_runTime
, CTD_tmpTime
);
550 afs_TruncateDaemonRunning
= 0;
551 afs_osi_Sleep((int *)afs_CacheTruncateDaemon
);
552 afs_TruncateDaemonRunning
= 1;
554 osi_GetTime(&ctd_stats
->CTD_afterSleep
);
555 afs_stats_GetDiff(CTD_tmpTime
, ctd_stats
->CTD_beforeSleep
,
556 ctd_stats
->CTD_afterSleep
);
557 afs_stats_AddTo(ctd_stats
->CTD_sleepTime
, CTD_tmpTime
);
559 if (afs_termState
== AFSOP_STOP_TRUNCDAEMON
) {
560 afs_termState
= AFSOP_STOP_AFSDB
;
561 afs_osi_Wakeup(&afs_termState
);
569 * Make adjustment for the new size in the disk cache entry
571 * \note Major Assumptions Here:
572 * Assumes that frag size is an integral power of two, less one,
573 * and that this is a two's complement machine. I don't
574 * know of any filesystems which violate this assumption...
576 * \param adc Ptr to dcache entry.
577 * \param anewsize New size desired.
582 afs_AdjustSize(struct dcache
*adc
, afs_int32 newSize
)
586 AFS_STATCNT(afs_AdjustSize
);
588 if (newSize
> afs_OtherCSize
&& !(adc
->f
.fid
.Fid
.Vnode
& 1)) {
589 /* No non-dir cache files should be larger than the chunk size.
590 * (Directory blobs are fetched in a single chunk file, so directories
591 * can be larger.) If someone is requesting that a chunk is larger than
592 * the chunk size, something strange is happening. Log a message about
593 * it, to give a hint to subsequent strange behavior, if any occurs. */
597 afs_warn("afs: Warning: dcache %d is very large (%d > %d). This "
598 "should not happen, but trying to continue regardless. If "
599 "AFS starts hanging or behaving strangely, this might be "
601 adc
->index
, newSize
, afs_OtherCSize
);
605 adc
->dflags
|= DFEntryMod
;
606 oldSize
= afs_round_to_fsfragsize(adc
->f
.chunkBytes
);
607 adc
->f
.chunkBytes
= newSize
;
610 newSize
= afs_round_to_fsfragsize(newSize
);
611 afs_DCAdjustSize(adc
, oldSize
, newSize
);
612 if ((newSize
> oldSize
) && !AFS_IS_DISCONNECTED
) {
614 /* We're growing the file, wakeup the daemon */
615 afs_MaybeWakeupTruncateDaemon();
617 afs_blocksUsed
+= (newSize
- oldSize
);
618 afs_stats_cmperf
.cacheBlocksInUse
= afs_blocksUsed
; /* XXX */
623 * This routine is responsible for moving at least one entry (but up
624 * to some number of them) from the LRU queue to the free queue.
626 * \param anumber Number of entries that should ideally be moved.
627 * \param aneedSpace How much space we need (1K blocks);
630 * The anumber parameter is just a hint; at least one entry MUST be
631 * moved, or we'll panic. We must be called with afs_xdcache
632 * write-locked. We should try to satisfy both anumber and aneedspace,
633 * whichever is more demanding - need to do several things:
634 * 1. only grab up to anumber victims if aneedSpace <= 0, not
635 * the whole set of MAXATONCE.
636 * 2. dynamically choose MAXATONCE to reflect severity of
637 * demand: something like (*aneedSpace >> (logChunk - 9))
639 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
640 * indicates that the cache is not properly configured/tuned or
641 * something. We should be able to automatically correct that problem.
644 #define MAXATONCE 16 /* max we can obtain at once */
646 afs_GetDownD(int anumber
, int *aneedSpace
, afs_int32 buckethint
)
650 struct VenusFid
*afid
;
655 afs_uint32 victims
[MAXATONCE
];
656 struct dcache
*victimDCs
[MAXATONCE
];
657 afs_hyper_t victimTimes
[MAXATONCE
]; /* youngest (largest LRU time) first */
658 afs_uint32 victimPtr
; /* next free item in victim arrays */
659 afs_hyper_t maxVictimTime
; /* youngest (largest LRU time) victim */
660 afs_uint32 maxVictimPtr
; /* where it is */
664 AFS_STATCNT(afs_GetDownD
);
666 if (CheckLock(&afs_xdcache
) != -1)
667 osi_Panic("getdownd nolock");
668 /* decrement anumber first for all dudes in free list */
669 /* SHOULD always decrement anumber first, even if aneedSpace >0,
670 * because we should try to free space even if anumber <=0 */
671 if (!aneedSpace
|| *aneedSpace
<= 0) {
672 anumber
-= afs_freeDCCount
;
674 return; /* enough already free */
678 /* bounds check parameter */
679 if (anumber
> MAXATONCE
)
680 anumber
= MAXATONCE
; /* all we can do */
682 /* rewrite so phases include a better eligiblity for gc test*/
684 * The phase variable manages reclaims. Set to 0, the first pass,
685 * we don't reclaim active entries, or other than target bucket.
686 * Set to 1, we reclaim even active ones in target bucket.
687 * Set to 2, we reclaim any inactive one.
688 * Set to 3, we reclaim even active ones. On Solaris, we also reclaim
689 * entries whose corresponding vcache has a nonempty multiPage list, when
698 for (i
= 0; i
< afs_cacheFiles
; i
++)
699 /* turn off all flags */
700 afs_indexFlags
[i
] &= ~IFFlag
;
702 while (anumber
> 0 || (aneedSpace
&& *aneedSpace
> 0)) {
703 /* find oldest entries for reclamation */
704 maxVictimPtr
= victimPtr
= 0;
705 hzero(maxVictimTime
);
706 curbucket
= afs_DCWhichBucket(phase
, buckethint
);
707 /* select victims from access time array */
708 for (i
= 0; i
< afs_cacheFiles
; i
++) {
709 if (afs_indexFlags
[i
] & (IFDataMod
| IFFree
| IFDiscarded
)) {
710 /* skip if dirty or already free */
713 tdc
= afs_indexTable
[i
];
714 if (tdc
&& (curbucket
!= tdc
->bucket
) && (phase
< 4))
716 /* Wrong bucket; can't use it! */
719 if (tdc
&& (tdc
->refCount
!= 0)) {
720 /* Referenced; can't use it! */
723 hset(vtime
, afs_indexTimes
[i
]);
725 /* if we've already looked at this one, skip it */
726 if (afs_indexFlags
[i
] & IFFlag
)
729 if (victimPtr
< MAXATONCE
) {
730 /* if there's at least one free victim slot left */
731 victims
[victimPtr
] = i
;
732 hset(victimTimes
[victimPtr
], vtime
);
733 if (hcmp(vtime
, maxVictimTime
) > 0) {
734 hset(maxVictimTime
, vtime
);
735 maxVictimPtr
= victimPtr
;
738 } else if (hcmp(vtime
, maxVictimTime
) < 0) {
740 * We're older than youngest victim, so we replace at
743 /* find youngest (largest LRU) victim */
746 osi_Panic("getdownd local");
748 hset(victimTimes
[j
], vtime
);
749 /* recompute maxVictimTime */
750 hset(maxVictimTime
, vtime
);
751 for (j
= 0; j
< victimPtr
; j
++)
752 if (hcmp(maxVictimTime
, victimTimes
[j
]) < 0) {
753 hset(maxVictimTime
, victimTimes
[j
]);
759 /* now really reclaim the victims */
760 j
= 0; /* flag to track if we actually got any of the victims */
761 /* first, hold all the victims, since we're going to release the lock
762 * during the truncate operation.
764 for (i
= 0; i
< victimPtr
; i
++) {
765 tdc
= afs_GetValidDSlot(victims
[i
]);
766 /* We got tdc->tlock(R) here */
767 if (tdc
&& tdc
->refCount
== 1)
772 ReleaseReadLock(&tdc
->tlock
);
777 for (i
= 0; i
< victimPtr
; i
++) {
778 /* q is first elt in dcache entry */
780 /* now, since we're dropping the afs_xdcache lock below, we
781 * have to verify, before proceeding, that there are no other
782 * references to this dcache entry, even now. Note that we
783 * compare with 1, since we bumped it above when we called
784 * afs_GetValidDSlot to preserve the entry's identity.
786 if (tdc
&& tdc
->refCount
== 1) {
787 unsigned char chunkFlags
;
788 afs_size_t tchunkoffset
= 0;
790 /* xdcache is lower than the xvcache lock */
791 ReleaseWriteLock(&afs_xdcache
);
792 ObtainReadLock(&afs_xvcache
);
793 tvc
= afs_FindVCache(afid
, 0 /* no stats, no vlru */ );
794 ReleaseReadLock(&afs_xvcache
);
795 ObtainWriteLock(&afs_xdcache
, 527);
797 if (tdc
->refCount
> 1)
800 tchunkoffset
= AFS_CHUNKTOBASE(tdc
->f
.chunk
);
801 chunkFlags
= afs_indexFlags
[tdc
->index
];
802 if (((phase
& 1) == 0) && osi_Active(tvc
))
804 if (((phase
& 1) == 1) && osi_Active(tvc
)
805 && (tvc
->f
.states
& CDCLock
)
806 && (chunkFlags
& IFAnyPages
))
808 if (chunkFlags
& IFDataMod
)
810 afs_Trace4(afs_iclSetp
, CM_TRACE_GETDOWND
,
811 ICL_TYPE_POINTER
, tvc
, ICL_TYPE_INT32
, skip
,
812 ICL_TYPE_INT32
, tdc
->index
, ICL_TYPE_OFFSET
,
813 ICL_HANDLE_OFFSET(tchunkoffset
));
815 #if defined(AFS_SUN5_ENV)
817 * Now we try to invalidate pages. We do this only for
818 * Solaris. For other platforms, it's OK to recycle a
819 * dcache entry out from under a page, because the strategy
820 * function can call afs_GetDCache().
822 if (!skip
&& (chunkFlags
& IFAnyPages
)) {
825 ReleaseWriteLock(&afs_xdcache
);
826 ObtainWriteLock(&tvc
->vlock
, 543);
827 if (!QEmpty(&tvc
->multiPage
)) {
828 if (phase
< 3 || osi_VM_MultiPageConflict(tvc
, tdc
)) {
833 /* block locking pages */
834 tvc
->vstates
|= VPageCleaning
;
835 /* block getting new pages */
837 ReleaseWriteLock(&tvc
->vlock
);
838 /* One last recheck */
839 ObtainWriteLock(&afs_xdcache
, 333);
840 chunkFlags
= afs_indexFlags
[tdc
->index
];
841 if (tdc
->refCount
> 1 || (chunkFlags
& IFDataMod
)
842 || (osi_Active(tvc
) && (tvc
->f
.states
& CDCLock
)
843 && (chunkFlags
& IFAnyPages
))) {
845 ReleaseWriteLock(&afs_xdcache
);
848 ReleaseWriteLock(&afs_xdcache
);
850 code
= osi_VM_GetDownD(tvc
, tdc
);
852 ObtainWriteLock(&afs_xdcache
, 269);
853 /* we actually removed all pages, clean and dirty */
855 afs_indexFlags
[tdc
->index
] &=
856 ~(IFDirtyPages
| IFAnyPages
);
859 ReleaseWriteLock(&afs_xdcache
);
861 ObtainWriteLock(&tvc
->vlock
, 544);
862 if (--tvc
->activeV
== 0
863 && (tvc
->vstates
& VRevokeWait
)) {
864 tvc
->vstates
&= ~VRevokeWait
;
865 afs_osi_Wakeup((char *)&tvc
->vstates
);
868 if (tvc
->vstates
& VPageCleaning
) {
869 tvc
->vstates
&= ~VPageCleaning
;
870 afs_osi_Wakeup((char *)&tvc
->vstates
);
873 ReleaseWriteLock(&tvc
->vlock
);
875 #endif /* AFS_SUN5_ENV */
877 ReleaseWriteLock(&afs_xdcache
);
880 afs_PutVCache(tvc
); /*XXX was AFS_FAST_RELE?*/
881 ObtainWriteLock(&afs_xdcache
, 528);
882 if (afs_indexFlags
[tdc
->index
] &
883 (IFDataMod
| IFDirtyPages
| IFAnyPages
))
885 if (tdc
->refCount
> 1)
888 #if defined(AFS_SUN5_ENV)
890 /* no vnode, so IFDirtyPages is spurious (we don't
891 * sweep dcaches on vnode recycling, so we can have
892 * DIRTYPAGES set even when all pages are gone). Just
894 * Hold vcache lock to prevent vnode from being
895 * created while we're clearing IFDirtyPages.
897 afs_indexFlags
[tdc
->index
] &=
898 ~(IFDirtyPages
| IFAnyPages
);
902 /* skip this guy and mark him as recently used */
903 afs_indexFlags
[tdc
->index
] |= IFFlag
;
904 afs_Trace4(afs_iclSetp
, CM_TRACE_GETDOWND
,
905 ICL_TYPE_POINTER
, tvc
, ICL_TYPE_INT32
, 2,
906 ICL_TYPE_INT32
, tdc
->index
, ICL_TYPE_OFFSET
,
907 ICL_HANDLE_OFFSET(tchunkoffset
));
909 /* flush this dude from the data cache and reclaim;
910 * first, make sure no one will care that we damage
911 * it, by removing it from all hash tables. Then,
912 * melt it down for parts. Note that any concurrent
913 * (new possibility!) calls to GetDownD won't touch
914 * this guy because his reference count is > 0. */
915 afs_Trace4(afs_iclSetp
, CM_TRACE_GETDOWND
,
916 ICL_TYPE_POINTER
, tvc
, ICL_TYPE_INT32
, 3,
917 ICL_TYPE_INT32
, tdc
->index
, ICL_TYPE_OFFSET
,
918 ICL_HANDLE_OFFSET(tchunkoffset
));
919 AFS_STATCNT(afs_gget
);
920 afs_HashOutDCache(tdc
, 1);
921 if (tdc
->f
.chunkBytes
!= 0) {
925 (tdc
->f
.chunkBytes
+ afs_fsfragsize
) >> 10;
930 afs_DiscardDCache(tdc
);
935 j
= 1; /* we reclaimed at least one victim */
940 } /* end of for victims loop */
943 /* Phase is 0 and no one was found, so try phase 1 (ignore
944 * osi_Active flag) */
947 for (i
= 0; i
< afs_cacheFiles
; i
++)
948 /* turn off all flags */
949 afs_indexFlags
[i
] &= ~IFFlag
;
952 /* found no one in phases 0-5, we're hosed */
956 } /* big while loop */
964 * Remove adc from any hash tables that would allow it to be located
965 * again by afs_FindDCache or afs_GetDCache.
967 * \param adc Pointer to dcache entry to remove from hash tables.
969 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
973 afs_HashOutDCache(struct dcache
*adc
, int zap
)
977 AFS_STATCNT(afs_glink
);
979 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
981 /* if this guy is in the hash table, pull him out */
982 if (adc
->f
.fid
.Fid
.Volume
!= 0) {
983 /* remove entry from first hash chains */
984 i
= DCHash(&adc
->f
.fid
, adc
->f
.chunk
);
985 us
= afs_dchashTbl
[i
];
986 if (us
== adc
->index
) {
987 /* first dude in the list */
988 afs_dchashTbl
[i
] = afs_dcnextTbl
[adc
->index
];
990 /* somewhere on the chain */
991 while (us
!= NULLIDX
) {
992 if (afs_dcnextTbl
[us
] == adc
->index
) {
993 /* found item pointing at the one to delete */
994 afs_dcnextTbl
[us
] = afs_dcnextTbl
[adc
->index
];
997 us
= afs_dcnextTbl
[us
];
1000 osi_Panic("dcache hc");
1002 /* remove entry from *other* hash chain */
1003 i
= DVHash(&adc
->f
.fid
);
1004 us
= afs_dvhashTbl
[i
];
1005 if (us
== adc
->index
) {
1006 /* first dude in the list */
1007 afs_dvhashTbl
[i
] = afs_dvnextTbl
[adc
->index
];
1009 /* somewhere on the chain */
1010 while (us
!= NULLIDX
) {
1011 if (afs_dvnextTbl
[us
] == adc
->index
) {
1012 /* found item pointing at the one to delete */
1013 afs_dvnextTbl
[us
] = afs_dvnextTbl
[adc
->index
];
1016 us
= afs_dvnextTbl
[us
];
1019 osi_Panic("dcache hv");
1024 /* prevent entry from being found on a reboot (it is already out of
1025 * the hash table, but after a crash, we just look at fid fields of
1026 * stable (old) entries).
1028 adc
->f
.fid
.Fid
.Volume
= 0; /* invalid */
1030 /* mark entry as modified */
1031 adc
->dflags
|= DFEntryMod
;
1036 } /*afs_HashOutDCache */
1039 * Flush the given dcache entry, pulling it from hash chains
1040 * and truncating the associated cache file.
1042 * \param adc Ptr to dcache entry to flush.
1044 * \note Environment:
1045 * This routine must be called with the afs_xdcache lock held
1049 afs_FlushDCache(struct dcache
*adc
)
1051 AFS_STATCNT(afs_FlushDCache
);
1053 * Bump the number of cache files flushed.
1055 afs_stats_cmperf
.cacheFlushes
++;
1057 /* remove from all hash tables */
1058 afs_HashOutDCache(adc
, 1);
1060 /* Free its space; special case null operation, since truncate operation
1061 * in UFS is slow even in this case, and this allows us to pre-truncate
1062 * these files at more convenient times with fewer locks set
1063 * (see afs_GetDownD).
1065 if (adc
->f
.chunkBytes
!= 0) {
1066 afs_DiscardDCache(adc
);
1067 afs_MaybeWakeupTruncateDaemon();
1069 afs_FreeDCache(adc
);
1071 } /*afs_FlushDCache */
1075 * Put a dcache entry on the free dcache entry list.
1077 * \param adc dcache entry to free.
1079 * \note Environment: called with afs_xdcache lock write-locked.
1082 afs_FreeDCache(struct dcache
*adc
)
1084 /* Thread on free list, update free list count and mark entry as
1085 * freed in its indexFlags element. Also, ensure DCache entry gets
1086 * written out (set DFEntryMod).
1089 afs_dvnextTbl
[adc
->index
] = afs_freeDCList
;
1090 afs_freeDCList
= adc
->index
;
1092 afs_indexFlags
[adc
->index
] |= IFFree
;
1093 adc
->dflags
|= DFEntryMod
;
1095 afs_WakeCacheWaitersIfDrained();
1096 } /* afs_FreeDCache */
1099 * Discard the cache element by moving it to the discardDCList.
1100 * This puts the cache element into a quasi-freed state, where
1101 * the space may be reused, but the file has not been truncated.
1103 * \note Major Assumptions Here:
1104 * Assumes that frag size is an integral power of two, less one,
1105 * and that this is a two's complement machine. I don't
1106 * know of any filesystems which violate this assumption...
1108 * \param adr Ptr to dcache entry.
1110 * \note Environment:
1111 * Must be called with afs_xdcache write-locked.
1115 afs_DiscardDCache(struct dcache
*adc
)
1119 AFS_STATCNT(afs_DiscardDCache
);
1121 osi_Assert(adc
->refCount
== 1);
1123 size
= afs_round_to_fsfragsize(adc
->f
.chunkBytes
);
1124 afs_blocksDiscarded
+= size
;
1125 afs_stats_cmperf
.cacheBlocksDiscarded
= afs_blocksDiscarded
;
1127 afs_dvnextTbl
[adc
->index
] = afs_discardDCList
;
1128 afs_discardDCList
= adc
->index
;
1129 afs_discardDCCount
++;
1131 adc
->f
.fid
.Fid
.Volume
= 0;
1132 adc
->dflags
|= DFEntryMod
;
1133 afs_indexFlags
[adc
->index
] |= IFDiscarded
;
1135 afs_WakeCacheWaitersIfDrained();
1136 } /*afs_DiscardDCache */
1139 * Get a dcache entry from the discard or free list
1141 * @param[out] adc On success, a dcache from the given list. Otherwise, NULL.
1142 * @param[in] indexp A pointer to the head of the dcache free list or discard
1143 * list (afs_freeDCList, or afs_discardDCList)
1145 * @return 0 on success. If there are no dcache slots available, return ENOSPC.
1146 * If we encountered an error in disk i/o while trying to find a
1147 * dcache, return EIO.
1149 * @pre afs_xdcache is write-locked
1152 afs_GetDSlotFromList(struct dcache
**adc
, afs_int32
*indexp
)
1158 if (*indexp
== NULLIDX
) {
1162 tdc
= afs_GetUnusedDSlot(*indexp
);
1167 osi_Assert(tdc
->refCount
== 1);
1168 ReleaseReadLock(&tdc
->tlock
);
1169 *indexp
= afs_dvnextTbl
[tdc
->index
];
1170 afs_dvnextTbl
[tdc
->index
] = NULLIDX
;
1177 * Free the next element on the list of discarded cache elements.
1179 * Returns -1 if we encountered an error preventing us from freeing a
1180 * discarded dcache, or 0 on success.
1183 afs_FreeDiscardedDCache(void)
1186 struct osi_file
*tfile
;
1189 AFS_STATCNT(afs_FreeDiscardedDCache
);
1191 ObtainWriteLock(&afs_xdcache
, 510);
1192 if (!afs_blocksDiscarded
) {
1193 ReleaseWriteLock(&afs_xdcache
);
1198 * Get an entry from the list of discarded cache elements
1200 (void)afs_GetDSlotFromList(&tdc
, &afs_discardDCList
);
1202 ReleaseWriteLock(&afs_xdcache
);
1206 afs_discardDCCount
--;
1207 size
= afs_round_to_fsfragsize(tdc
->f
.chunkBytes
);
1208 afs_blocksDiscarded
-= size
;
1209 afs_stats_cmperf
.cacheBlocksDiscarded
= afs_blocksDiscarded
;
1210 /* We can lock because we just took it off the free list */
1211 ObtainWriteLock(&tdc
->lock
, 626);
1212 ReleaseWriteLock(&afs_xdcache
);
1215 * Truncate the element to reclaim its space
1217 tfile
= afs_CFileOpen(&tdc
->f
.inode
);
1219 afs_CFileTruncate(tfile
, 0);
1220 afs_CFileClose(tfile
);
1221 afs_AdjustSize(tdc
, 0);
1222 afs_DCMoveBucket(tdc
, 0, 0);
1225 * Free the element we just truncated
1227 ObtainWriteLock(&afs_xdcache
, 511);
1228 afs_indexFlags
[tdc
->index
] &= ~IFDiscarded
;
1229 afs_FreeDCache(tdc
);
1230 tdc
->f
.states
&= ~(DRO
|DBackup
|DRW
);
1231 ReleaseWriteLock(&tdc
->lock
);
1233 ReleaseWriteLock(&afs_xdcache
);
1239 * Free as many entries from the list of discarded cache elements
1240 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1245 afs_MaybeFreeDiscardedDCache(void)
1248 AFS_STATCNT(afs_MaybeFreeDiscardedDCache
);
1250 while (afs_blocksDiscarded
1251 && (afs_blocksUsed
>
1252 PERCENT(CM_WAITFORDRAINPCT
, afs_cacheBlocks
))) {
1253 int code
= afs_FreeDiscardedDCache();
1255 /* Callers depend on us to get the afs_blocksDiscarded count down.
1256 * If we cannot do that, the callers can spin by calling us over
1257 * and over. Panic for now until we can figure out something
1259 osi_Panic("Error freeing discarded dcache");
1266 * Try to free up a certain number of disk slots.
1268 * \param anumber Targeted number of disk slots to free up.
1270 * \note Environment:
1271 * Must be called with afs_xdcache write-locked.
1275 afs_GetDownDSlot(int anumber
)
1277 struct afs_q
*tq
, *nq
;
1282 AFS_STATCNT(afs_GetDownDSlot
);
1283 if (cacheDiskType
== AFS_FCACHE_TYPE_MEM
)
1284 osi_Panic("diskless getdowndslot");
1286 if (CheckLock(&afs_xdcache
) != -1)
1287 osi_Panic("getdowndslot nolock");
1289 /* decrement anumber first for all dudes in free list */
1290 for (tdc
= afs_freeDSList
; tdc
; tdc
= (struct dcache
*)tdc
->lruq
.next
)
1293 return; /* enough already free */
1295 for (cnt
= 0, tq
= afs_DLRU
.prev
; tq
!= &afs_DLRU
&& anumber
> 0;
1297 tdc
= (struct dcache
*)tq
; /* q is first elt in dcache entry */
1298 nq
= QPrev(tq
); /* in case we remove it */
1299 if (tdc
->refCount
== 0) {
1300 if ((ix
= tdc
->index
) == NULLIDX
)
1301 osi_Panic("getdowndslot");
1303 /* write-through if modified */
1304 if (tdc
->dflags
& DFEntryMod
) {
1305 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1307 * ask proxy to do this for us - we don't have the stack space
1309 while (tdc
->dflags
& DFEntryMod
) {
1312 s
= SPLOCK(afs_sgibklock
);
1313 if (afs_sgibklist
== NULL
) {
1314 /* if slot is free, grab it. */
1315 afs_sgibklist
= tdc
;
1316 SV_SIGNAL(&afs_sgibksync
);
1318 /* wait for daemon to (start, then) finish. */
1319 SP_WAIT(afs_sgibklock
, s
, &afs_sgibkwait
, PINOD
);
1325 code
= afs_WriteDCache(tdc
, 1);
1328 * We couldn't flush it at this time; return early because
1329 * if afs_WriteDCache() failed once it is likely to
1330 * continue failing for subsequent dcaches.
1334 tdc
->dflags
&= ~DFEntryMod
;
1338 /* pull the entry out of the lruq and put it on the free list */
1339 QRemove(&tdc
->lruq
);
1340 afs_indexTable
[ix
] = NULL
;
1341 afs_indexFlags
[ix
] &= ~IFEverUsed
;
1342 tdc
->index
= NULLIDX
;
1343 tdc
->lruq
.next
= (struct afs_q
*)afs_freeDSList
;
1344 afs_freeDSList
= tdc
;
1348 } /*afs_GetDownDSlot */
1355 * Increment the reference count on a disk cache entry,
1356 * which already has a non-zero refcount. In order to
1357 * increment the refcount of a zero-reference entry, you
1358 * have to hold afs_xdcache.
1361 * adc : Pointer to the dcache entry to increment.
1364 * Nothing interesting.
1367 afs_RefDCache(struct dcache
*adc
)
1369 ObtainWriteLock(&adc
->tlock
, 627);
1370 if (adc
->refCount
< 0)
1371 osi_Panic("RefDCache: negative refcount");
1373 ReleaseWriteLock(&adc
->tlock
);
1382 * Decrement the reference count on a disk cache entry.
1385 * ad : Ptr to the dcache entry to decrement.
1388 * Nothing interesting.
1391 afs_PutDCache(struct dcache
*adc
)
1393 AFS_STATCNT(afs_PutDCache
);
1394 ObtainWriteLock(&adc
->tlock
, 276);
1395 if (adc
->refCount
<= 0)
1396 osi_Panic("putdcache");
1398 ReleaseWriteLock(&adc
->tlock
);
1407 * Try to discard all data associated with this file from the
1411 * avc : Pointer to the cache info for the file.
1414 * Both pvnLock and lock are write held.
1417 afs_TryToSmush(struct vcache
*avc
, afs_ucred_t
*acred
, int sync
)
1422 AFS_STATCNT(afs_TryToSmush
);
1423 afs_Trace2(afs_iclSetp
, CM_TRACE_TRYTOSMUSH
, ICL_TYPE_POINTER
, avc
,
1424 ICL_TYPE_OFFSET
, ICL_HANDLE_OFFSET(avc
->f
.m
.Length
));
1425 sync
= 1; /* XX Temp testing XX */
1427 #if defined(AFS_SUN5_ENV)
1428 ObtainWriteLock(&avc
->vlock
, 573);
1429 avc
->activeV
++; /* block new getpages */
1430 ReleaseWriteLock(&avc
->vlock
);
1433 /* Flush VM pages */
1434 osi_VM_TryToSmush(avc
, acred
, sync
);
1437 * Get the hash chain containing all dce's for this fid
1439 i
= DVHash(&avc
->f
.fid
);
1440 ObtainWriteLock(&afs_xdcache
, 277);
1441 for (index
= afs_dvhashTbl
[i
]; index
!= NULLIDX
; index
= i
) {
1442 i
= afs_dvnextTbl
[index
]; /* next pointer this hash table */
1443 if (afs_indexUnique
[index
] == avc
->f
.fid
.Fid
.Unique
) {
1444 int releaseTlock
= 1;
1445 tdc
= afs_GetValidDSlot(index
);
1447 /* afs_TryToSmush is best-effort; we may not actually discard
1448 * everything, so failure to discard dcaches due to an i/o
1452 if (!FidCmp(&tdc
->f
.fid
, &avc
->f
.fid
)) {
1454 if ((afs_indexFlags
[index
] & IFDataMod
) == 0
1455 && tdc
->refCount
== 1) {
1456 ReleaseReadLock(&tdc
->tlock
);
1458 afs_FlushDCache(tdc
);
1461 afs_indexTable
[index
] = 0;
1464 ReleaseReadLock(&tdc
->tlock
);
1468 #if defined(AFS_SUN5_ENV)
1469 ObtainWriteLock(&avc
->vlock
, 545);
1470 if (--avc
->activeV
== 0 && (avc
->vstates
& VRevokeWait
)) {
1471 avc
->vstates
&= ~VRevokeWait
;
1472 afs_osi_Wakeup((char *)&avc
->vstates
);
1474 ReleaseWriteLock(&avc
->vlock
);
1476 ReleaseWriteLock(&afs_xdcache
);
1478 * It's treated like a callback so that when we do lookups we'll
1479 * invalidate the unique bit if any
1480 * trytoSmush occured during the lookup call
1486 * afs_DCacheMissingChunks
1489 * Given the cached info for a file, return the number of chunks that
1490 * are not available from the dcache.
1493 * avc: Pointer to the (held) vcache entry to look in.
1496 * The number of chunks which are not currently cached.
1499 * The vcache entry is held upon entry.
1503 afs_DCacheMissingChunks(struct vcache
*avc
)
1506 afs_size_t totalLength
= 0;
1507 afs_uint32 totalChunks
= 0;
1510 totalLength
= avc
->f
.m
.Length
;
1511 if (avc
->f
.truncPos
< totalLength
)
1512 totalLength
= avc
->f
.truncPos
;
1514 /* Length is 0, no chunk missing. */
1515 if (totalLength
== 0)
1518 /* If totalLength is a multiple of chunksize, the last byte appears
1519 * as being part of the next chunk, which does not exist.
1520 * Decrementing totalLength by one fixes that.
1523 totalChunks
= (AFS_CHUNK(totalLength
) + 1);
1525 /* If we're a directory, we only ever have one chunk, regardless of
1526 * the size of the dir.
1528 if (avc
->f
.fid
.Fid
.Vnode
& 1 || vType(avc
) == VDIR
)
1532 printf("Should have %d chunks for %u bytes\n",
1533 totalChunks, (totalLength + 1));
1535 i
= DVHash(&avc
->f
.fid
);
1536 ObtainWriteLock(&afs_xdcache
, 1001);
1537 for (index
= afs_dvhashTbl
[i
]; index
!= NULLIDX
; index
= i
) {
1538 i
= afs_dvnextTbl
[index
];
1539 if (afs_indexUnique
[index
] == avc
->f
.fid
.Fid
.Unique
) {
1540 tdc
= afs_GetValidDSlot(index
);
1544 if (!FidCmp(&tdc
->f
.fid
, &avc
->f
.fid
)) {
1547 ReleaseReadLock(&tdc
->tlock
);
1551 ReleaseWriteLock(&afs_xdcache
);
1553 /*printf("Missing %d chunks\n", totalChunks);*/
1555 return (totalChunks
);
1562 * Given the cached info for a file and a byte offset into the
1563 * file, make sure the dcache entry for that file and containing
1564 * the given byte is available, returning it to our caller.
1567 * avc : Pointer to the (held) vcache entry to look in.
1568 * abyte : Which byte we want to get to.
1571 * Pointer to the dcache entry covering the file & desired byte,
1572 * or NULL if not found.
1575 * The vcache entry is held upon entry.
1579 afs_FindDCache(struct vcache
*avc
, afs_size_t abyte
)
1583 struct dcache
*tdc
= NULL
;
1585 AFS_STATCNT(afs_FindDCache
);
1586 chunk
= AFS_CHUNK(abyte
);
1589 * Hash on the [fid, chunk] and get the corresponding dcache index
1590 * after write-locking the dcache.
1592 i
= DCHash(&avc
->f
.fid
, chunk
);
1593 ObtainWriteLock(&afs_xdcache
, 278);
1594 for (index
= afs_dchashTbl
[i
]; index
!= NULLIDX
; index
= afs_dcnextTbl
[index
]) {
1595 if (afs_indexUnique
[index
] == avc
->f
.fid
.Fid
.Unique
) {
1596 tdc
= afs_GetValidDSlot(index
);
1598 /* afs_FindDCache is best-effort; we may not find the given
1599 * file/offset, so if we cannot find the given dcache due to
1600 * i/o errors, that is okay. */
1604 ReleaseReadLock(&tdc
->tlock
);
1605 if (!FidCmp(&tdc
->f
.fid
, &avc
->f
.fid
) && chunk
== tdc
->f
.chunk
) {
1606 break; /* leaving refCount high for caller */
1611 if (index
!= NULLIDX
) {
1612 hset(afs_indexTimes
[tdc
->index
], afs_indexCounter
);
1613 hadd32(afs_indexCounter
, 1);
1614 ReleaseWriteLock(&afs_xdcache
);
1617 ReleaseWriteLock(&afs_xdcache
);
1619 } /*afs_FindDCache */
1621 /* only call these from afs_AllocDCache() */
1623 afs_AllocFreeDSlot(struct dcache
**adc
)
1628 code
= afs_GetDSlotFromList(&tdc
, &afs_freeDCList
);
1632 afs_indexFlags
[tdc
->index
] &= ~IFFree
;
1633 ObtainWriteLock(&tdc
->lock
, 604);
1640 afs_AllocDiscardDSlot(struct dcache
**adc
, afs_int32 lock
)
1644 afs_uint32 size
= 0;
1645 struct osi_file
*file
;
1647 code
= afs_GetDSlotFromList(&tdc
, &afs_discardDCList
);
1651 afs_indexFlags
[tdc
->index
] &= ~IFDiscarded
;
1652 ObtainWriteLock(&tdc
->lock
, 605);
1653 afs_discardDCCount
--;
1654 size
= afs_round_to_fsfragsize(tdc
->f
.chunkBytes
);
1655 tdc
->f
.states
&= ~(DRO
|DBackup
|DRW
);
1656 afs_DCMoveBucket(tdc
, size
, 0);
1657 afs_blocksDiscarded
-= size
;
1658 afs_stats_cmperf
.cacheBlocksDiscarded
= afs_blocksDiscarded
;
1660 /* Truncate the chunk so zeroes get filled properly */
1661 file
= afs_CFileOpen(&tdc
->f
.inode
);
1663 afs_CFileTruncate(file
, 0);
1664 afs_CFileClose(file
);
1665 afs_AdjustSize(tdc
, 0);
1673 * Get a fresh dcache from the free or discarded list.
1675 * \param adc Set to the new dcache on success, and NULL on error.
1676 * \param avc Who's dcache is this going to be?
1677 * \param chunk The position where it will be placed in.
1678 * \param lock How are locks held.
1679 * \param ashFid If this dcache going to be used for a shadow dir,
1682 * \note Required locks:
1684 * - avc (R if (lock & 1) set and W otherwise)
1685 * \note It write locks the new dcache. The caller must unlock it.
1687 * \return If we're out of dslots, ENOSPC. If we encountered disk errors, EIO.
1688 * On success, return 0.
1691 afs_AllocDCache(struct dcache
**adc
, struct vcache
*avc
, afs_int32 chunk
,
1692 afs_int32 lock
, struct VenusFid
*ashFid
)
1695 struct dcache
*tdc
= NULL
;
1699 /* if (lock & 2), prefer 'free' dcaches; otherwise, prefer 'discard'
1700 * dcaches. In either case, try both if our first choice doesn't work due
1703 code
= afs_AllocFreeDSlot(&tdc
);
1704 if (code
== ENOSPC
) {
1705 code
= afs_AllocDiscardDSlot(&tdc
, lock
);
1708 code
= afs_AllocDiscardDSlot(&tdc
, lock
);
1709 if (code
== ENOSPC
) {
1710 code
= afs_AllocFreeDSlot(&tdc
);
1719 * avc->lock(R) if setLocks
1720 * avc->lock(W) if !setLocks
1726 * Fill in the newly-allocated dcache record.
1728 afs_indexFlags
[tdc
->index
] &= ~(IFDirtyPages
| IFAnyPages
);
1730 /* Use shadow fid if provided. */
1731 tdc
->f
.fid
= *ashFid
;
1733 /* Use normal vcache's fid otherwise. */
1734 tdc
->f
.fid
= avc
->f
.fid
;
1735 if (avc
->f
.states
& CRO
)
1736 tdc
->f
.states
= DRO
;
1737 else if (avc
->f
.states
& CBackup
)
1738 tdc
->f
.states
= DBackup
;
1740 tdc
->f
.states
= DRW
;
1741 afs_DCMoveBucket(tdc
, 0, afs_DCGetBucket(avc
));
1742 afs_indexUnique
[tdc
->index
] = tdc
->f
.fid
.Fid
.Unique
;
1744 hones(tdc
->f
.versionNo
); /* invalid value */
1745 tdc
->f
.chunk
= chunk
;
1746 tdc
->validPos
= AFS_CHUNKTOBASE(chunk
);
1748 if (tdc
->lruq
.prev
== &tdc
->lruq
)
1749 osi_Panic("lruq 1");
1756 IsDCacheSizeOK(struct dcache
*adc
, struct vcache
*avc
, afs_int32 chunk_bytes
,
1757 afs_size_t file_length
, afs_uint32 versionNo
, int from_net
)
1759 afs_size_t expected_bytes
;
1760 afs_size_t chunk_start
= AFS_CHUNKTOBASE(adc
->f
.chunk
);
1762 if (vType(avc
) == VDIR
) {
1764 * Directory blobs may be constructed locally (see afs_LocalHero), and
1765 * the size of the blob may differ slightly compared to what's on the
1766 * fileserver. So, skip size checks for directories.
1771 if ((avc
->f
.states
& CDirty
)) {
1773 * Our vcache may have writes that are local to our cache, but not yet
1774 * written to the fileserver. In such a situation, we may have dcaches
1775 * for that file that are "short". For example:
1777 * Say we have a file that is 0 bytes long. A process opens that file,
1778 * and writes some data to offset 5M (keeping the file open). Another
1779 * process comes along and reads data from offset 1M. We'll try to
1780 * fetch data at offset 1M, and the fileserver will respond with 0
1781 * bytes, since our locally-written data hasn't been written to the
1782 * fileserver yet (on the fileserver, the file is still 0-bytes long).
1783 * So our dcache at offset 1M will have 0 bytes.
1785 * So if CDirty is set, don't do any size/length checks at all, since
1786 * we have no idea if the avc length is valid.
1791 if (!from_net
&& (adc
->f
.states
& DRW
)) {
1793 * The dcache data we're looking at is from our local cache (not from a
1794 * fileserver), and it's for data in an RW volume. For cached RW data,
1795 * there are some edge cases that can cause the below length checks to
1796 * trigger false positives.
1798 * For example: if the local client writes 4 bytes to a new file at
1799 * offset 0, and then 4 bytes at offset 0x400000, the file will be
1800 * 0x400004 bytes long, but the first dcache chunk will only contain 4
1801 * bytes. If such a file is fetched from a fileserver, the first chunk
1802 * will have a full chunk of data (most of it zeroes), but on the
1803 * client that did the write, the sparse data will not appear in the
1806 * Such false positives should only be possible with RW data, since
1807 * non-RW data is never generated locally. So to avoid the false
1808 * positives, assume the dcache length is OK for RW data if the dcache
1809 * came from our local cache (and not directly from a fileserver).
1814 if (file_length
< chunk_start
) {
1818 expected_bytes
= file_length
- chunk_start
;
1820 if (vType(avc
) != VDIR
&& expected_bytes
> AFS_CHUNKTOSIZE(adc
->f
.chunk
)) {
1821 /* A non-dir chunk cannot have more bytes than the chunksize. */
1822 expected_bytes
= AFS_CHUNKTOSIZE(adc
->f
.chunk
);
1826 if (chunk_bytes
!= expected_bytes
) {
1827 static const afs_uint32 one_hour
= 60 * 60;
1828 static afs_uint32 last_warn
;
1829 afs_uint32 now
= osi_Time();
1831 if (now
< last_warn
) {
1832 /* clock went backwards */
1836 if (now
- last_warn
> one_hour
) {
1837 unsigned int mtime
= adc
->f
.modTime
;
1843 * The dcache we're looking at didn't come from the cache, but is
1844 * being populated from the net. Don't print out its mtime in that
1845 * case; that would be misleading since that's the mtime from the
1846 * last time this dcache slot was written to.
1851 afs_warn("afs: Detected corrupt dcache for file %d.%u.%u.%u: chunk %d "
1852 "(offset %lu) has %d bytes, but it should have %lu bytes\n",
1854 adc
->f
.fid
.Fid
.Volume
,
1855 adc
->f
.fid
.Fid
.Vnode
,
1856 adc
->f
.fid
.Fid
.Unique
,
1858 (unsigned long)chunk_start
,
1860 (unsigned long)expected_bytes
);
1861 afs_warn("afs: (dcache %p, file length %lu, DV %u, dcache mtime %u, "
1862 "index %d, dflags 0x%x, mflags 0x%x, states 0x%x, vcache "
1865 (unsigned long)file_length
,
1869 (unsigned)adc
->dflags
,
1870 (unsigned)adc
->mflags
,
1871 (unsigned)adc
->f
.states
,
1873 afs_warn("afs: Ignoring the dcache for now, but this may indicate "
1874 "corruption in the AFS cache, or a bug.\n");
1882 * Check if a dcache is "fresh". That is, if the dcache's DV matches the DV of
1883 * the vcache for that file, and the dcache looks "sane" (its length makes
1884 * sense, when considering the length of the given avc).
1886 * \param adc The dcache to check
1887 * \param avc The vcache for adc
1889 * \return 1 if the dcache is "fresh". 0 otherwise.
1892 afs_IsDCacheFresh(struct dcache
*adc
, struct vcache
*avc
)
1894 if (!hsame(adc
->f
.versionNo
, avc
->f
.m
.DataVersion
)) {
1899 * If we've reached here, the DV in adc matches the DV of our avc. Check if
1900 * the number of bytes in adc agrees with the avc file length, as a sanity
1901 * check. If they don't match, we'll pretend the DVs don't match, so the
1902 * bad dcache data will not be used, and we'll probably re-fetch the chunk
1903 * data, replacing the bad chunk.
1906 if (!IsDCacheSizeOK(adc
, avc
, adc
->f
.chunkBytes
, avc
->f
.m
.Length
,
1907 hgetlo(adc
->f
.versionNo
), 0)) {
1918 * This function is called to obtain a reference to data stored in
1919 * the disk cache, locating a chunk of data containing the desired
1920 * byte and returning a reference to the disk cache entry, with its
1921 * reference count incremented.
1925 * avc : Ptr to a vcache entry (unlocked)
1926 * abyte : Byte position in the file desired
1927 * areq : Request structure identifying the requesting user.
1928 * aflags : Settings as follows:
1930 * 2 : Return after creating entry.
1931 * 4 : called from afs_vnop_write.c
1932 * *alen contains length of data to be written.
1934 * aoffset : Set to the offset within the chunk where the resident
1936 * alen : Set to the number of bytes of data after the desired
1937 * byte (including the byte itself) which can be read
1941 * The vcache entry pointed to by avc is unlocked upon entry.
1945 * Update the vnode-to-dcache hint if we can get the vnode lock
1946 * right away. Assumes dcache entry is at least read-locked.
1949 updateV2DC(int lockVc
, struct vcache
*v
, struct dcache
*d
, int src
)
1951 if (!lockVc
|| 0 == NBObtainWriteLock(&v
->lock
, src
)) {
1952 if (afs_IsDCacheFresh(d
, v
) && v
->callback
)
1955 ReleaseWriteLock(&v
->lock
);
1959 /* avc - Write-locked unless aflags & 1 */
1961 afs_GetDCache(struct vcache
*avc
, afs_size_t abyte
,
1962 struct vrequest
*areq
, afs_size_t
* aoffset
,
1963 afs_size_t
* alen
, int aflags
)
1965 afs_int32 i
, code
, shortcut
;
1966 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1967 afs_int32 adjustsize
= 0;
1973 afs_size_t Position
= 0;
1974 afs_int32 size
, tlen
; /* size of segment to transfer */
1975 struct afs_FetchOutput
*tsmall
= 0;
1977 struct osi_file
*file
;
1978 struct afs_conn
*tc
;
1980 struct server
*newCallback
= NULL
;
1981 char setNewCallback
;
1982 char setVcacheStatus
;
1983 char doVcacheUpdate
;
1985 int doAdjustSize
= 0;
1986 int doReallyAdjustSize
= 0;
1987 int overWriteWholeChunk
= 0;
1988 struct rx_connection
*rxconn
;
1991 struct afs_stats_AccessInfo
*accP
; /*Ptr to access record in stats */
1992 int fromReplica
; /*Are we reading from a replica? */
1993 int numFetchLoops
; /*# times around the fetch/analyze loop */
1994 #endif /* AFS_NOSTATS */
1996 AFS_STATCNT(afs_GetDCache
);
2000 setLocks
= aflags
& 1;
2003 * Determine the chunk number and offset within the chunk corresponding
2004 * to the desired byte.
2006 if (avc
->f
.fid
.Fid
.Vnode
& 1) { /* if (vType(avc) == VDIR) */
2009 chunk
= AFS_CHUNK(abyte
);
2012 /* come back to here if we waited for the cache to drain. */
2015 setNewCallback
= setVcacheStatus
= 0;
2019 ObtainWriteLock(&avc
->lock
, 616);
2021 ObtainReadLock(&avc
->lock
);
2026 * avc->lock(R) if setLocks && !slowPass
2027 * avc->lock(W) if !setLocks || slowPass
2032 /* check hints first! (might could use bcmp or some such...) */
2033 if ((tdc
= avc
->dchint
)) {
2037 * The locking order between afs_xdcache and dcache lock matters.
2038 * The hint dcache entry could be anywhere, even on the free list.
2039 * Locking afs_xdcache ensures that noone is trying to pull dcache
2040 * entries from the free list, and thereby assuming them to be not
2041 * referenced and not locked.
2043 ObtainReadLock(&afs_xdcache
);
2044 dcLocked
= (0 == NBObtainSharedLock(&tdc
->lock
, 601));
2046 if (dcLocked
&& (tdc
->index
!= NULLIDX
)
2047 && !FidCmp(&tdc
->f
.fid
, &avc
->f
.fid
) && chunk
== tdc
->f
.chunk
2048 && !(afs_indexFlags
[tdc
->index
] & (IFFree
| IFDiscarded
))) {
2049 /* got the right one. It might not be the right version, and it
2050 * might be fetching, but it's the right dcache entry.
2052 /* All this code should be integrated better with what follows:
2053 * I can save a good bit more time under a write lock if I do..
2055 ObtainWriteLock(&tdc
->tlock
, 603);
2057 ReleaseWriteLock(&tdc
->tlock
);
2059 ReleaseReadLock(&afs_xdcache
);
2062 if (afs_IsDCacheFresh(tdc
, avc
)
2063 && !(tdc
->dflags
& DFFetching
)) {
2065 afs_stats_cmperf
.dcacheHits
++;
2066 ObtainWriteLock(&afs_xdcache
, 559);
2067 QRemove(&tdc
->lruq
);
2068 QAdd(&afs_DLRU
, &tdc
->lruq
);
2069 ReleaseWriteLock(&afs_xdcache
);
2072 * avc->lock(R) if setLocks && !slowPass
2073 * avc->lock(W) if !setLocks || slowPass
2080 ReleaseSharedLock(&tdc
->lock
);
2081 ReleaseReadLock(&afs_xdcache
);
2089 * avc->lock(R) if setLocks && !slowPass
2090 * avc->lock(W) if !setLocks || slowPass
2091 * tdc->lock(S) if tdc
2094 if (!tdc
) { /* If the hint wasn't the right dcache entry */
2095 int dslot_error
= 0;
2097 * Hash on the [fid, chunk] and get the corresponding dcache index
2098 * after write-locking the dcache.
2103 * avc->lock(R) if setLocks && !slowPass
2104 * avc->lock(W) if !setLocks || slowPass
2107 i
= DCHash(&avc
->f
.fid
, chunk
);
2108 /* check to make sure our space is fine */
2109 afs_MaybeWakeupTruncateDaemon();
2111 ObtainWriteLock(&afs_xdcache
, 280);
2113 for (index
= afs_dchashTbl
[i
]; index
!= NULLIDX
; us
= index
, index
= afs_dcnextTbl
[index
]) {
2114 if (afs_indexUnique
[index
] == avc
->f
.fid
.Fid
.Unique
) {
2115 tdc
= afs_GetValidDSlot(index
);
2117 /* we got an i/o error when trying to get the given dslot.
2118 * it's possible the dslot we're looking for is elsewhere,
2119 * but most likely the disk cache is currently unusable, so
2120 * all afs_GetValidDSlot calls will fail, so just bail out. */
2125 ReleaseReadLock(&tdc
->tlock
);
2128 * avc->lock(R) if setLocks && !slowPass
2129 * avc->lock(W) if !setLocks || slowPass
2132 if (!FidCmp(&tdc
->f
.fid
, &avc
->f
.fid
) && chunk
== tdc
->f
.chunk
) {
2133 /* Move it up in the beginning of the list */
2134 if (afs_dchashTbl
[i
] != index
) {
2135 afs_dcnextTbl
[us
] = afs_dcnextTbl
[index
];
2136 afs_dcnextTbl
[index
] = afs_dchashTbl
[i
];
2137 afs_dchashTbl
[i
] = index
;
2139 ReleaseWriteLock(&afs_xdcache
);
2140 ObtainSharedLock(&tdc
->lock
, 606);
2141 break; /* leaving refCount high for caller */
2149 * If we didn't find the entry, we'll create one.
2151 if (index
== NULLIDX
) {
2154 * avc->lock(R) if setLocks
2155 * avc->lock(W) if !setLocks
2158 afs_Trace2(afs_iclSetp
, CM_TRACE_GETDCACHE1
, ICL_TYPE_POINTER
,
2159 avc
, ICL_TYPE_INT32
, chunk
);
2162 /* We couldn't find the dcache we want, but we hit some i/o
2163 * errors when trying to find it, so we're not sure if the
2164 * dcache we want is in the cache or not. Error out, so we
2165 * don't try to possibly create 2 separate dcaches for the
2166 * same exact data. */
2167 ReleaseWriteLock(&afs_xdcache
);
2171 if (afs_discardDCList
== NULLIDX
&& afs_freeDCList
== NULLIDX
) {
2173 avc
->f
.states
|= CDCLock
;
2174 /* just need slots */
2175 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc
));
2177 avc
->f
.states
&= ~CDCLock
;
2179 code
= afs_AllocDCache(&tdc
, avc
, chunk
, aflags
, NULL
);
2181 ReleaseWriteLock(&afs_xdcache
);
2182 if (code
== ENOSPC
) {
2183 /* It looks like afs_AllocDCache failed because we don't
2184 * have any free dslots to use. Maybe if we wait a little
2185 * while, we'll be able to free up some slots, so try for 5
2186 * minutes, then bail out. */
2187 if (++downDCount
> 300) {
2188 afs_warn("afs: Unable to get free cache space for file "
2189 "%u:%u.%u.%u for 5 minutes; failing with an i/o error\n",
2191 avc
->f
.fid
.Fid
.Volume
,
2192 avc
->f
.fid
.Fid
.Vnode
,
2193 avc
->f
.fid
.Fid
.Unique
);
2196 afs_osi_Wait(1000, 0, 0);
2200 /* afs_AllocDCache failed, but not because we're out of free
2201 * dslots. Something must be screwy with the cache, so bail out
2202 * immediately without waiting. */
2203 afs_warn("afs: Error while alloc'ing cache slot for file "
2204 "%u:%u.%u.%u; failing with an i/o error\n",
2206 avc
->f
.fid
.Fid
.Volume
,
2207 avc
->f
.fid
.Fid
.Vnode
,
2208 avc
->f
.fid
.Fid
.Unique
);
2214 * avc->lock(R) if setLocks
2215 * avc->lock(W) if !setLocks
2221 * Now add to the two hash chains - note that i is still set
2222 * from the above DCHash call.
2224 afs_dcnextTbl
[tdc
->index
] = afs_dchashTbl
[i
];
2225 afs_dchashTbl
[i
] = tdc
->index
;
2226 i
= DVHash(&avc
->f
.fid
);
2227 afs_dvnextTbl
[tdc
->index
] = afs_dvhashTbl
[i
];
2228 afs_dvhashTbl
[i
] = tdc
->index
;
2229 tdc
->dflags
= DFEntryMod
;
2231 afs_MaybeWakeupTruncateDaemon();
2232 ReleaseWriteLock(&afs_xdcache
);
2233 ConvertWToSLock(&tdc
->lock
);
2238 /* vcache->dcache hint failed */
2241 * avc->lock(R) if setLocks && !slowPass
2242 * avc->lock(W) if !setLocks || slowPass
2245 afs_Trace4(afs_iclSetp
, CM_TRACE_GETDCACHE2
, ICL_TYPE_POINTER
, avc
,
2246 ICL_TYPE_POINTER
, tdc
, ICL_TYPE_INT32
,
2247 hgetlo(tdc
->f
.versionNo
), ICL_TYPE_INT32
,
2248 hgetlo(avc
->f
.m
.DataVersion
));
2250 * Here we have the entry in tdc, with its refCount incremented.
2251 * Note: we don't use the S-lock on avc; it costs concurrency when
2252 * storing a file back to the server.
2256 * Not a newly created file so we need to check the file's length and
2257 * compare data versions since someone could have changed the data or we're
2258 * reading a file written elsewhere. We only want to bypass doing no-op
2259 * read rpcs on newly created files (dv of 0) since only then we guarantee
2260 * that this chunk's data hasn't been filled by another client.
2262 size
= AFS_CHUNKSIZE(abyte
);
2263 if (aflags
& 4) /* called from write */
2265 else /* called from read */
2266 tlen
= tdc
->validPos
- abyte
;
2267 Position
= AFS_CHUNKTOBASE(chunk
);
2268 afs_Trace4(afs_iclSetp
, CM_TRACE_GETDCACHE3
, ICL_TYPE_INT32
, tlen
,
2269 ICL_TYPE_INT32
, aflags
, ICL_TYPE_OFFSET
,
2270 ICL_HANDLE_OFFSET(abyte
), ICL_TYPE_OFFSET
,
2271 ICL_HANDLE_OFFSET(Position
));
2272 if ((aflags
& 4) && (hiszero(avc
->f
.m
.DataVersion
)))
2274 if ((AFS_CHUNKTOBASE(chunk
) >= avc
->f
.m
.Length
) ||
2275 ((aflags
& 4) && (abyte
== Position
) && (tlen
>= size
)))
2276 overWriteWholeChunk
= 1;
2277 if (doAdjustSize
|| overWriteWholeChunk
) {
2278 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
2282 #else /* AFS_SGI_ENV */
2285 #endif /* AFS_SGI_ENV */
2286 if (AFS_CHUNKTOBASE(chunk
) + adjustsize
>= avc
->f
.m
.Length
&&
2287 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2288 #if defined(AFS_SUN5_ENV)
2289 if ((doAdjustSize
|| (AFS_CHUNKTOBASE(chunk
) >= avc
->f
.m
.Length
)) &&
2291 if (AFS_CHUNKTOBASE(chunk
) >= avc
->f
.m
.Length
&&
2293 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2294 !afs_IsDCacheFresh(tdc
, avc
))
2295 doReallyAdjustSize
= 1;
2297 if (doReallyAdjustSize
|| overWriteWholeChunk
) {
2298 /* no data in file to read at this position */
2299 UpgradeSToWLock(&tdc
->lock
, 607);
2300 file
= afs_CFileOpen(&tdc
->f
.inode
);
2302 afs_CFileTruncate(file
, 0);
2303 afs_CFileClose(file
);
2304 afs_AdjustSize(tdc
, 0);
2305 hset(tdc
->f
.versionNo
, avc
->f
.m
.DataVersion
);
2306 tdc
->dflags
|= DFEntryMod
;
2308 ConvertWToSLock(&tdc
->lock
);
2313 * We must read in the whole chunk if the version number doesn't
2317 /* don't need data, just a unique dcache entry */
2318 ObtainWriteLock(&afs_xdcache
, 608);
2319 hset(afs_indexTimes
[tdc
->index
], afs_indexCounter
);
2320 hadd32(afs_indexCounter
, 1);
2321 ReleaseWriteLock(&afs_xdcache
);
2323 updateV2DC(setLocks
, avc
, tdc
, 553);
2324 if (vType(avc
) == VDIR
)
2327 *aoffset
= AFS_CHUNKOFFSET(abyte
);
2328 if (tdc
->validPos
< abyte
)
2329 *alen
= (afs_size_t
) 0;
2331 *alen
= tdc
->validPos
- abyte
;
2332 ReleaseSharedLock(&tdc
->lock
);
2335 ReleaseWriteLock(&avc
->lock
);
2337 ReleaseReadLock(&avc
->lock
);
2339 return tdc
; /* check if we're done */
2344 * avc->lock(R) if setLocks && !slowPass
2345 * avc->lock(W) if !setLocks || slowPass
2348 osi_Assert((setLocks
&& !slowPass
) || WriteLocked(&avc
->lock
));
2350 setNewCallback
= setVcacheStatus
= 0;
2354 * avc->lock(R) if setLocks && !slowPass
2355 * avc->lock(W) if !setLocks || slowPass
2358 if (!afs_IsDCacheFresh(tdc
, avc
) && !overWriteWholeChunk
) {
2360 * Version number mismatch.
2363 * If we are disconnected, then we can't do much of anything
2364 * because the data doesn't match the file.
2366 if (AFS_IS_DISCONNECTED
) {
2367 ReleaseSharedLock(&tdc
->lock
);
2370 ReleaseWriteLock(&avc
->lock
);
2372 ReleaseReadLock(&avc
->lock
);
2374 /* Flush the Dcache */
2379 UpgradeSToWLock(&tdc
->lock
, 609);
2382 * If data ever existed for this vnode, and this is a text object,
2383 * do some clearing. Now, you'd think you need only do the flush
2384 * when VTEXT is on, but VTEXT is turned off when the text object
2385 * is freed, while pages are left lying around in memory marked
2386 * with this vnode. If we would reactivate (create a new text
2387 * object from) this vnode, we could easily stumble upon some of
2388 * these old pages in pagein. So, we always flush these guys.
2389 * Sun has a wonderful lack of useful invariants in this system.
2391 * avc->flushDV is the data version # of the file at the last text
2392 * flush. Clearly, at least, we don't have to flush the file more
2393 * often than it changes
2395 if (hcmp(avc
->flushDV
, avc
->f
.m
.DataVersion
) < 0) {
2397 * By here, the cache entry is always write-locked. We can
2398 * deadlock if we call osi_Flush with the cache entry locked...
2399 * Unlock the dcache too.
2401 ReleaseWriteLock(&tdc
->lock
);
2402 if (setLocks
&& !slowPass
)
2403 ReleaseReadLock(&avc
->lock
);
2405 ReleaseWriteLock(&avc
->lock
);
2409 * Call osi_FlushPages in open, read/write, and map, since it
2410 * is too hard here to figure out if we should lock the
2413 if (setLocks
&& !slowPass
)
2414 ObtainReadLock(&avc
->lock
);
2416 ObtainWriteLock(&avc
->lock
, 66);
2417 ObtainWriteLock(&tdc
->lock
, 610);
2422 * avc->lock(R) if setLocks && !slowPass
2423 * avc->lock(W) if !setLocks || slowPass
2427 /* Watch for standard race condition around osi_FlushText */
2428 if (afs_IsDCacheFresh(tdc
, avc
)) {
2429 updateV2DC(setLocks
, avc
, tdc
, 569); /* set hint */
2430 afs_stats_cmperf
.dcacheHits
++;
2431 ConvertWToSLock(&tdc
->lock
);
2435 /* Sleep here when cache needs to be drained. */
2436 if (setLocks
&& !slowPass
2437 && (afs_blocksUsed
>
2438 PERCENT(CM_WAITFORDRAINPCT
, afs_cacheBlocks
))) {
2439 /* Make sure truncate daemon is running */
2440 afs_MaybeWakeupTruncateDaemon();
2441 ObtainWriteLock(&tdc
->tlock
, 614);
2442 tdc
->refCount
--; /* we'll re-obtain the dcache when we re-try. */
2443 ReleaseWriteLock(&tdc
->tlock
);
2444 ReleaseWriteLock(&tdc
->lock
);
2445 ReleaseReadLock(&avc
->lock
);
2446 while ((afs_blocksUsed
- afs_blocksDiscarded
) >
2447 PERCENT(CM_WAITFORDRAINPCT
, afs_cacheBlocks
)) {
2448 afs_MaybeWaitForCacheDrain();
2450 afs_MaybeFreeDiscardedDCache();
2451 /* need to check if someone else got the chunk first. */
2452 goto RetryGetDCache
;
2455 Position
= AFS_CHUNKBASE(abyte
);
2456 if (vType(avc
) == VDIR
) {
2457 size
= avc
->f
.m
.Length
;
2458 if (size
> tdc
->f
.chunkBytes
) {
2459 /* pre-reserve space for file */
2460 afs_AdjustSize(tdc
, size
);
2462 size
= 999999999; /* max size for transfer */
2464 afs_size_t maxGoodLength
;
2466 /* estimate how much data we're expecting back from the server,
2467 * and reserve space in the dcache entry for it */
2469 maxGoodLength
= avc
->f
.m
.Length
;
2470 if (avc
->f
.truncPos
< maxGoodLength
)
2471 maxGoodLength
= avc
->f
.truncPos
;
2473 size
= AFS_CHUNKSIZE(abyte
); /* expected max size */
2474 if (Position
> maxGoodLength
) { /* If we're beyond EOF */
2476 } else if (Position
+ size
> maxGoodLength
) {
2477 size
= maxGoodLength
- Position
;
2479 osi_Assert(size
>= 0);
2481 if (size
> tdc
->f
.chunkBytes
) {
2482 /* pre-reserve estimated space for file */
2483 afs_AdjustSize(tdc
, size
); /* changes chunkBytes */
2487 /* For the actual fetch, do not limit the request to the
2488 * length of the file. If this results in a read past EOF on
2489 * the server, the server will just reply with less data than
2490 * requested. If we limit ourselves to only requesting data up
2491 * to the avc file length, we open ourselves up to races if the
2492 * file is extended on the server at about the same time.
2494 * However, we must restrict ourselves to the avc->f.truncPos
2495 * length, since this represents an outstanding local
2496 * truncation of the file that will be committed to the
2497 * fileserver when we actually write the fileserver contents.
2498 * If we do not restrict the fetch length based on
2499 * avc->f.truncPos, a different truncate operation extending
2500 * the file length could cause the old data after
2501 * avc->f.truncPos to reappear, instead of extending the file
2502 * with NUL bytes. */
2503 size
= AFS_CHUNKSIZE(abyte
);
2504 if (Position
> avc
->f
.truncPos
) {
2506 } else if (Position
+ size
> avc
->f
.truncPos
) {
2507 size
= avc
->f
.truncPos
- Position
;
2509 osi_Assert(size
>= 0);
2512 if (afs_mariner
&& !tdc
->f
.chunk
)
2513 afs_MarinerLog("fetch$Fetching", avc
); /* , Position, size, afs_indexCounter ); */
2515 * Right now, we only have one tool, and it's a hammer. So, we
2516 * fetch the whole file.
2518 DZap(tdc
); /* pages in cache may be old */
2519 file
= afs_CFileOpen(&tdc
->f
.inode
);
2521 /* We can't access the file in the disk cache backing this dcache;
2523 ReleaseWriteLock(&tdc
->lock
);
2528 afs_RemoveVCB(&avc
->f
.fid
);
2529 tdc
->f
.states
|= DWriting
;
2530 tdc
->dflags
|= DFFetching
;
2531 tdc
->validPos
= Position
; /* which is AFS_CHUNKBASE(abyte) */
2532 if (tdc
->mflags
& DFFetchReq
) {
2533 tdc
->mflags
&= ~DFFetchReq
;
2534 if (afs_osi_Wakeup(&tdc
->validPos
) == 0)
2535 afs_Trace4(afs_iclSetp
, CM_TRACE_DCACHEWAKE
, ICL_TYPE_STRING
,
2536 __FILE__
, ICL_TYPE_INT32
, __LINE__
,
2537 ICL_TYPE_POINTER
, tdc
, ICL_TYPE_INT32
,
2540 tsmall
= osi_AllocLargeSpace(sizeof(struct afs_FetchOutput
));
2541 setVcacheStatus
= 0;
2544 * Remember if we are doing the reading from a replicated volume,
2545 * and how many times we've zipped around the fetch/analyze loop.
2547 fromReplica
= (avc
->f
.states
& CRO
) ? 1 : 0;
2549 accP
= &(afs_stats_cmfullperf
.accessinf
);
2551 (accP
->replicatedRefs
)++;
2553 (accP
->unreplicatedRefs
)++;
2554 #endif /* AFS_NOSTATS */
2555 /* this is a cache miss */
2556 afs_Trace4(afs_iclSetp
, CM_TRACE_FETCHPROC
, ICL_TYPE_POINTER
, avc
,
2557 ICL_TYPE_FID
, &(avc
->f
.fid
), ICL_TYPE_OFFSET
,
2558 ICL_HANDLE_OFFSET(Position
), ICL_TYPE_INT32
, size
);
2561 afs_stats_cmperf
.dcacheMisses
++;
2563 * Dynamic root support: fetch data from local memory.
2565 if (afs_IsDynroot(avc
)) {
2569 afs_GetDynroot(&dynrootDir
, &dynrootLen
, &tsmall
->OutStatus
);
2571 dynrootDir
+= Position
;
2572 dynrootLen
-= Position
;
2573 if (size
> dynrootLen
)
2577 code
= afs_CFileWrite(file
, 0, dynrootDir
, size
);
2585 tdc
->validPos
= Position
+ size
;
2586 afs_CFileTruncate(file
, size
); /* prune it */
2587 } else if (afs_IsDynrootMount(avc
)) {
2591 afs_GetDynrootMount(&dynrootDir
, &dynrootLen
, &tsmall
->OutStatus
);
2593 dynrootDir
+= Position
;
2594 dynrootLen
-= Position
;
2595 if (size
> dynrootLen
)
2599 code
= afs_CFileWrite(file
, 0, dynrootDir
, size
);
2607 tdc
->validPos
= Position
+ size
;
2608 afs_CFileTruncate(file
, size
); /* prune it */
2611 * Not a dynamic vnode: do the real fetch.
2616 * avc->lock(R) if setLocks && !slowPass
2617 * avc->lock(W) if !setLocks || slowPass
2621 tc
= afs_Conn(&avc
->f
.fid
, areq
, SHARED_LOCK
, &rxconn
);
2626 (accP
->numReplicasAccessed
)++;
2628 #endif /* AFS_NOSTATS */
2629 if (!setLocks
|| slowPass
) {
2630 avc
->callback
= tc
->parent
->srvr
->server
;
2632 newCallback
= tc
->parent
->srvr
->server
;
2635 code
= afs_CacheFetchProc(tc
, rxconn
, file
, Position
, tdc
,
2641 /* callback could have been broken (or expired) in a race here,
2642 * but we return the data anyway. It's as good as we knew about
2643 * when we started. */
2645 * validPos is updated by CacheFetchProc, and can only be
2646 * modifed under a dcache write lock, which we've blocked out
2650 size
= tdc
->validPos
- Position
; /* actual segment size */
2653 afs_CFileTruncate(file
, size
); /* prune it */
2655 /* Check that the amount of data that we fetched for the
2656 * dcache makes sense. */
2657 FillInt64(length
, tsmall
->OutStatus
.Length_hi
, tsmall
->OutStatus
.Length
);
2658 if (!IsDCacheSizeOK(tdc
, avc
, size
,
2660 tsmall
->OutStatus
.DataVersion
, 1)) {
2665 if (!setLocks
|| slowPass
) {
2666 afs_StaleVCacheFlags(avc
, AFS_STALEVC_CLEARCB
, CUnique
);
2668 /* Something lost. Forget about performance, and go
2669 * back with a vcache write lock.
2671 afs_CFileTruncate(file
, 0);
2672 afs_AdjustSize(tdc
, 0);
2673 afs_CFileClose(file
);
2674 osi_FreeLargeSpace(tsmall
);
2676 ReleaseWriteLock(&tdc
->lock
);
2681 * Call afs_Analyze to manage the connection references
2682 * and handle the error code (possibly mark servers
2683 * down, etc). We are going to retry getting the
2684 * dcache regardless, so we just ignore the retry hint
2685 * returned by afs_Analyze on this call.
2687 (void)afs_Analyze(tc
, rxconn
, code
, &avc
->f
.fid
, areq
,
2688 AFS_STATS_FS_RPCIDX_FETCHDATA
, SHARED_LOCK
, NULL
);
2690 ReleaseReadLock(&avc
->lock
);
2693 goto RetryGetDCache
;
2697 } while (afs_Analyze
2698 (tc
, rxconn
, code
, &avc
->f
.fid
, areq
,
2699 AFS_STATS_FS_RPCIDX_FETCHDATA
, SHARED_LOCK
, NULL
));
2703 * avc->lock(R) if setLocks && !slowPass
2704 * avc->lock(W) if !setLocks || slowPass
2710 * In the case of replicated access, jot down info on the number of
2711 * attempts it took before we got through or gave up.
2714 if (numFetchLoops
<= 1)
2715 (accP
->refFirstReplicaOK
)++;
2716 if (numFetchLoops
> accP
->maxReplicasPerRef
)
2717 accP
->maxReplicasPerRef
= numFetchLoops
;
2719 #endif /* AFS_NOSTATS */
2721 tdc
->dflags
&= ~DFFetching
;
2722 if (afs_osi_Wakeup(&tdc
->validPos
) == 0)
2723 afs_Trace4(afs_iclSetp
, CM_TRACE_DCACHEWAKE
, ICL_TYPE_STRING
,
2724 __FILE__
, ICL_TYPE_INT32
, __LINE__
, ICL_TYPE_POINTER
,
2725 tdc
, ICL_TYPE_INT32
, tdc
->dflags
);
2726 if (avc
->execsOrWriters
== 0)
2727 tdc
->f
.states
&= ~DWriting
;
2729 /* now, if code != 0, we have an error and should punt.
2730 * note that we have the vcache write lock, either because
2731 * !setLocks or slowPass.
2734 afs_CFileTruncate(file
, 0);
2735 afs_AdjustSize(tdc
, 0);
2736 afs_CFileClose(file
);
2737 ZapDCE(tdc
); /* sets DFEntryMod */
2738 if (vType(avc
) == VDIR
) {
2741 tdc
->f
.states
&= ~(DRO
|DBackup
|DRW
);
2742 afs_DCMoveBucket(tdc
, 0, 0);
2743 ReleaseWriteLock(&tdc
->lock
);
2745 if (!afs_IsDynroot(avc
)) {
2746 afs_StaleVCacheFlags(avc
, 0, CUnique
);
2749 * avc->lock(W); assert(!setLocks || slowPass)
2751 osi_Assert(!setLocks
|| slowPass
);
2757 /* otherwise we copy in the just-fetched info */
2758 afs_CFileClose(file
);
2759 afs_AdjustSize(tdc
, size
); /* new size */
2761 * Copy appropriate fields into vcache. Status is
2762 * copied later where we selectively acquire the
2763 * vcache write lock.
2766 afs_ProcessFS(avc
, &tsmall
->OutStatus
, areq
);
2768 setVcacheStatus
= 1;
2769 hset64(tdc
->f
.versionNo
, tsmall
->OutStatus
.dataVersionHigh
,
2770 tsmall
->OutStatus
.DataVersion
);
2771 tdc
->dflags
|= DFEntryMod
;
2772 afs_indexFlags
[tdc
->index
] |= IFEverUsed
;
2773 ConvertWToSLock(&tdc
->lock
);
2774 } /*Data version numbers don't match */
2777 * Data version numbers match.
2779 afs_stats_cmperf
.dcacheHits
++;
2780 } /*Data version numbers match */
2782 updateV2DC(setLocks
, avc
, tdc
, 335); /* set hint */
2786 * avc->lock(R) if setLocks && !slowPass
2787 * avc->lock(W) if !setLocks || slowPass
2788 * tdc->lock(S) if tdc
2792 * See if this was a reference to a file in the local cell.
2794 if (afs_IsPrimaryCellNum(avc
->f
.fid
.Cell
))
2795 afs_stats_cmperf
.dlocalAccesses
++;
2797 afs_stats_cmperf
.dremoteAccesses
++;
2799 /* Fix up LRU info */
2802 ObtainWriteLock(&afs_xdcache
, 602);
2803 hset(afs_indexTimes
[tdc
->index
], afs_indexCounter
);
2804 hadd32(afs_indexCounter
, 1);
2805 ReleaseWriteLock(&afs_xdcache
);
2807 /* return the data */
2808 if (vType(avc
) == VDIR
)
2811 *aoffset
= AFS_CHUNKOFFSET(abyte
);
2812 *alen
= (tdc
->f
.chunkBytes
- *aoffset
);
2813 ReleaseSharedLock(&tdc
->lock
);
2818 * avc->lock(R) if setLocks && !slowPass
2819 * avc->lock(W) if !setLocks || slowPass
2822 /* Fix up the callback and status values in the vcache */
2824 if (setLocks
&& !slowPass
) {
2827 * This is our dirty little secret to parallel fetches.
2828 * We don't write-lock the vcache while doing the fetch,
2829 * but potentially we'll need to update the vcache after
2830 * the fetch is done.
2832 * Drop the read lock and try to re-obtain the write
2833 * lock. If the vcache still has the same DV, it's
2834 * ok to go ahead and install the new data.
2836 afs_hyper_t currentDV
, statusDV
;
2838 hset(currentDV
, avc
->f
.m
.DataVersion
);
2840 if (setNewCallback
&& avc
->callback
!= newCallback
)
2844 hset64(statusDV
, tsmall
->OutStatus
.dataVersionHigh
,
2845 tsmall
->OutStatus
.DataVersion
);
2847 if (setVcacheStatus
&& avc
->f
.m
.Length
!= tsmall
->OutStatus
.Length
)
2849 if (setVcacheStatus
&& !hsame(currentDV
, statusDV
))
2853 ReleaseReadLock(&avc
->lock
);
2855 if (doVcacheUpdate
) {
2856 ObtainWriteLock(&avc
->lock
, 615);
2857 if (!hsame(avc
->f
.m
.DataVersion
, currentDV
)) {
2858 /* We lose. Someone will beat us to it. */
2860 ReleaseWriteLock(&avc
->lock
);
2865 /* With slow pass, we've already done all the updates */
2867 ReleaseWriteLock(&avc
->lock
);
2870 /* Check if we need to perform any last-minute fixes with a write-lock */
2871 if (!setLocks
|| doVcacheUpdate
) {
2873 avc
->callback
= newCallback
;
2874 if (tsmall
&& setVcacheStatus
)
2875 afs_ProcessFS(avc
, &tsmall
->OutStatus
, areq
);
2877 ReleaseWriteLock(&avc
->lock
);
2881 osi_FreeLargeSpace(tsmall
);
2884 } /*afs_GetDCache */
2888 * afs_WriteThroughDSlots
2891 * Sweep through the dcache slots and write out any modified
2892 * in-memory data back on to our caching store.
2898 * The afs_xdcache is write-locked through this whole affair.
2901 afs_WriteThroughDSlots(void)
2904 afs_int32 i
, touchedit
= 0;
2907 struct afs_q DirtyQ
, *tq
;
2909 AFS_STATCNT(afs_WriteThroughDSlots
);
2912 * Because of lock ordering, we can't grab dcache locks while
2913 * holding afs_xdcache. So we enter xdcache, get a reference
2914 * for every dcache entry, and exit xdcache.
2916 ObtainWriteLock(&afs_xdcache
, 283);
2918 for (i
= 0; i
< afs_cacheFiles
; i
++) {
2919 tdc
= afs_indexTable
[i
];
2921 /* Grab tlock in case the existing refcount isn't zero */
2922 if (tdc
&& !(afs_indexFlags
[i
] & (IFFree
| IFDiscarded
))) {
2923 ObtainWriteLock(&tdc
->tlock
, 623);
2925 ReleaseWriteLock(&tdc
->tlock
);
2927 QAdd(&DirtyQ
, &tdc
->dirty
);
2930 ReleaseWriteLock(&afs_xdcache
);
2933 * Now, for each dcache entry we found, check if it's dirty.
2934 * If so, get write-lock, get afs_xdcache, which protects
2935 * afs_cacheInodep, and flush it. Don't forget to put back
2939 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2941 for (tq
= DirtyQ
.prev
; tq
!= &DirtyQ
&& code
== 0; tq
= QPrev(tq
)) {
2943 if (tdc
->dflags
& DFEntryMod
) {
2946 wrLock
= (0 == NBObtainWriteLock(&tdc
->lock
, 619));
2948 /* Now that we have the write lock, double-check */
2949 if (wrLock
&& (tdc
->dflags
& DFEntryMod
)) {
2950 tdc
->dflags
&= ~DFEntryMod
;
2951 ObtainWriteLock(&afs_xdcache
, 620);
2952 code
= afs_WriteDCache(tdc
, 1);
2953 ReleaseWriteLock(&afs_xdcache
);
2955 /* We didn't successfully write out the dslot; make sure we
2956 * try again later */
2957 tdc
->dflags
|= DFEntryMod
;
2963 ReleaseWriteLock(&tdc
->lock
);
2973 ObtainWriteLock(&afs_xdcache
, 617);
2974 if (!touchedit
&& (cacheDiskType
!= AFS_FCACHE_TYPE_MEM
)) {
2975 /* Touch the file to make sure that the mtime on the file is kept
2976 * up-to-date to avoid losing cached files on cold starts because
2977 * their mtime seems old...
2979 struct afs_fheader theader
;
2981 afs_InitFHeader(&theader
);
2982 afs_osi_Write(afs_cacheInodep
, 0, &theader
, sizeof(theader
));
2984 ReleaseWriteLock(&afs_xdcache
);
2992 * Return a pointer to an freshly initialized dcache entry using
2993 * a memory-based cache. The tlock will be read-locked.
2996 * aslot : Dcache slot to look at.
2997 * type : What 'type' of dslot to get; see the dslot_state enum
3000 * Must be called with afs_xdcache write-locked.
3004 afs_MemGetDSlot(afs_int32 aslot
, dslot_state type
)
3009 AFS_STATCNT(afs_MemGetDSlot
);
3010 if (CheckLock(&afs_xdcache
) != -1)
3011 osi_Panic("getdslot nolock");
3012 if (aslot
< 0 || aslot
>= afs_cacheFiles
)
3013 osi_Panic("getdslot slot %d (of %d)", aslot
, afs_cacheFiles
);
3014 tdc
= afs_indexTable
[aslot
];
3016 QRemove(&tdc
->lruq
); /* move to queue head */
3017 QAdd(&afs_DLRU
, &tdc
->lruq
);
3018 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
3019 ObtainWriteLock(&tdc
->tlock
, 624);
3021 ConvertWToRLock(&tdc
->tlock
);
3025 /* if we got here, the given slot is not in memory in our list of known
3026 * slots. for memcache, the only place a dslot can exist is in memory, so
3027 * if the caller is expecting to get back a known dslot, and we've reached
3028 * here, something is very wrong. DSLOT_NEW is the only type of dslot that
3029 * may not exist; for all others, the caller assumes the given dslot
3030 * already exists. so, 'type' had better be DSLOT_NEW here, or something is
3032 osi_Assert(type
== DSLOT_NEW
);
3034 if (!afs_freeDSList
)
3035 afs_GetDownDSlot(4);
3036 if (!afs_freeDSList
) {
3037 /* none free, making one is better than a panic */
3038 afs_stats_cmperf
.dcacheXAllocs
++; /* count in case we have a leak */
3039 tdc
= afs_osi_Alloc(sizeof(struct dcache
));
3040 osi_Assert(tdc
!= NULL
);
3041 #ifdef KERNEL_HAVE_PIN
3042 pin((char *)tdc
, sizeof(struct dcache
)); /* XXX */
3045 tdc
= afs_freeDSList
;
3046 afs_freeDSList
= (struct dcache
*)tdc
->lruq
.next
;
3049 tdc
->dflags
= 0; /* up-to-date, not in free q */
3051 QAdd(&afs_DLRU
, &tdc
->lruq
);
3052 if (tdc
->lruq
.prev
== &tdc
->lruq
)
3053 osi_Panic("lruq 3");
3055 /* initialize entry */
3056 tdc
->f
.fid
.Cell
= 0;
3057 tdc
->f
.fid
.Fid
.Volume
= 0;
3059 hones(tdc
->f
.versionNo
);
3060 tdc
->f
.inode
.mem
= aslot
;
3061 tdc
->dflags
|= DFEntryMod
;
3064 afs_indexUnique
[aslot
] = tdc
->f
.fid
.Fid
.Unique
;
3067 osi_Assert(0 == NBObtainWriteLock(&tdc
->lock
, 674));
3068 osi_Assert(0 == NBObtainWriteLock(&tdc
->mflock
, 675));
3069 osi_Assert(0 == NBObtainWriteLock(&tdc
->tlock
, 676));
3072 AFS_RWLOCK_INIT(&tdc
->lock
, "dcache lock");
3073 AFS_RWLOCK_INIT(&tdc
->tlock
, "dcache tlock");
3074 AFS_RWLOCK_INIT(&tdc
->mflock
, "dcache flock");
3075 ObtainReadLock(&tdc
->tlock
);
3077 afs_indexTable
[aslot
] = tdc
;
3080 } /*afs_MemGetDSlot */
3083 LogCacheError(int aslot
, int off
, int code
, int target_size
)
3085 struct osi_stat tstat
;
3088 if (afs_osi_Stat(afs_cacheInodep
, &tstat
)) {
3092 procname
= osi_AllocSmallSpace(AFS_SMALLOCSIZ
);
3093 if (procname
!= NULL
) {
3094 osi_procname(procname
, AFS_SMALLOCSIZ
);
3095 procname
[AFS_SMALLOCSIZ
-1] = '\0';
3098 afs_warn("afs: disk cache read error in CacheItems slot %d "
3099 "off %d/%d code %d/%d pid %d (%s)\n",
3100 aslot
, off
, (int)tstat
.size
, code
, target_size
,
3101 (int)MyPidxx2Pid(MyPidxx
),
3102 procname
? procname
: "");
3104 if (procname
!= NULL
) {
3105 osi_FreeSmallSpace(procname
);
3110 unsigned int last_error
= 0, lasterrtime
= 0;
3116 * Return a pointer to an freshly initialized dcache entry using
3117 * a UFS-based disk cache. The dcache tlock will be read-locked.
3120 * aslot : Dcache slot to look at.
3121 * type : What 'type' of dslot to get; see the dslot_state enum
3124 * afs_xdcache lock write-locked.
3127 afs_UFSGetDSlot(afs_int32 aslot
, dslot_state type
)
3135 AFS_STATCNT(afs_UFSGetDSlot
);
3136 if (CheckLock(&afs_xdcache
) != -1)
3137 osi_Panic("getdslot nolock");
3138 if (aslot
< 0 || aslot
>= afs_cacheFiles
)
3139 osi_Panic("getdslot slot %d (of %d)", aslot
, afs_cacheFiles
);
3140 tdc
= afs_indexTable
[aslot
];
3142 QRemove(&tdc
->lruq
); /* move to queue head */
3143 QAdd(&afs_DLRU
, &tdc
->lruq
);
3144 /* Grab tlock in case refCount != 0 */
3145 ObtainWriteLock(&tdc
->tlock
, 625);
3147 ConvertWToRLock(&tdc
->tlock
);
3151 /* otherwise we should read it in from the cache file */
3152 if (!afs_freeDSList
)
3153 afs_GetDownDSlot(4);
3154 if (!afs_freeDSList
) {
3155 /* none free, making one is better than a panic */
3156 afs_stats_cmperf
.dcacheXAllocs
++; /* count in case we have a leak */
3157 tdc
= afs_osi_Alloc(sizeof(struct dcache
));
3158 osi_Assert(tdc
!= NULL
);
3159 #ifdef KERNEL_HAVE_PIN
3160 pin((char *)tdc
, sizeof(struct dcache
)); /* XXX */
3163 tdc
= afs_freeDSList
;
3164 afs_freeDSList
= (struct dcache
*)tdc
->lruq
.next
;
3167 tdc
->dflags
= 0; /* up-to-date, not in free q */
3169 QAdd(&afs_DLRU
, &tdc
->lruq
);
3170 if (tdc
->lruq
.prev
== &tdc
->lruq
)
3171 osi_Panic("lruq 3");
3174 * Seek to the aslot'th entry and read it in.
3176 off
= sizeof(struct fcache
)*aslot
+ sizeof(struct afs_fheader
);
3178 afs_osi_Read(afs_cacheInodep
,
3179 off
, (char *)(&tdc
->f
),
3180 sizeof(struct fcache
));
3182 if (code
!= sizeof(struct fcache
)) {
3184 #if defined(KERNEL_HAVE_UERROR)
3185 last_error
= getuerror();
3189 lasterrtime
= osi_Time();
3190 if (type
!= DSLOT_NEW
) {
3191 /* If we are requesting a non-DSLOT_NEW slot, this is an error.
3192 * non-DSLOT_NEW slots are supposed to already exist, so if we
3193 * failed to read in the slot, something is wrong. */
3194 LogCacheError(aslot
, off
, code
, sizeof(struct fcache
));
3196 /* put tdc back on the free dslot list */
3197 QRemove(&tdc
->lruq
);
3198 tdc
->index
= NULLIDX
;
3199 tdc
->lruq
.next
= (struct afs_q
*)afs_freeDSList
;
3200 afs_freeDSList
= tdc
;
3204 if (!afs_CellNumValid(tdc
->f
.fid
.Cell
)) {
3206 if (type
== DSLOT_VALID
) {
3207 osi_Panic("afs: needed valid dcache but index %d off %d has "
3208 "invalid cell num %d\n",
3209 (int)aslot
, off
, (int)tdc
->f
.fid
.Cell
);
3213 if (type
== DSLOT_VALID
&& tdc
->f
.fid
.Fid
.Volume
== 0) {
3214 osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
3218 if (type
== DSLOT_UNUSED
) {
3219 /* the requested dslot is known to exist, but contain invalid data
3220 * (this happens when we're using a dslot from the free or discard
3221 * list). be sure not to re-use the data in it, so force invalidation.
3227 tdc
->f
.fid
.Cell
= 0;
3228 tdc
->f
.fid
.Fid
.Volume
= 0;
3230 hones(tdc
->f
.versionNo
);
3231 tdc
->dflags
|= DFEntryMod
;
3232 afs_indexUnique
[aslot
] = tdc
->f
.fid
.Fid
.Unique
;
3233 tdc
->f
.states
&= ~(DRO
|DBackup
|DRW
);
3234 afs_DCMoveBucket(tdc
, 0, 0);
3236 if (tdc
->f
.states
& DRO
) {
3237 afs_DCMoveBucket(tdc
, 0, 2);
3238 } else if (tdc
->f
.states
& DBackup
) {
3239 afs_DCMoveBucket(tdc
, 0, 1);
3241 afs_DCMoveBucket(tdc
, 0, 1);
3246 if (tdc
->f
.chunk
>= 0)
3247 tdc
->validPos
= AFS_CHUNKTOBASE(tdc
->f
.chunk
) + tdc
->f
.chunkBytes
;
3252 osi_Assert(0 == NBObtainWriteLock(&tdc
->lock
, 674));
3253 osi_Assert(0 == NBObtainWriteLock(&tdc
->mflock
, 675));
3254 osi_Assert(0 == NBObtainWriteLock(&tdc
->tlock
, 676));
3257 AFS_RWLOCK_INIT(&tdc
->lock
, "dcache lock");
3258 AFS_RWLOCK_INIT(&tdc
->tlock
, "dcache tlock");
3259 AFS_RWLOCK_INIT(&tdc
->mflock
, "dcache flock");
3260 ObtainReadLock(&tdc
->tlock
);
3263 * If we didn't read into a temporary dcache region, update the
3264 * slot pointer table.
3266 afs_indexTable
[aslot
] = tdc
;
3269 } /*afs_UFSGetDSlot */
3274 * Write a particular dcache entry back to its home in the
3277 * \param adc Pointer to the dcache entry to write.
3278 * \param atime If true, set the modtime on the file to the current time.
3280 * \note Environment:
3281 * Must be called with the afs_xdcache lock at least read-locked,
3282 * and dcache entry at least read-locked.
3283 * The reference count is not changed.
3287 afs_WriteDCache(struct dcache
*adc
, int atime
)
3291 if (cacheDiskType
== AFS_FCACHE_TYPE_MEM
)
3293 AFS_STATCNT(afs_WriteDCache
);
3294 osi_Assert(WriteLocked(&afs_xdcache
));
3296 adc
->f
.modTime
= osi_Time();
3298 if ((afs_indexFlags
[adc
->index
] & (IFFree
| IFDiscarded
)) == 0 &&
3299 adc
->f
.fid
.Fid
.Volume
== 0) {
3300 /* If a dcache slot is not on the free or discard list, it must be
3301 * in the hash table. Thus, the volume must be non-zero, since that
3302 * is how we determine whether or not to unhash the entry when kicking
3303 * it out of the cache. Do this check now, since otherwise this can
3304 * cause hash table corruption and a panic later on after we read the
3306 osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
3307 adc
->index
, (unsigned)afs_indexFlags
[adc
->index
]);
3311 * Seek to the right dcache slot and write the in-memory image out to disk.
3313 afs_cellname_write();
3315 afs_osi_Write(afs_cacheInodep
,
3316 sizeof(struct fcache
) * adc
->index
+
3317 sizeof(struct afs_fheader
), (char *)(&adc
->f
),
3318 sizeof(struct fcache
));
3319 if (code
!= sizeof(struct fcache
)) {
3320 afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
3321 (long)(sizeof(struct fcache
) * adc
->index
+ sizeof(struct afs_fheader
)),
3322 (int)code
, (int)sizeof(struct fcache
));
3331 * Wake up users of a particular file waiting for stores to take
3334 * \param avc Ptr to related vcache entry.
3336 * \note Environment:
3337 * Nothing interesting.
3340 afs_wakeup(struct vcache
*avc
)
3343 struct brequest
*tb
;
3345 AFS_STATCNT(afs_wakeup
);
3346 for (i
= 0; i
< NBRS
; i
++, tb
++) {
3347 /* if request is valid and for this file, we've found it */
3348 if (tb
->refCount
> 0 && avc
== tb
->vc
) {
3351 * If CSafeStore is on, then we don't awaken the guy
3352 * waiting for the store until the whole store has finished.
3353 * Otherwise, we do it now. Note that if CSafeStore is on,
3354 * the BStore routine actually wakes up the user, instead
3356 * I think this is redundant now because this sort of thing
3357 * is already being handled by the higher-level code.
3359 if ((avc
->f
.states
& CSafeStore
) == 0) {
3360 tb
->code_raw
= tb
->code_checkcode
= 0;
3361 tb
->flags
|= BUVALID
;
3362 if (tb
->flags
& BUWAIT
) {
3363 tb
->flags
&= ~BUWAIT
;
3374 * Given a file name and inode, set up that file to be an
3375 * active member in the AFS cache. This also involves checking
3376 * the usability of its data.
3378 * \param afile Name of the cache file to initialize.
3379 * \param ainode Inode of the file.
3381 * \note Environment:
3382 * This function is called only during initialization.
3385 afs_InitCacheFile(char *afile
, ino_t ainode
)
3390 struct osi_file
*tfile
;
3391 struct osi_stat tstat
;
3394 AFS_STATCNT(afs_InitCacheFile
);
3395 index
= afs_stats_cmperf
.cacheNumEntries
;
3396 if (index
>= afs_cacheFiles
)
3399 ObtainWriteLock(&afs_xdcache
, 282);
3400 tdc
= afs_GetNewDSlot(index
);
3401 ReleaseReadLock(&tdc
->tlock
);
3402 ReleaseWriteLock(&afs_xdcache
);
3404 ObtainWriteLock(&tdc
->lock
, 621);
3405 ObtainWriteLock(&afs_xdcache
, 622);
3406 if (!afile
&& !ainode
) {
3411 code
= afs_LookupInodeByPath(afile
, &tdc
->f
.inode
.ufs
, NULL
);
3413 ReleaseWriteLock(&afs_xdcache
);
3414 ReleaseWriteLock(&tdc
->lock
);
3419 /* Add any other 'complex' inode types here ... */
3420 #if !defined(AFS_LINUX_ENV) && !defined(AFS_CACHE_VNODE_PATH)
3421 tdc
->f
.inode
.ufs
= ainode
;
3423 osi_Panic("Can't init cache with inode numbers when complex inodes are "
3428 if ((tdc
->f
.states
& DWriting
) || tdc
->f
.fid
.Fid
.Volume
== 0)
3430 tfile
= osi_UFSOpen(&tdc
->f
.inode
);
3432 ReleaseWriteLock(&afs_xdcache
);
3433 ReleaseWriteLock(&tdc
->lock
);
3438 code
= afs_osi_Stat(tfile
, &tstat
);
3440 osi_Panic("initcachefile stat");
3443 * If file size doesn't match the cache info file, it's probably bad.
3445 if (tdc
->f
.chunkBytes
!= tstat
.size
)
3448 * If file changed within T (120?) seconds of cache info file, it's
3449 * probably bad. In addition, if slot changed within last T seconds,
3450 * the cache info file may be incorrectly identified, and so slot
3453 if (cacheInfoModTime
< tstat
.mtime
+ 120)
3455 if (cacheInfoModTime
< tdc
->f
.modTime
+ 120)
3457 /* In case write through is behind, make sure cache items entry is
3458 * at least as new as the chunk.
3460 if (tdc
->f
.modTime
< tstat
.mtime
)
3463 tdc
->f
.chunkBytes
= 0;
3466 tdc
->f
.fid
.Fid
.Volume
= 0; /* not in the hash table */
3467 if (tfile
&& tstat
.size
!= 0)
3468 osi_UFSTruncate(tfile
, 0);
3469 tdc
->f
.states
&= ~(DRO
|DBackup
|DRW
);
3470 afs_DCMoveBucket(tdc
, 0, 0);
3471 /* put entry in free cache slot list */
3472 afs_dvnextTbl
[tdc
->index
] = afs_freeDCList
;
3473 afs_freeDCList
= index
;
3475 afs_indexFlags
[index
] |= IFFree
;
3476 afs_indexUnique
[index
] = 0;
3479 * We must put this entry in the appropriate hash tables.
3480 * Note that i is still set from the above DCHash call
3482 code
= DCHash(&tdc
->f
.fid
, tdc
->f
.chunk
);
3483 afs_dcnextTbl
[tdc
->index
] = afs_dchashTbl
[code
];
3484 afs_dchashTbl
[code
] = tdc
->index
;
3485 code
= DVHash(&tdc
->f
.fid
);
3486 afs_dvnextTbl
[tdc
->index
] = afs_dvhashTbl
[code
];
3487 afs_dvhashTbl
[code
] = tdc
->index
;
3488 afs_AdjustSize(tdc
, tstat
.size
); /* adjust to new size */
3490 /* has nontrivial amt of data */
3491 afs_indexFlags
[index
] |= IFEverUsed
;
3492 afs_stats_cmperf
.cacheFilesReused
++;
3494 * Initialize index times to file's mod times; init indexCounter
3497 hset32(afs_indexTimes
[index
], tstat
.atime
);
3498 if (hgetlo(afs_indexCounter
) < tstat
.atime
) {
3499 hset32(afs_indexCounter
, tstat
.atime
);
3501 afs_indexUnique
[index
] = tdc
->f
.fid
.Fid
.Unique
;
3502 } /*File is not bad */
3505 osi_UFSClose(tfile
);
3506 tdc
->f
.states
&= ~DWriting
;
3507 tdc
->dflags
&= ~DFEntryMod
;
3508 /* don't set f.modTime; we're just cleaning up */
3509 osi_Assert(afs_WriteDCache(tdc
, 0) == 0);
3510 ReleaseWriteLock(&afs_xdcache
);
3511 ReleaseWriteLock(&tdc
->lock
);
3513 afs_stats_cmperf
.cacheNumEntries
++;
3518 /*Max # of struct dcache's resident at any time*/
3520 * If 'dchint' is enabled then in-memory dcache min is increased because of
3526 * Initialize dcache related variables.
3536 afs_dcacheInit(int afiles
, int ablocks
, int aDentries
, int achunk
, int aflags
)
3543 afs_freeDCList
= NULLIDX
;
3544 afs_discardDCList
= NULLIDX
;
3545 afs_freeDCCount
= 0;
3546 afs_freeDSList
= NULL
;
3547 hzero(afs_indexCounter
);
3549 LOCK_INIT(&afs_xdcache
, "afs_xdcache");
3555 if (achunk
< 0 || achunk
> 30)
3556 achunk
= 13; /* Use default */
3557 AFS_SETCHUNKSIZE(achunk
);
3563 /* afs_dhashsize defaults to 1024 */
3564 if (aDentries
> 512)
3565 afs_dhashsize
= 2048;
3566 /* Try to keep the average chain length around two unless the table
3567 * would be ridiculously big. */
3568 if (aDentries
> 4096) {
3569 afs_dhashbits
= opr_fls(aDentries
) - 3;
3570 /* Cap the hash tables to 32k entries. */
3571 if (afs_dhashbits
> 15)
3573 afs_dhashsize
= opr_jhash_size(afs_dhashbits
);
3575 /* initialize hash tables */
3576 afs_dvhashTbl
= afs_osi_Alloc(afs_dhashsize
* sizeof(afs_int32
));
3577 osi_Assert(afs_dvhashTbl
!= NULL
);
3578 afs_dchashTbl
= afs_osi_Alloc(afs_dhashsize
* sizeof(afs_int32
));
3579 osi_Assert(afs_dchashTbl
!= NULL
);
3580 for (i
= 0; i
< afs_dhashsize
; i
++) {
3581 afs_dvhashTbl
[i
] = NULLIDX
;
3582 afs_dchashTbl
[i
] = NULLIDX
;
3584 afs_dvnextTbl
= afs_osi_Alloc(afiles
* sizeof(afs_int32
));
3585 osi_Assert(afs_dvnextTbl
!= NULL
);
3586 afs_dcnextTbl
= afs_osi_Alloc(afiles
* sizeof(afs_int32
));
3587 osi_Assert(afs_dcnextTbl
!= NULL
);
3588 for (i
= 0; i
< afiles
; i
++) {
3589 afs_dvnextTbl
[i
] = NULLIDX
;
3590 afs_dcnextTbl
[i
] = NULLIDX
;
3593 /* Allocate and zero the pointer array to the dcache entries */
3594 afs_indexTable
= afs_osi_Alloc(sizeof(struct dcache
*) * afiles
);
3595 osi_Assert(afs_indexTable
!= NULL
);
3596 memset(afs_indexTable
, 0, sizeof(struct dcache
*) * afiles
);
3597 afs_indexTimes
= afs_osi_Alloc(afiles
* sizeof(afs_hyper_t
));
3598 osi_Assert(afs_indexTimes
!= NULL
);
3599 memset(afs_indexTimes
, 0, afiles
* sizeof(afs_hyper_t
));
3600 afs_indexUnique
= afs_osi_Alloc(afiles
* sizeof(afs_uint32
));
3601 osi_Assert(afs_indexUnique
!= NULL
);
3602 memset(afs_indexUnique
, 0, afiles
* sizeof(afs_uint32
));
3603 afs_indexFlags
= afs_osi_Alloc(afiles
* sizeof(u_char
));
3604 osi_Assert(afs_indexFlags
!= NULL
);
3605 memset(afs_indexFlags
, 0, afiles
* sizeof(char));
3607 /* Allocate and thread the struct dcache entries themselves */
3608 tdp
= afs_Initial_freeDSList
=
3609 afs_osi_Alloc(aDentries
* sizeof(struct dcache
));
3610 osi_Assert(tdp
!= NULL
);
3611 memset(tdp
, 0, aDentries
* sizeof(struct dcache
));
3612 #ifdef KERNEL_HAVE_PIN
3613 pin((char *)afs_indexTable
, sizeof(struct dcache
*) * afiles
); /* XXX */
3614 pin((char *)afs_indexTimes
, sizeof(afs_hyper_t
) * afiles
); /* XXX */
3615 pin((char *)afs_indexFlags
, sizeof(char) * afiles
); /* XXX */
3616 pin((char *)afs_indexUnique
, sizeof(afs_int32
) * afiles
); /* XXX */
3617 pin((char *)tdp
, aDentries
* sizeof(struct dcache
)); /* XXX */
3618 pin((char *)afs_dvhashTbl
, sizeof(afs_int32
) * afs_dhashsize
); /* XXX */
3619 pin((char *)afs_dchashTbl
, sizeof(afs_int32
) * afs_dhashsize
); /* XXX */
3620 pin((char *)afs_dcnextTbl
, sizeof(afs_int32
) * afiles
); /* XXX */
3621 pin((char *)afs_dvnextTbl
, sizeof(afs_int32
) * afiles
); /* XXX */
3624 afs_freeDSList
= &tdp
[0];
3625 for (i
= 0; i
< aDentries
- 1; i
++) {
3626 tdp
[i
].lruq
.next
= (struct afs_q
*)(&tdp
[i
+ 1]);
3627 AFS_RWLOCK_INIT(&tdp
[i
].lock
, "dcache lock");
3628 AFS_RWLOCK_INIT(&tdp
[i
].tlock
, "dcache tlock");
3629 AFS_RWLOCK_INIT(&tdp
[i
].mflock
, "dcache flock");
3631 tdp
[aDentries
- 1].lruq
.next
= (struct afs_q
*)0;
3632 AFS_RWLOCK_INIT(&tdp
[aDentries
- 1].lock
, "dcache lock");
3633 AFS_RWLOCK_INIT(&tdp
[aDentries
- 1].tlock
, "dcache tlock");
3634 AFS_RWLOCK_INIT(&tdp
[aDentries
- 1].mflock
, "dcache flock");
3636 afs_stats_cmperf
.cacheBlocksOrig
= afs_stats_cmperf
.cacheBlocksTotal
=
3637 afs_cacheBlocks
= ablocks
;
3638 afs_ComputeCacheParms(); /* compute parms based on cache size */
3640 afs_dcentries
= aDentries
;
3642 afs_stats_cmperf
.cacheBucket0_Discarded
=
3643 afs_stats_cmperf
.cacheBucket1_Discarded
=
3644 afs_stats_cmperf
.cacheBucket2_Discarded
= 0;
3648 if (aflags
& AFSCALL_INIT_MEMCACHE
) {
3650 * Use a memory cache instead of a disk cache
3652 cacheDiskType
= AFS_FCACHE_TYPE_MEM
;
3653 afs_cacheType
= &afs_MemCacheOps
;
3654 afiles
= (afiles
< aDentries
) ? afiles
: aDentries
; /* min */
3655 code
= afs_InitMemCache(afiles
, AFS_FIRSTCSIZE
, aflags
);
3657 afs_warn("afsd: memory cache too large for available memory.\n");
3658 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3662 afs_warn("Memory cache: Allocating %d dcache entries...",
3665 cacheDiskType
= AFS_FCACHE_TYPE_UFS
;
3666 afs_cacheType
= &afs_UfsCacheOps
;
3672 * Shuts down the cache.
3676 shutdown_dcache(void)
3680 #ifdef AFS_CACHE_VNODE_PATH
3681 if (cacheDiskType
!= AFS_FCACHE_TYPE_MEM
) {
3683 for (i
= 0; i
< afs_cacheFiles
; i
++) {
3684 tdc
= afs_indexTable
[i
];
3686 afs_osi_FreeStr(tdc
->f
.inode
.ufs
);
3692 afs_osi_Free(afs_dvnextTbl
, afs_cacheFiles
* sizeof(afs_int32
));
3693 afs_osi_Free(afs_dcnextTbl
, afs_cacheFiles
* sizeof(afs_int32
));
3694 afs_osi_Free(afs_indexTable
, afs_cacheFiles
* sizeof(struct dcache
*));
3695 afs_osi_Free(afs_indexTimes
, afs_cacheFiles
* sizeof(afs_hyper_t
));
3696 afs_osi_Free(afs_indexUnique
, afs_cacheFiles
* sizeof(afs_uint32
));
3697 afs_osi_Free(afs_indexFlags
, afs_cacheFiles
* sizeof(u_char
));
3698 afs_osi_Free(afs_Initial_freeDSList
,
3699 afs_dcentries
* sizeof(struct dcache
));
3700 #ifdef KERNEL_HAVE_PIN
3701 unpin((char *)afs_dcnextTbl
, afs_cacheFiles
* sizeof(afs_int32
));
3702 unpin((char *)afs_dvnextTbl
, afs_cacheFiles
* sizeof(afs_int32
));
3703 unpin((char *)afs_indexTable
, afs_cacheFiles
* sizeof(struct dcache
*));
3704 unpin((char *)afs_indexTimes
, afs_cacheFiles
* sizeof(afs_hyper_t
));
3705 unpin((char *)afs_indexUnique
, afs_cacheFiles
* sizeof(afs_uint32
));
3706 unpin((u_char
*) afs_indexFlags
, afs_cacheFiles
* sizeof(u_char
));
3707 unpin(afs_Initial_freeDSList
, afs_dcentries
* sizeof(struct dcache
));
3711 for (i
= 0; i
< afs_dhashsize
; i
++) {
3712 afs_dvhashTbl
[i
] = NULLIDX
;
3713 afs_dchashTbl
[i
] = NULLIDX
;
3716 afs_osi_Free(afs_dvhashTbl
, afs_dhashsize
* sizeof(afs_int32
));
3717 afs_osi_Free(afs_dchashTbl
, afs_dhashsize
* sizeof(afs_int32
));
3719 afs_blocksUsed
= afs_dcentries
= 0;
3720 afs_stats_cmperf
.cacheBucket0_Discarded
=
3721 afs_stats_cmperf
.cacheBucket1_Discarded
=
3722 afs_stats_cmperf
.cacheBucket2_Discarded
= 0;
3723 hzero(afs_indexCounter
);
3725 afs_freeDCCount
= 0;
3726 afs_freeDCList
= NULLIDX
;
3727 afs_discardDCList
= NULLIDX
;
3728 afs_freeDSList
= afs_Initial_freeDSList
= 0;
3730 LOCK_INIT(&afs_xdcache
, "afs_xdcache");
3736 * Get a dcache ready for writing, respecting the current cache size limits
3738 * len is required because afs_GetDCache with flag == 4 expects the length
3739 * field to be filled. It decides from this whether it's necessary to fetch
3740 * data into the chunk before writing or not (when the whole chunk is
3743 * \param avc The vcache to fetch a dcache for
3744 * \param filePos The start of the section to be written
3745 * \param len The length of the section to be written
3749 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3750 * must be released and afs_PutDCache() called to free dcache.
3753 * \note avc->lock must be held on entry. Function may release and reobtain
3754 * avc->lock and GLOCK.
3758 afs_ObtainDCacheForWriting(struct vcache
*avc
, afs_size_t filePos
,
3759 afs_size_t len
, struct vrequest
*areq
,
3762 struct dcache
*tdc
= NULL
;
3765 /* read the cached info */
3767 tdc
= afs_FindDCache(avc
, filePos
);
3769 ObtainWriteLock(&tdc
->lock
, 657);
3770 } else if (afs_blocksUsed
>
3771 PERCENT(CM_WAITFORDRAINPCT
, afs_cacheBlocks
)) {
3772 tdc
= afs_FindDCache(avc
, filePos
);
3774 ObtainWriteLock(&tdc
->lock
, 658);
3775 if (!afs_IsDCacheFresh(tdc
, avc
)
3776 || (tdc
->dflags
& DFFetching
)) {
3777 ReleaseWriteLock(&tdc
->lock
);
3783 afs_MaybeWakeupTruncateDaemon();
3784 while (afs_blocksUsed
>
3785 PERCENT(CM_WAITFORDRAINPCT
, afs_cacheBlocks
)) {
3786 ReleaseWriteLock(&avc
->lock
);
3787 afs_MaybeWaitForCacheDrain();
3788 afs_MaybeFreeDiscardedDCache();
3789 afs_MaybeWakeupTruncateDaemon();
3790 ObtainWriteLock(&avc
->lock
, 509);
3792 avc
->f
.states
|= CDirty
;
3793 tdc
= afs_GetDCache(avc
, filePos
, areq
, &offset
, &len
, 4);
3795 ObtainWriteLock(&tdc
->lock
, 659);
3798 tdc
= afs_GetDCache(avc
, filePos
, areq
, &offset
, &len
, 4);
3800 ObtainWriteLock(&tdc
->lock
, 660);
3803 if (!(afs_indexFlags
[tdc
->index
] & IFDataMod
)) {
3804 afs_stats_cmperf
.cacheCurrDirtyChunks
++;
3805 afs_indexFlags
[tdc
->index
] |= IFDataMod
; /* so it doesn't disappear */
3807 if (!(tdc
->f
.states
& DWriting
)) {
3808 /* don't mark entry as mod if we don't have to */
3809 tdc
->f
.states
|= DWriting
;
3810 tdc
->dflags
|= DFEntryMod
;
3817 * Make a shadow copy of a dir's dcache. It's used for disconnected
3818 * operations like remove/create/rename to keep the original directory data.
3819 * On reconnection, we can diff the original data with the server and get the
3820 * server changes and with the local data to get the local changes.
3822 * \param avc The dir vnode.
3823 * \param adc The dir dcache.
3825 * \return 0 for success.
3827 * \note The vcache entry must be write locked.
3828 * \note The dcache entry must be read locked.
3831 afs_MakeShadowDir(struct vcache
*avc
, struct dcache
*adc
)
3833 int i
, code
, ret_code
= 0, written
, trans_size
;
3834 struct dcache
*new_dc
= NULL
;
3835 struct osi_file
*tfile_src
= NULL
, *tfile_dst
= NULL
;
3836 struct VenusFid shadow_fid
;
3839 /* Is this a dir? */
3840 if (vType(avc
) != VDIR
)
3843 if (avc
->f
.shadow
.vnode
|| avc
->f
.shadow
.unique
)
3846 /* Generate a fid for the shadow dir. */
3847 shadow_fid
.Cell
= avc
->f
.fid
.Cell
;
3848 shadow_fid
.Fid
.Volume
= avc
->f
.fid
.Fid
.Volume
;
3849 afs_GenShadowFid(&shadow_fid
);
3851 ObtainWriteLock(&afs_xdcache
, 716);
3853 /* Get a fresh dcache. */
3854 (void)afs_AllocDCache(&new_dc
, avc
, 0, 0, &shadow_fid
);
3857 ObtainReadLock(&adc
->mflock
);
3859 /* Set up the new fid. */
3860 /* Copy interesting data from original dir dcache. */
3861 new_dc
->mflags
= adc
->mflags
;
3862 new_dc
->dflags
= adc
->dflags
;
3863 new_dc
->f
.modTime
= adc
->f
.modTime
;
3864 new_dc
->f
.versionNo
= adc
->f
.versionNo
;
3865 new_dc
->f
.states
= adc
->f
.states
;
3866 new_dc
->f
.chunk
= adc
->f
.chunk
;
3867 new_dc
->f
.chunkBytes
= adc
->f
.chunkBytes
;
3869 ReleaseReadLock(&adc
->mflock
);
3871 /* Now add to the two hash chains */
3872 i
= DCHash(&shadow_fid
, 0);
3873 afs_dcnextTbl
[new_dc
->index
] = afs_dchashTbl
[i
];
3874 afs_dchashTbl
[i
] = new_dc
->index
;
3876 i
= DVHash(&shadow_fid
);
3877 afs_dvnextTbl
[new_dc
->index
] = afs_dvhashTbl
[i
];
3878 afs_dvhashTbl
[i
] = new_dc
->index
;
3880 ReleaseWriteLock(&afs_xdcache
);
3882 /* Make sure and flush dir buffers back into the disk cache */
3885 /* Alloc a 4k block. */
3886 data
= afs_osi_Alloc(4096);
3888 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3893 /* Open the files. */
3894 tfile_src
= afs_CFileOpen(&adc
->f
.inode
);
3900 tfile_dst
= afs_CFileOpen(&new_dc
->f
.inode
);
3906 /* And now copy dir dcache data into this dcache,
3910 while (written
< adc
->f
.chunkBytes
) {
3911 trans_size
= adc
->f
.chunkBytes
- written
;
3912 if (trans_size
> 4096)
3915 /* Read a chunk from the dcache. */
3916 code
= afs_CFileRead(tfile_src
, written
, data
, trans_size
);
3917 if (code
< trans_size
) {
3922 /* Write it to the new dcache. */
3923 code
= afs_CFileWrite(tfile_dst
, written
, data
, trans_size
);
3924 if (code
< trans_size
) {
3929 written
+=trans_size
;
3934 afs_CFileClose(tfile_dst
);
3936 afs_CFileClose(tfile_src
);
3939 afs_osi_Free(data
, 4096);
3941 ReleaseWriteLock(&new_dc
->lock
);
3942 afs_PutDCache(new_dc
);
3945 ObtainWriteLock(&afs_xvcache
, 763);
3946 ObtainWriteLock(&afs_disconDirtyLock
, 765);
3947 QAdd(&afs_disconShadow
, &avc
->shadowq
);
3948 osi_Assert((afs_RefVCache(avc
) == 0));
3949 ReleaseWriteLock(&afs_disconDirtyLock
);
3950 ReleaseWriteLock(&afs_xvcache
);
3952 avc
->f
.shadow
.vnode
= shadow_fid
.Fid
.Vnode
;
3953 avc
->f
.shadow
.unique
= shadow_fid
.Fid
.Unique
;
3960 * Delete the dcaches of a shadow dir.
3962 * \param avc The vcache containing the shadow fid.
3964 * \note avc must be write locked.
3967 afs_DeleteShadowDir(struct vcache
*avc
)
3970 struct VenusFid shadow_fid
;
3972 shadow_fid
.Cell
= avc
->f
.fid
.Cell
;
3973 shadow_fid
.Fid
.Volume
= avc
->f
.fid
.Fid
.Volume
;
3974 shadow_fid
.Fid
.Vnode
= avc
->f
.shadow
.vnode
;
3975 shadow_fid
.Fid
.Unique
= avc
->f
.shadow
.unique
;
3977 tdc
= afs_FindDCacheByFid(&shadow_fid
);
3979 afs_HashOutDCache(tdc
, 1);
3980 afs_DiscardDCache(tdc
);
3983 avc
->f
.shadow
.vnode
= avc
->f
.shadow
.unique
= 0;
3984 ObtainWriteLock(&afs_disconDirtyLock
, 708);
3985 QRemove(&avc
->shadowq
);
3986 ReleaseWriteLock(&afs_disconDirtyLock
);
3987 afs_PutVCache(avc
); /* Because we held it when we added to the queue */
3991 * Populate a dcache with empty chunks up to a given file size,
3992 * used before extending a file in order to avoid 'holes' which
3993 * we can't access in disconnected mode.
3995 * \param avc The vcache which is being extended (locked)
3996 * \param alen The new length of the file
4000 afs_PopulateDCache(struct vcache
*avc
, afs_size_t apos
, struct vrequest
*areq
)
4003 afs_size_t len
, offset
;
4004 afs_int32 start
, end
;
4006 /* We're doing this to deal with the situation where we extend
4007 * by writing after lseek()ing past the end of the file . If that
4008 * extension skips chunks, then those chunks won't be created, and
4009 * GetDCache will assume that they have to be fetched from the server.
4010 * So, for each chunk between the current file position, and the new
4011 * length we GetDCache for that chunk.
4014 if (AFS_CHUNK(apos
) == 0 || apos
<= avc
->f
.m
.Length
)
4017 if (avc
->f
.m
.Length
== 0)
4020 start
= AFS_CHUNK(avc
->f
.m
.Length
)+1;
4022 end
= AFS_CHUNK(apos
);
4025 len
= AFS_CHUNKTOSIZE(start
);
4026 tdc
= afs_GetDCache(avc
, AFS_CHUNKTOBASE(start
), areq
, &offset
, &len
, 4);