2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
13 #include <afsconfig.h>
14 #include "afs/param.h"
19 #include "afs/sysincludes.h" /*Standard vendor system headers */
20 #include "afsincludes.h" /*AFS-based standard headers */
21 #include "afs/afs_stats.h" /* statistics */
22 #include "afs/afs_cbqueue.h"
23 #include "afs/afs_osidnlc.h"
25 /* Forward declarations. */
26 static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
27 static void afs_FreeDiscardedDCache(void);
28 static void afs_DiscardDCache(struct dcache *);
29 static void afs_FreeDCache(struct dcache *);
31 static afs_int32 afs_DCGetBucket(struct vcache *);
32 static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
33 static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
34 static void afs_DCSizeInit(void);
35 static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);
39 * --------------------- Exported definitions ---------------------
42 afs_int32 afs_blocksUsed_0; /*1K blocks in cache - in theory is zero */
43 afs_int32 afs_blocksUsed_1; /*1K blocks in cache */
44 afs_int32 afs_blocksUsed_2; /*1K blocks in cache */
45 afs_int32 afs_pct1 = -1;
46 afs_int32 afs_pct2 = -1;
47 afs_uint32 afs_tpct1 = 0;
48 afs_uint32 afs_tpct2 = 0;
49 afs_uint32 splitdcache = 0;
51 afs_lock_t afs_xdcache; /*Lock: alloc new disk cache entries */
52 afs_int32 afs_freeDCList; /*Free list for disk cache entries */
53 afs_int32 afs_freeDCCount; /*Count of elts in freeDCList */
54 afs_int32 afs_discardDCList; /*Discarded disk cache entries */
55 afs_int32 afs_discardDCCount; /*Count of elts in discardDCList */
56 struct dcache *afs_freeDSList; /*Free list for disk slots */
57 struct dcache *afs_Initial_freeDSList; /*Initial list for above */
58 ino_t cacheInode; /*Inode for CacheItems file */
59 struct osi_file *afs_cacheInodep = 0; /* file for CacheItems inode */
60 struct afs_q afs_DLRU; /*dcache LRU */
61 afs_int32 afs_dhashsize = 1024;
62 afs_int32 *afs_dvhashTbl; /*Data cache hash table */
63 afs_int32 *afs_dchashTbl; /*Data cache hash table */
64 afs_int32 *afs_dvnextTbl; /*Dcache hash table links */
65 afs_int32 *afs_dcnextTbl; /*Dcache hash table links */
66 struct dcache **afs_indexTable; /*Pointers to dcache entries */
67 afs_hyper_t *afs_indexTimes; /*Dcache entry Access times */
68 afs_int32 *afs_indexUnique; /*dcache entry Fid.Unique */
69 unsigned char *afs_indexFlags; /*(only one) Is there data there? */
70 afs_hyper_t afs_indexCounter; /*Fake time for marking index
72 afs_int32 afs_cacheFiles = 0; /*Size of afs_indexTable */
73 afs_int32 afs_cacheBlocks; /*1K blocks in cache */
74 afs_int32 afs_cacheStats; /*Stat entries in cache */
75 afs_int32 afs_blocksUsed; /*Number of blocks in use */
76 afs_int32 afs_blocksDiscarded; /*Blocks freed but not truncated */
77 afs_int32 afs_fsfragsize = 1023; /*Underlying Filesystem minimum unit
78 *of disk allocation usually 1K
79 *this value is (truefrag -1 ) to
80 *save a bunch of subtracts... */
81 #ifdef AFS_64BIT_CLIENT
82 #ifdef AFS_VM_RDWR_ENV
83 afs_size_t afs_vmMappingEnd; /* for large files (>= 2GB) the VM
84 * mapping an 32bit addressing machines
85 * can only be used below the 2 GB
86 * line. From this point upwards we
87 * must do direct I/O into the cache
88 * files. The value should be on a
90 #endif /* AFS_VM_RDWR_ENV */
91 #endif /* AFS_64BIT_CLIENT */
93 /* The following is used to ensure that new dcache's aren't obtained when
94 * the cache is nearly full.
96 int afs_WaitForCacheDrain = 0;
97 int afs_TruncateDaemonRunning = 0;
98 int afs_CacheTooFull = 0;
100 afs_int32 afs_dcentries; /* In-memory dcache entries */
103 int dcacheDisabled = 0;
105 static int afs_UFSCacheFetchProc(), afs_UFSCacheStoreProc();
106 struct afs_cacheOps afs_UfsCacheOps = {
114 afs_UFSCacheFetchProc,
115 afs_UFSCacheStoreProc,
121 struct afs_cacheOps afs_MemCacheOps = {
123 afs_MemCacheTruncate,
129 afs_MemCacheFetchProc,
130 afs_MemCacheStoreProc,
136 int cacheDiskType; /*Type of backing disk for cache */
137 struct afs_cacheOps *afs_cacheType;
140 afs_DCGetBucket(struct vcache *avc)
145 /* This should be replaced with some sort of user configurable function */
146 if (avc->states & CRO) {
148 } else if (avc->states & CBackup) {
158 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
160 afs_int32 adjustSize = newSize - oldSize;
168 afs_blocksUsed_0 += adjustSize;
169 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
172 afs_blocksUsed_1 += adjustSize;
173 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
176 afs_blocksUsed_2 += adjustSize;
177 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
185 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
193 afs_blocksUsed_0 -= size;
196 afs_blocksUsed_1 -= size;
199 afs_blocksUsed_2 -= size;
203 adc->bucket = newBucket;
208 afs_blocksUsed_0 += size;
211 afs_blocksUsed_1 += size;
214 afs_blocksUsed_2 += size;
224 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
228 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
233 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
234 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
236 /* Short cut: if we don't know about it, try to kill it */
237 if (phase < 2 && afs_blocksUsed_0)
240 if (afs_pct1 > afs_tpct1)
242 if (afs_pct2 > afs_tpct2)
244 return 0; /* unlikely */
252 * Warn about failing to store a file.
255 * acode : Associated error code.
256 * avolume : Volume involved.
257 * aflags : How to handle the output:
258 * aflags & 1: Print out on console
259 * aflags & 2: Print out on controlling tty
262 * Call this from close call when vnodeops is RCS unlocked.
266 afs_StoreWarn(register afs_int32 acode, afs_int32 avolume,
267 register afs_int32 aflags)
269 static char problem_fmt[] =
270 "afs: failed to store file in volume %d (%s)\n";
271 static char problem_fmt_w_error[] =
272 "afs: failed to store file in volume %d (error %d)\n";
273 static char netproblems[] = "network problems";
274 static char partfull[] = "partition full";
275 static char overquota[] = "over quota";
277 AFS_STATCNT(afs_StoreWarn);
283 afs_warn(problem_fmt, avolume, netproblems);
285 afs_warnuser(problem_fmt, avolume, netproblems);
286 } else if (acode == ENOSPC) {
291 afs_warn(problem_fmt, avolume, partfull);
293 afs_warnuser(problem_fmt, avolume, partfull);
296 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
297 * Instead ENOSPC will be sent...
299 if (acode == EDQUOT) {
304 afs_warn(problem_fmt, avolume, overquota);
306 afs_warnuser(problem_fmt, avolume, overquota);
314 afs_warn(problem_fmt_w_error, avolume, acode);
316 afs_warnuser(problem_fmt_w_error, avolume, acode);
321 afs_MaybeWakeupTruncateDaemon(void)
323 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
324 afs_CacheTooFull = 1;
325 if (!afs_TruncateDaemonRunning)
326 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
327 } else if (!afs_TruncateDaemonRunning
328 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
329 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
333 /* Keep statistics on run time for afs_CacheTruncateDaemon. This is a
334 * struct so we need only export one symbol for AIX.
336 static struct CTD_stats {
337 osi_timeval_t CTD_beforeSleep;
338 osi_timeval_t CTD_afterSleep;
339 osi_timeval_t CTD_sleepTime;
340 osi_timeval_t CTD_runTime;
344 u_int afs_min_cache = 0;
346 afs_CacheTruncateDaemon(void)
348 osi_timeval_t CTD_tmpTime;
352 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
354 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
356 osi_GetuTime(&CTD_stats.CTD_afterSleep);
357 afs_TruncateDaemonRunning = 1;
359 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
360 MObtainWriteLock(&afs_xdcache, 266);
361 if (afs_CacheTooFull) {
362 int space_needed, slots_needed;
363 /* if we get woken up, we should try to clean something out */
364 for (counter = 0; counter < 10; counter++) {
366 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
368 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
369 afs_GetDownD(slots_needed, &space_needed, 0);
370 if ((space_needed <= 0) && (slots_needed <= 0)) {
373 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
376 if (!afs_CacheIsTooFull())
377 afs_CacheTooFull = 0;
379 MReleaseWriteLock(&afs_xdcache);
382 * This is a defensive check to try to avoid starving threads
383 * that may need the global lock so thay can help free some
384 * cache space. If this thread won't be sleeping or truncating
385 * any cache files then give up the global lock so other
386 * threads get a chance to run.
388 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
389 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
390 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
394 * This is where we free the discarded cache elements.
396 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
397 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
398 afs_FreeDiscardedDCache();
401 /* See if we need to continue to run. Someone may have
402 * signalled us while we were executing.
404 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
405 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
406 /* Collect statistics on truncate daemon. */
407 CTD_stats.CTD_nSleeps++;
408 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
409 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
410 CTD_stats.CTD_beforeSleep);
411 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
413 afs_TruncateDaemonRunning = 0;
414 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
415 afs_TruncateDaemonRunning = 1;
417 osi_GetuTime(&CTD_stats.CTD_afterSleep);
418 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
419 CTD_stats.CTD_afterSleep);
420 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
422 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
424 afs_termState = AFSOP_STOP_AFSDB;
426 afs_termState = AFSOP_STOP_RXEVENT;
428 afs_osi_Wakeup(&afs_termState);
439 * Make adjustment for the new size in the disk cache entry
441 * Major Assumptions Here:
442 * Assumes that frag size is an integral power of two, less one,
443 * and that this is a two's complement machine. I don't
444 * know of any filesystems which violate this assumption...
447 * adc : Ptr to dcache entry.
448 * anewsize : New size desired.
452 afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
454 register afs_int32 oldSize;
456 AFS_STATCNT(afs_AdjustSize);
458 adc->dflags |= DFEntryMod;
459 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
460 adc->f.chunkBytes = newSize;
463 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
464 afs_DCAdjustSize(adc, oldSize, newSize);
465 if (newSize > oldSize) {
466 /* We're growing the file, wakeup the daemon */
467 afs_MaybeWakeupTruncateDaemon();
469 afs_blocksUsed += (newSize - oldSize);
470 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
478 * This routine is responsible for moving at least one entry (but up
479 * to some number of them) from the LRU queue to the free queue.
482 * anumber : Number of entries that should ideally be moved.
483 * aneedSpace : How much space we need (1K blocks);
486 * The anumber parameter is just a hint; at least one entry MUST be
487 * moved, or we'll panic. We must be called with afs_xdcache
488 * write-locked. We should try to satisfy both anumber and aneedspace,
489 * whichever is more demanding - need to do several things:
490 * 1. only grab up to anumber victims if aneedSpace <= 0, not
491 * the whole set of MAXATONCE.
492 * 2. dynamically choose MAXATONCE to reflect severity of
493 * demand: something like (*aneedSpace >> (logChunk - 9))
494 * N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
495 * indicates that the cache is not properly configured/tuned or
496 * something. We should be able to automatically correct that problem.
499 #define MAXATONCE 16 /* max we can obtain at once */
501 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
505 struct VenusFid *afid;
509 register struct vcache *tvc;
510 afs_uint32 victims[MAXATONCE];
511 struct dcache *victimDCs[MAXATONCE];
512 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
513 afs_uint32 victimPtr; /* next free item in victim arrays */
514 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
515 afs_uint32 maxVictimPtr; /* where it is */
520 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
521 vfslocked = VFS_LOCK_GIANT(afs_globalVFS);
524 AFS_STATCNT(afs_GetDownD);
526 if (CheckLock(&afs_xdcache) != -1)
527 osi_Panic("getdownd nolock");
528 /* decrement anumber first for all dudes in free list */
529 /* SHOULD always decrement anumber first, even if aneedSpace >0,
530 * because we should try to free space even if anumber <=0 */
531 if (!aneedSpace || *aneedSpace <= 0) {
532 anumber -= afs_freeDCCount;
534 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
535 VFS_UNLOCK_GIANT(vfslocked);
537 return; /* enough already free */
541 /* bounds check parameter */
542 if (anumber > MAXATONCE)
543 anumber = MAXATONCE; /* all we can do */
545 /* rewrite so phases include a better eligiblity for gc test*/
547 * The phase variable manages reclaims. Set to 0, the first pass,
548 * we don't reclaim active entries, or other than target bucket.
549 * Set to 1, we reclaim even active ones in target bucket.
550 * Set to 2, we reclaim any inactive one.
551 * Set to 3, we reclaim even active ones.
559 for (i = 0; i < afs_cacheFiles; i++)
560 /* turn off all flags */
561 afs_indexFlags[i] &= ~IFFlag;
563 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
564 /* find oldest entries for reclamation */
565 maxVictimPtr = victimPtr = 0;
566 hzero(maxVictimTime);
567 curbucket = afs_DCWhichBucket(phase, buckethint);
568 /* select victims from access time array */
569 for (i = 0; i < afs_cacheFiles; i++) {
570 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
571 /* skip if dirty or already free */
574 tdc = afs_indexTable[i];
575 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
577 /* Wrong bucket; can't use it! */
580 if (tdc && (tdc->refCount != 0)) {
581 /* Referenced; can't use it! */
584 hset(vtime, afs_indexTimes[i]);
586 /* if we've already looked at this one, skip it */
587 if (afs_indexFlags[i] & IFFlag)
590 if (victimPtr < MAXATONCE) {
591 /* if there's at least one free victim slot left */
592 victims[victimPtr] = i;
593 hset(victimTimes[victimPtr], vtime);
594 if (hcmp(vtime, maxVictimTime) > 0) {
595 hset(maxVictimTime, vtime);
596 maxVictimPtr = victimPtr;
599 } else if (hcmp(vtime, maxVictimTime) < 0) {
601 * We're older than youngest victim, so we replace at
604 /* find youngest (largest LRU) victim */
607 osi_Panic("getdownd local");
609 hset(victimTimes[j], vtime);
610 /* recompute maxVictimTime */
611 hset(maxVictimTime, vtime);
612 for (j = 0; j < victimPtr; j++)
613 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
614 hset(maxVictimTime, victimTimes[j]);
620 /* now really reclaim the victims */
621 j = 0; /* flag to track if we actually got any of the victims */
622 /* first, hold all the victims, since we're going to release the lock
623 * during the truncate operation.
625 for (i = 0; i < victimPtr; i++) {
626 tdc = afs_GetDSlot(victims[i], 0);
627 /* We got tdc->tlock(R) here */
628 if (tdc->refCount == 1)
632 ReleaseReadLock(&tdc->tlock);
636 for (i = 0; i < victimPtr; i++) {
637 /* q is first elt in dcache entry */
639 /* now, since we're dropping the afs_xdcache lock below, we
640 * have to verify, before proceeding, that there are no other
641 * references to this dcache entry, even now. Note that we
642 * compare with 1, since we bumped it above when we called
643 * afs_GetDSlot to preserve the entry's identity.
645 if (tdc && tdc->refCount == 1) {
646 unsigned char chunkFlags;
647 afs_size_t tchunkoffset = 0;
649 /* xdcache is lower than the xvcache lock */
650 MReleaseWriteLock(&afs_xdcache);
651 MObtainReadLock(&afs_xvcache);
652 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
653 MReleaseReadLock(&afs_xvcache);
654 MObtainWriteLock(&afs_xdcache, 527);
656 if (tdc->refCount > 1)
659 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
660 chunkFlags = afs_indexFlags[tdc->index];
661 if (((phase & 1) == 0) && osi_Active(tvc))
663 if (((phase & 1) == 1) && osi_Active(tvc)
664 && (tvc->states & CDCLock)
665 && (chunkFlags & IFAnyPages))
667 if (chunkFlags & IFDataMod)
669 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
670 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
671 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
672 ICL_HANDLE_OFFSET(tchunkoffset));
674 #if defined(AFS_SUN5_ENV)
676 * Now we try to invalidate pages. We do this only for
677 * Solaris. For other platforms, it's OK to recycle a
678 * dcache entry out from under a page, because the strategy
679 * function can call afs_GetDCache().
681 if (!skip && (chunkFlags & IFAnyPages)) {
684 MReleaseWriteLock(&afs_xdcache);
685 MObtainWriteLock(&tvc->vlock, 543);
686 if (tvc->multiPage) {
690 /* block locking pages */
691 tvc->vstates |= VPageCleaning;
692 /* block getting new pages */
694 MReleaseWriteLock(&tvc->vlock);
695 /* One last recheck */
696 MObtainWriteLock(&afs_xdcache, 333);
697 chunkFlags = afs_indexFlags[tdc->index];
698 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
699 || (osi_Active(tvc) && (tvc->states & CDCLock)
700 && (chunkFlags & IFAnyPages))) {
702 MReleaseWriteLock(&afs_xdcache);
705 MReleaseWriteLock(&afs_xdcache);
707 code = osi_VM_GetDownD(tvc, tdc);
709 MObtainWriteLock(&afs_xdcache, 269);
710 /* we actually removed all pages, clean and dirty */
712 afs_indexFlags[tdc->index] &=
713 ~(IFDirtyPages | IFAnyPages);
716 MReleaseWriteLock(&afs_xdcache);
718 MObtainWriteLock(&tvc->vlock, 544);
719 if (--tvc->activeV == 0
720 && (tvc->vstates & VRevokeWait)) {
721 tvc->vstates &= ~VRevokeWait;
722 afs_osi_Wakeup((char *)&tvc->vstates);
725 if (tvc->vstates & VPageCleaning) {
726 tvc->vstates &= ~VPageCleaning;
727 afs_osi_Wakeup((char *)&tvc->vstates);
730 MReleaseWriteLock(&tvc->vlock);
732 #endif /* AFS_SUN5_ENV */
734 MReleaseWriteLock(&afs_xdcache);
737 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
738 MObtainWriteLock(&afs_xdcache, 528);
739 if (afs_indexFlags[tdc->index] &
740 (IFDataMod | IFDirtyPages | IFAnyPages))
742 if (tdc->refCount > 1)
745 #if defined(AFS_SUN5_ENV)
747 /* no vnode, so IFDirtyPages is spurious (we don't
748 * sweep dcaches on vnode recycling, so we can have
749 * DIRTYPAGES set even when all pages are gone). Just
751 * Hold vcache lock to prevent vnode from being
752 * created while we're clearing IFDirtyPages.
754 afs_indexFlags[tdc->index] &=
755 ~(IFDirtyPages | IFAnyPages);
759 /* skip this guy and mark him as recently used */
760 afs_indexFlags[tdc->index] |= IFFlag;
761 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
762 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
763 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
764 ICL_HANDLE_OFFSET(tchunkoffset));
766 /* flush this dude from the data cache and reclaim;
767 * first, make sure no one will care that we damage
768 * it, by removing it from all hash tables. Then,
769 * melt it down for parts. Note that any concurrent
770 * (new possibility!) calls to GetDownD won't touch
771 * this guy because his reference count is > 0. */
772 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
773 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
774 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
775 ICL_HANDLE_OFFSET(tchunkoffset));
776 AFS_STATCNT(afs_gget);
777 afs_HashOutDCache(tdc);
778 if (tdc->f.chunkBytes != 0) {
782 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
787 afs_DiscardDCache(tdc);
792 j = 1; /* we reclaimed at least one victim */
799 /* Phase is 0 and no one was found, so try phase 1 (ignore
800 * osi_Active flag) */
803 for (i = 0; i < afs_cacheFiles; i++)
804 /* turn off all flags */
805 afs_indexFlags[i] &= ~IFFlag;
808 /* found no one in phases 0-5, we're hosed */
812 } /* big while loop */
814 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
815 VFS_UNLOCK_GIANT(vfslocked);
824 * Description: remove adc from any hash tables that would allow it to be located
825 * again by afs_FindDCache or afs_GetDCache.
827 * Parameters: adc -- pointer to dcache entry to remove from hash tables.
829 * Locks: Must have the afs_xdcache lock write-locked to call this function.
832 afs_HashOutDCache(struct dcache *adc)
836 AFS_STATCNT(afs_glink);
837 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
839 /* if this guy is in the hash table, pull him out */
840 if (adc->f.fid.Fid.Volume != 0) {
841 /* remove entry from first hash chains */
842 i = DCHash(&adc->f.fid, adc->f.chunk);
843 us = afs_dchashTbl[i];
844 if (us == adc->index) {
845 /* first dude in the list */
846 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
848 /* somewhere on the chain */
849 while (us != NULLIDX) {
850 if (afs_dcnextTbl[us] == adc->index) {
851 /* found item pointing at the one to delete */
852 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
855 us = afs_dcnextTbl[us];
858 osi_Panic("dcache hc");
860 /* remove entry from *other* hash chain */
861 i = DVHash(&adc->f.fid);
862 us = afs_dvhashTbl[i];
863 if (us == adc->index) {
864 /* first dude in the list */
865 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
867 /* somewhere on the chain */
868 while (us != NULLIDX) {
869 if (afs_dvnextTbl[us] == adc->index) {
870 /* found item pointing at the one to delete */
871 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
874 us = afs_dvnextTbl[us];
877 osi_Panic("dcache hv");
881 /* prevent entry from being found on a reboot (it is already out of
882 * the hash table, but after a crash, we just look at fid fields of
883 * stable (old) entries).
885 adc->f.fid.Fid.Volume = 0; /* invalid */
887 /* mark entry as modified */
888 adc->dflags |= DFEntryMod;
892 } /*afs_HashOutDCache */
899 * Flush the given dcache entry, pulling it from hash chains
900 * and truncating the associated cache file.
903 * adc: Ptr to dcache entry to flush.
906 * This routine must be called with the afs_xdcache lock held
911 afs_FlushDCache(register struct dcache *adc)
913 AFS_STATCNT(afs_FlushDCache);
915 * Bump the number of cache files flushed.
917 afs_stats_cmperf.cacheFlushes++;
919 /* remove from all hash tables */
920 afs_HashOutDCache(adc);
922 /* Free its space; special case null operation, since truncate operation
923 * in UFS is slow even in this case, and this allows us to pre-truncate
924 * these files at more convenient times with fewer locks set
925 * (see afs_GetDownD).
927 if (adc->f.chunkBytes != 0) {
928 afs_DiscardDCache(adc);
929 afs_MaybeWakeupTruncateDaemon();
934 if (afs_WaitForCacheDrain) {
935 if (afs_blocksUsed <=
936 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
937 afs_WaitForCacheDrain = 0;
938 afs_osi_Wakeup(&afs_WaitForCacheDrain);
941 } /*afs_FlushDCache */
947 * Description: put a dcache entry on the free dcache entry list.
949 * Parameters: adc -- dcache entry to free
951 * Environment: called with afs_xdcache lock write-locked.
954 afs_FreeDCache(register struct dcache *adc)
956 /* Thread on free list, update free list count and mark entry as
957 * freed in its indexFlags element. Also, ensure DCache entry gets
958 * written out (set DFEntryMod).
961 afs_dvnextTbl[adc->index] = afs_freeDCList;
962 afs_freeDCList = adc->index;
964 afs_indexFlags[adc->index] |= IFFree;
965 adc->dflags |= DFEntryMod;
967 if (afs_WaitForCacheDrain) {
968 if ((afs_blocksUsed - afs_blocksDiscarded) <=
969 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
970 afs_WaitForCacheDrain = 0;
971 afs_osi_Wakeup(&afs_WaitForCacheDrain);
980 * Discard the cache element by moving it to the discardDCList.
981 * This puts the cache element into a quasi-freed state, where
982 * the space may be reused, but the file has not been truncated.
984 * Major Assumptions Here:
985 * Assumes that frag size is an integral power of two, less one,
986 * and that this is a two's complement machine. I don't
987 * know of any filesystems which violate this assumption...
990 * adc : Ptr to dcache entry.
993 * Must be called with afs_xdcache write-locked.
997 afs_DiscardDCache(register struct dcache *adc)
999 register afs_int32 size;
1001 AFS_STATCNT(afs_DiscardDCache);
1003 osi_Assert(adc->refCount == 1);
1005 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1006 afs_blocksDiscarded += size;
1007 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1009 afs_dvnextTbl[adc->index] = afs_discardDCList;
1010 afs_discardDCList = adc->index;
1011 afs_discardDCCount++;
1013 adc->f.fid.Fid.Volume = 0;
1014 adc->dflags |= DFEntryMod;
1015 afs_indexFlags[adc->index] |= IFDiscarded;
1017 if (afs_WaitForCacheDrain) {
1018 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1019 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1020 afs_WaitForCacheDrain = 0;
1021 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1025 } /*afs_DiscardDCache */
1028 * afs_FreeDiscardedDCache
1031 * Free the next element on the list of discarded cache elements.
1034 afs_FreeDiscardedDCache(void)
1036 register struct dcache *tdc;
1037 register struct osi_file *tfile;
1038 register afs_int32 size;
1040 AFS_STATCNT(afs_FreeDiscardedDCache);
1042 MObtainWriteLock(&afs_xdcache, 510);
1043 if (!afs_blocksDiscarded) {
1044 MReleaseWriteLock(&afs_xdcache);
1049 * Get an entry from the list of discarded cache elements
1051 tdc = afs_GetDSlot(afs_discardDCList, 0);
1052 osi_Assert(tdc->refCount == 1);
1053 ReleaseReadLock(&tdc->tlock);
1055 afs_discardDCList = afs_dvnextTbl[tdc->index];
1056 afs_dvnextTbl[tdc->index] = NULLIDX;
1057 afs_discardDCCount--;
1058 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1059 afs_blocksDiscarded -= size;
1060 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1061 /* We can lock because we just took it off the free list */
1062 ObtainWriteLock(&tdc->lock, 626);
1063 MReleaseWriteLock(&afs_xdcache);
1066 * Truncate the element to reclaim its space
1068 tfile = afs_CFileOpen(tdc->f.inode);
1069 afs_CFileTruncate(tfile, 0);
1070 afs_CFileClose(tfile);
1071 afs_AdjustSize(tdc, 0);
1072 afs_DCMoveBucket(tdc, 0, 0);
1075 * Free the element we just truncated
1077 MObtainWriteLock(&afs_xdcache, 511);
1078 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1079 afs_FreeDCache(tdc);
1080 tdc->f.states &= ~(DRO|DBackup|DRW);
1081 ReleaseWriteLock(&tdc->lock);
1083 MReleaseWriteLock(&afs_xdcache);
1087 * afs_MaybeFreeDiscardedDCache
1090 * Free as many entries from the list of discarded cache elements
1091 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1097 afs_MaybeFreeDiscardedDCache(void)
1100 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1102 while (afs_blocksDiscarded
1103 && (afs_blocksUsed >
1104 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1105 afs_FreeDiscardedDCache();
1114 * Try to free up a certain number of disk slots.
1117 * anumber : Targeted number of disk slots to free up.
1120 * Must be called with afs_xdcache write-locked.
1123 afs_GetDownDSlot(int anumber)
1125 struct afs_q *tq, *nq;
1130 AFS_STATCNT(afs_GetDownDSlot);
1131 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1132 osi_Panic("diskless getdowndslot");
1134 if (CheckLock(&afs_xdcache) != -1)
1135 osi_Panic("getdowndslot nolock");
1137 /* decrement anumber first for all dudes in free list */
1138 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1141 return; /* enough already free */
1143 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1145 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1146 nq = QPrev(tq); /* in case we remove it */
1147 if (tdc->refCount == 0) {
1148 if ((ix = tdc->index) == NULLIDX)
1149 osi_Panic("getdowndslot");
1150 /* pull the entry out of the lruq and put it on the free list */
1151 QRemove(&tdc->lruq);
1153 /* write-through if modified */
1154 if (tdc->dflags & DFEntryMod) {
1155 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1157 * ask proxy to do this for us - we don't have the stack space
1159 while (tdc->dflags & DFEntryMod) {
1162 s = SPLOCK(afs_sgibklock);
1163 if (afs_sgibklist == NULL) {
1164 /* if slot is free, grab it. */
1165 afs_sgibklist = tdc;
1166 SV_SIGNAL(&afs_sgibksync);
1168 /* wait for daemon to (start, then) finish. */
1169 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1173 tdc->dflags &= ~DFEntryMod;
1174 afs_WriteDCache(tdc, 1);
1178 /* finally put the entry in the free list */
1179 afs_indexTable[ix] = NULL;
1180 afs_indexFlags[ix] &= ~IFEverUsed;
1181 tdc->index = NULLIDX;
1182 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1183 afs_freeDSList = tdc;
1187 } /*afs_GetDownDSlot */
1194 * Increment the reference count on a disk cache entry,
1195 * which already has a non-zero refcount. In order to
1196 * increment the refcount of a zero-reference entry, you
1197 * have to hold afs_xdcache.
1200 * adc : Pointer to the dcache entry to increment.
1203 * Nothing interesting.
1206 afs_RefDCache(struct dcache *adc)
1208 ObtainWriteLock(&adc->tlock, 627);
1209 if (adc->refCount < 0)
1210 osi_Panic("RefDCache: negative refcount");
1212 ReleaseWriteLock(&adc->tlock);
1221 * Decrement the reference count on a disk cache entry.
1224 * ad : Ptr to the dcache entry to decrement.
1227 * Nothing interesting.
1230 afs_PutDCache(register struct dcache *adc)
1232 AFS_STATCNT(afs_PutDCache);
1233 ObtainWriteLock(&adc->tlock, 276);
1234 if (adc->refCount <= 0)
1235 osi_Panic("putdcache");
1237 ReleaseWriteLock(&adc->tlock);
1246 * Try to discard all data associated with this file from the
1250 * avc : Pointer to the cache info for the file.
1253 * Both pvnLock and lock are write held.
1256 afs_TryToSmush(register struct vcache *avc, struct AFS_UCRED *acred, int sync)
1258 register struct dcache *tdc;
1261 AFS_STATCNT(afs_TryToSmush);
1262 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1263 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length));
1264 sync = 1; /* XX Temp testing XX */
1266 #if defined(AFS_SUN5_ENV)
1267 ObtainWriteLock(&avc->vlock, 573);
1268 avc->activeV++; /* block new getpages */
1269 ReleaseWriteLock(&avc->vlock);
1272 /* Flush VM pages */
1273 osi_VM_TryToSmush(avc, acred, sync);
1276 * Get the hash chain containing all dce's for this fid
1278 i = DVHash(&avc->fid);
1279 MObtainWriteLock(&afs_xdcache, 277);
1280 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1281 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1282 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1283 int releaseTlock = 1;
1284 tdc = afs_GetDSlot(index, NULL);
1285 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1287 if ((afs_indexFlags[index] & IFDataMod) == 0
1288 && tdc->refCount == 1) {
1289 ReleaseReadLock(&tdc->tlock);
1291 afs_FlushDCache(tdc);
1294 afs_indexTable[index] = 0;
1297 ReleaseReadLock(&tdc->tlock);
1301 #if defined(AFS_SUN5_ENV)
1302 ObtainWriteLock(&avc->vlock, 545);
1303 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1304 avc->vstates &= ~VRevokeWait;
1305 afs_osi_Wakeup((char *)&avc->vstates);
1307 ReleaseWriteLock(&avc->vlock);
1309 MReleaseWriteLock(&afs_xdcache);
1311 * It's treated like a callback so that when we do lookups we'll
1312 * invalidate the unique bit if any
1313 * trytoSmush occured during the lookup call
1319 * afs_DCacheMissingChunks
1322 * Given the cached info for a file, return the number of chunks that
1323 * are not available from the dcache.
1326 * avc: Pointer to the (held) vcache entry to look in.
1329 * The number of chunks which are not currently cached.
1332 * The vcache entry is held upon entry.
1336 afs_DCacheMissingChunks(struct vcache *avc)
1339 afs_size_t totalLength;
1340 afs_uint32 totalChunks;
1343 totalLength = avc->m.Length;
1344 if (avc->truncPos < totalLength)
1345 totalLength = avc->truncPos;
1347 totalChunks = AFS_CHUNK(totalLength) + 1;
1349 /*printf("Should have %d chunks for %d bytes\n", totalChunks, totalLength);*/
1351 i = DVHash(&avc->fid);
1352 MObtainWriteLock(&afs_xdcache, 1001);
1353 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1354 i = afs_dvnextTbl[index];
1355 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1356 tdc = afs_GetDSlot(index, NULL);
1357 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1360 ReleaseReadLock(&tdc->tlock);
1364 MReleaseWriteLock(&afs_xdcache);
1366 /*printf("Missing %d chunks\n", totalChunks);*/
1368 return (totalChunks);
1375 * Given the cached info for a file and a byte offset into the
1376 * file, make sure the dcache entry for that file and containing
1377 * the given byte is available, returning it to our caller.
1380 * avc : Pointer to the (held) vcache entry to look in.
1381 * abyte : Which byte we want to get to.
1384 * Pointer to the dcache entry covering the file & desired byte,
1385 * or NULL if not found.
1388 * The vcache entry is held upon entry.
1392 afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1395 register afs_int32 i, index;
1396 register struct dcache *tdc = NULL;
1398 AFS_STATCNT(afs_FindDCache);
1399 chunk = AFS_CHUNK(abyte);
1402 * Hash on the [fid, chunk] and get the corresponding dcache index
1403 * after write-locking the dcache.
1405 i = DCHash(&avc->fid, chunk);
1406 MObtainWriteLock(&afs_xdcache, 278);
1407 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1408 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1409 tdc = afs_GetDSlot(index, NULL);
1410 ReleaseReadLock(&tdc->tlock);
1411 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1412 break; /* leaving refCount high for caller */
1416 index = afs_dcnextTbl[index];
1418 if (index != NULLIDX) {
1419 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1420 hadd32(afs_indexCounter, 1);
1421 MReleaseWriteLock(&afs_xdcache);
1424 MReleaseWriteLock(&afs_xdcache);
1426 } /*afs_FindDCache */
1430 * afs_UFSCacheStoreProc
1433 * Called upon store.
1436 * acall : Ptr to the Rx call structure involved.
1437 * afile : Ptr to the related file descriptor.
1438 * alen : Size of the file in bytes.
1439 * avc : Ptr to the vcache entry.
1440 * shouldWake : is it "safe" to return early from close() ?
1441 * abytesToXferP : Set to the number of bytes to xfer.
1442 * NOTE: This parameter is only used if AFS_NOSTATS
1444 * abytesXferredP : Set to the number of bytes actually xferred.
1445 * NOTE: This parameter is only used if AFS_NOSTATS
1449 * Nothing interesting.
1452 afs_UFSCacheStoreProc(register struct rx_call *acall, struct osi_file *afile,
1453 register afs_int32 alen, struct vcache *avc,
1454 int *shouldWake, afs_size_t * abytesToXferP,
1455 afs_size_t * abytesXferredP)
1457 afs_int32 code, got;
1458 register char *tbuffer;
1461 AFS_STATCNT(UFS_CacheStoreProc);
1465 * In this case, alen is *always* the amount of data we'll be trying
1468 (*abytesToXferP) = alen;
1469 (*abytesXferredP) = 0;
1470 #endif /* AFS_NOSTATS */
1472 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1473 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1474 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1475 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1477 tlen = (alen > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : alen);
1478 got = afs_osi_Read(afile, -1, tbuffer, tlen);
1480 #if defined(KERNEL_HAVE_UERROR)
1481 || (got != tlen && getuerror())
1484 osi_FreeLargeSpace(tbuffer);
1487 afs_Trace2(afs_iclSetp, CM_TRACE_STOREPROC2, ICL_TYPE_OFFSET,
1488 ICL_HANDLE_OFFSET(*tbuffer), ICL_TYPE_INT32, got);
1490 code = rx_Write(acall, tbuffer, got); /* writing 0 bytes will
1491 * push a short packet. Is that really what we want, just because the
1492 * data didn't come back from the disk yet? Let's try it and see. */
1495 (*abytesXferredP) += code;
1496 #endif /* AFS_NOSTATS */
1498 code = rx_Error(acall);
1499 osi_FreeLargeSpace(tbuffer);
1500 return code ? code : -33;
1504 * If file has been locked on server, we can allow the store
1507 if (shouldWake && *shouldWake && (rx_GetRemoteStatus(acall) & 1)) {
1508 *shouldWake = 0; /* only do this once */
1512 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1513 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1514 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1515 osi_FreeLargeSpace(tbuffer);
1518 } /* afs_UFSCacheStoreProc */
1522 * afs_UFSCacheFetchProc
1525 * Routine called on fetch; also tells people waiting for data
1526 * that more has arrived.
1529 * acall : Ptr to the Rx call structure.
1530 * afile : File descriptor for the cache file.
1531 * abase : Base offset to fetch.
1532 * adc : Ptr to the dcache entry for the file, write-locked.
1533 * avc : Ptr to the vcache entry for the file.
1534 * abytesToXferP : Set to the number of bytes to xfer.
1535 * NOTE: This parameter is only used if AFS_NOSTATS
1537 * abytesXferredP : Set to the number of bytes actually xferred.
1538 * NOTE: This parameter is only used if AFS_NOSTATS
1542 * Nothing interesting.
1546 afs_UFSCacheFetchProc(register struct rx_call *acall, struct osi_file *afile,
1547 afs_size_t abase, struct dcache *adc,
1548 struct vcache *avc, afs_size_t * abytesToXferP,
1549 afs_size_t * abytesXferredP, afs_int32 lengthFound)
1552 register afs_int32 code;
1553 register char *tbuffer;
1557 AFS_STATCNT(UFS_CacheFetchProc);
1558 osi_Assert(WriteLocked(&adc->lock));
1559 afile->offset = 0; /* Each time start from the beginning */
1560 length = lengthFound;
1562 (*abytesToXferP) = 0;
1563 (*abytesXferredP) = 0;
1564 #endif /* AFS_NOSTATS */
1565 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1566 adc->validPos = abase;
1570 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
1572 length = ntohl(length);
1573 if (code != sizeof(afs_int32)) {
1574 osi_FreeLargeSpace(tbuffer);
1575 code = rx_Error(acall);
1576 return (code ? code : -1); /* try to return code, not -1 */
1580 * The fetch protocol is extended for the AFS/DFS translator
1581 * to allow multiple blocks of data, each with its own length,
1582 * to be returned. As long as the top bit is set, there are more
1585 * We do not do this for AFS file servers because they sometimes
1586 * return large negative numbers as the transfer size.
1588 if (avc->states & CForeign) {
1589 moredata = length & 0x80000000;
1590 length &= ~0x80000000;
1595 (*abytesToXferP) += length;
1596 #endif /* AFS_NOSTATS */
1597 while (length > 0) {
1598 tlen = (length > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : length);
1599 #ifdef RX_KERNEL_TRACE
1600 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1604 code = rx_Read(acall, tbuffer, tlen);
1606 #ifdef RX_KERNEL_TRACE
1607 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1611 (*abytesXferredP) += code;
1612 #endif /* AFS_NOSTATS */
1614 osi_FreeLargeSpace(tbuffer);
1615 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64READ,
1616 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
1617 ICL_TYPE_INT32, length);
1620 code = afs_osi_Write(afile, -1, tbuffer, tlen);
1622 osi_FreeLargeSpace(tbuffer);
1627 adc->validPos = abase;
1628 if (afs_osi_Wakeup(&adc->validPos) == 0)
1629 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
1630 __FILE__, ICL_TYPE_INT32, __LINE__,
1631 ICL_TYPE_POINTER, adc, ICL_TYPE_INT32,
1635 osi_FreeLargeSpace(tbuffer);
1638 } /* afs_UFSCacheFetchProc */
1644 * This function is called to obtain a reference to data stored in
1645 * the disk cache, locating a chunk of data containing the desired
1646 * byte and returning a reference to the disk cache entry, with its
1647 * reference count incremented.
1651 * avc : Ptr to a vcache entry (unlocked)
1652 * abyte : Byte position in the file desired
1653 * areq : Request structure identifying the requesting user.
1654 * aflags : Settings as follows:
1656 * 2 : Return after creating entry.
1657 * 4 : called from afs_vnop_write.c
1658 * *alen contains length of data to be written.
1660 * aoffset : Set to the offset within the chunk where the resident
1662 * alen : Set to the number of bytes of data after the desired
1663 * byte (including the byte itself) which can be read
1667 * The vcache entry pointed to by avc is unlocked upon entry.
1671 struct AFSVolSync tsync;
1672 struct AFSFetchStatus OutStatus;
1673 struct AFSCallBack CallBack;
1677 * Update the vnode-to-dcache hint if we can get the vnode lock
1678 * right away. Assumes dcache entry is at least read-locked.
1681 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1683 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1684 if (hsame(v->m.DataVersion, d->f.versionNo) && v->callback)
1687 ReleaseWriteLock(&v->lock);
1691 /* avc - Write-locked unless aflags & 1 */
1693 afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1694 register struct vrequest *areq, afs_size_t * aoffset,
1695 afs_size_t * alen, int aflags)
1697 register afs_int32 i, code, code1 = 0, shortcut;
1698 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1699 register afs_int32 adjustsize = 0;
1705 afs_size_t maxGoodLength; /* amount of good data at server */
1706 struct rx_call *tcall;
1707 afs_size_t Position = 0;
1708 #ifdef AFS_64BIT_CLIENT
1710 afs_size_t lengthFound; /* as returned from server */
1711 #endif /* AFS_64BIT_CLIENT */
1712 afs_int32 size, tlen; /* size of segment to transfer */
1713 struct tlocal1 *tsmall = 0;
1714 register struct dcache *tdc;
1715 register struct osi_file *file;
1716 register struct conn *tc;
1718 struct server *newCallback = NULL;
1719 char setNewCallback;
1720 char setVcacheStatus;
1721 char doVcacheUpdate;
1723 int doAdjustSize = 0;
1724 int doReallyAdjustSize = 0;
1725 int overWriteWholeChunk = 0;
1729 struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
1730 osi_timeval_t xferStartTime, /*FS xfer start time */
1731 xferStopTime; /*FS xfer stop time */
1732 afs_size_t bytesToXfer; /* # bytes to xfer */
1733 afs_size_t bytesXferred; /* # bytes actually xferred */
1734 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1735 int fromReplica; /*Are we reading from a replica? */
1736 int numFetchLoops; /*# times around the fetch/analyze loop */
1737 #endif /* AFS_NOSTATS */
1739 AFS_STATCNT(afs_GetDCache);
1744 setLocks = aflags & 1;
1747 * Determine the chunk number and offset within the chunk corresponding
1748 * to the desired byte.
1750 if (avc->fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1753 chunk = AFS_CHUNK(abyte);
1756 /* come back to here if we waited for the cache to drain. */
1759 setNewCallback = setVcacheStatus = 0;
1763 ObtainWriteLock(&avc->lock, 616);
1765 ObtainReadLock(&avc->lock);
1770 * avc->lock(R) if setLocks && !slowPass
1771 * avc->lock(W) if !setLocks || slowPass
1776 /* check hints first! (might could use bcmp or some such...) */
1777 if ((tdc = avc->dchint)) {
1781 * The locking order between afs_xdcache and dcache lock matters.
1782 * The hint dcache entry could be anywhere, even on the free list.
1783 * Locking afs_xdcache ensures that noone is trying to pull dcache
1784 * entries from the free list, and thereby assuming them to be not
1785 * referenced and not locked.
1787 MObtainReadLock(&afs_xdcache);
1788 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1790 if (dcLocked && (tdc->index != NULLIDX)
1791 && !FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk
1792 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1793 /* got the right one. It might not be the right version, and it
1794 * might be fetching, but it's the right dcache entry.
1796 /* All this code should be integrated better with what follows:
1797 * I can save a good bit more time under a write lock if I do..
1799 ObtainWriteLock(&tdc->tlock, 603);
1801 ReleaseWriteLock(&tdc->tlock);
1803 MReleaseReadLock(&afs_xdcache);
1806 if (hsame(tdc->f.versionNo, avc->m.DataVersion)
1807 && !(tdc->dflags & DFFetching)) {
1809 afs_stats_cmperf.dcacheHits++;
1810 MObtainWriteLock(&afs_xdcache, 559);
1811 QRemove(&tdc->lruq);
1812 QAdd(&afs_DLRU, &tdc->lruq);
1813 MReleaseWriteLock(&afs_xdcache);
1816 * avc->lock(R) if setLocks && !slowPass
1817 * avc->lock(W) if !setLocks || slowPass
1824 ReleaseSharedLock(&tdc->lock);
1825 MReleaseReadLock(&afs_xdcache);
1833 * avc->lock(R) if setLocks && !slowPass
1834 * avc->lock(W) if !setLocks || slowPass
1835 * tdc->lock(S) if tdc
1838 if (!tdc) { /* If the hint wasn't the right dcache entry */
1840 * Hash on the [fid, chunk] and get the corresponding dcache index
1841 * after write-locking the dcache.
1846 * avc->lock(R) if setLocks && !slowPass
1847 * avc->lock(W) if !setLocks || slowPass
1850 i = DCHash(&avc->fid, chunk);
1851 /* check to make sure our space is fine */
1852 afs_MaybeWakeupTruncateDaemon();
1854 MObtainWriteLock(&afs_xdcache, 280);
1856 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1857 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1858 tdc = afs_GetDSlot(index, NULL);
1859 ReleaseReadLock(&tdc->tlock);
1862 * avc->lock(R) if setLocks && !slowPass
1863 * avc->lock(W) if !setLocks || slowPass
1866 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1867 /* Move it up in the beginning of the list */
1868 if (afs_dchashTbl[i] != index) {
1869 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1870 afs_dcnextTbl[index] = afs_dchashTbl[i];
1871 afs_dchashTbl[i] = index;
1873 MReleaseWriteLock(&afs_xdcache);
1874 ObtainSharedLock(&tdc->lock, 606);
1875 break; /* leaving refCount high for caller */
1881 index = afs_dcnextTbl[index];
1885 * If we didn't find the entry, we'll create one.
1887 if (index == NULLIDX) {
1888 /* If we're disconnected, we can't do anything */
1889 if (AFS_IS_DISCONNECTED) {
1890 MReleaseWriteLock(&afs_xdcache);
1893 ReleaseWriteLock(&avc->lock);
1895 ReleaseReadLock(&avc->lock);
1901 * avc->lock(R) if setLocks
1902 * avc->lock(W) if !setLocks
1905 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1906 avc, ICL_TYPE_INT32, chunk);
1908 /* Make sure there is a free dcache entry for us to use */
1909 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1912 avc->states |= CDCLock;
1913 /* just need slots */
1914 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1916 avc->states &= ~CDCLock;
1917 if (afs_discardDCList != NULLIDX
1918 || afs_freeDCList != NULLIDX)
1920 /* If we can't get space for 5 mins we give up and panic */
1921 if (++downDCount > 300)
1922 osi_Panic("getdcache");
1923 MReleaseWriteLock(&afs_xdcache);
1926 * avc->lock(R) if setLocks
1927 * avc->lock(W) if !setLocks
1929 afs_osi_Wait(1000, 0, 0);
1934 if (afs_discardDCList == NULLIDX
1935 || ((aflags & 2) && afs_freeDCList != NULLIDX)) {
1937 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1938 tdc = afs_GetDSlot(afs_freeDCList, 0);
1939 osi_Assert(tdc->refCount == 1);
1940 ReleaseReadLock(&tdc->tlock);
1941 ObtainWriteLock(&tdc->lock, 604);
1942 afs_freeDCList = afs_dvnextTbl[tdc->index];
1945 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1946 tdc = afs_GetDSlot(afs_discardDCList, 0);
1947 osi_Assert(tdc->refCount == 1);
1948 ReleaseReadLock(&tdc->tlock);
1949 ObtainWriteLock(&tdc->lock, 605);
1950 afs_discardDCList = afs_dvnextTbl[tdc->index];
1951 afs_discardDCCount--;
1953 ((tdc->f.chunkBytes +
1954 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1955 tdc->f.states &= ~(DRO|DBackup|DRW);
1956 afs_DCMoveBucket(tdc, size, 0);
1957 afs_blocksDiscarded -= size;
1958 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1960 /* Truncate the chunk so zeroes get filled properly */
1961 file = afs_CFileOpen(tdc->f.inode);
1962 afs_CFileTruncate(file, 0);
1963 afs_CFileClose(file);
1964 afs_AdjustSize(tdc, 0);
1970 * avc->lock(R) if setLocks
1971 * avc->lock(W) if !setLocks
1977 * Fill in the newly-allocated dcache record.
1979 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1980 tdc->f.fid = avc->fid;
1981 if (avc->states & CRO)
1982 tdc->f.states = DRO;
1983 else if (avc->states & CBackup)
1984 tdc->f.states = DBackup;
1986 tdc->f.states = DRW;
1987 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1988 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1989 hones(tdc->f.versionNo); /* invalid value */
1990 tdc->f.chunk = chunk;
1991 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1993 if (tdc->lruq.prev == &tdc->lruq)
1994 osi_Panic("lruq 1");
1997 * Now add to the two hash chains - note that i is still set
1998 * from the above DCHash call.
2000 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
2001 afs_dchashTbl[i] = tdc->index;
2002 i = DVHash(&avc->fid);
2003 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
2004 afs_dvhashTbl[i] = tdc->index;
2005 tdc->dflags = DFEntryMod;
2007 afs_MaybeWakeupTruncateDaemon();
2008 MReleaseWriteLock(&afs_xdcache);
2009 ConvertWToSLock(&tdc->lock);
2014 /* vcache->dcache hint failed */
2017 * avc->lock(R) if setLocks && !slowPass
2018 * avc->lock(W) if !setLocks || slowPass
2021 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
2022 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2023 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
2024 hgetlo(avc->m.DataVersion));
2026 * Here we have the entry in tdc, with its refCount incremented.
2027 * Note: we don't use the S-lock on avc; it costs concurrency when
2028 * storing a file back to the server.
2032 * Not a newly created file so we need to check the file's length and
2033 * compare data versions since someone could have changed the data or we're
2034 * reading a file written elsewhere. We only want to bypass doing no-op
2035 * read rpcs on newly created files (dv of 0) since only then we guarantee
2036 * that this chunk's data hasn't been filled by another client.
2038 size = AFS_CHUNKSIZE(abyte);
2039 if (aflags & 4) /* called from write */
2041 else /* called from read */
2042 tlen = tdc->validPos - abyte;
2043 Position = AFS_CHUNKTOBASE(chunk);
2044 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
2045 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
2046 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
2047 ICL_HANDLE_OFFSET(Position));
2048 if ((aflags & 4) && (hiszero(avc->m.DataVersion)))
2050 if ((AFS_CHUNKTOBASE(chunk) >= avc->m.Length) ||
2051 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
2052 overWriteWholeChunk = 1;
2053 if (doAdjustSize || overWriteWholeChunk) {
2054 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
2056 #ifdef AFS_SGI64_ENV
2059 #else /* AFS_SGI64_ENV */
2062 #endif /* AFS_SGI64_ENV */
2063 #else /* AFS_SGI_ENV */
2066 #endif /* AFS_SGI_ENV */
2067 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->m.Length &&
2068 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2069 #if defined(AFS_SUN5_ENV) || defined(AFS_OSF_ENV)
2070 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
2072 if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
2074 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2075 !hsame(avc->m.DataVersion, tdc->f.versionNo))
2076 doReallyAdjustSize = 1;
2078 if (doReallyAdjustSize || overWriteWholeChunk) {
2079 /* no data in file to read at this position */
2080 UpgradeSToWLock(&tdc->lock, 607);
2082 file = afs_CFileOpen(tdc->f.inode);
2083 afs_CFileTruncate(file, 0);
2084 afs_CFileClose(file);
2085 afs_AdjustSize(tdc, 0);
2086 hset(tdc->f.versionNo, avc->m.DataVersion);
2087 tdc->dflags |= DFEntryMod;
2089 ConvertWToSLock(&tdc->lock);
2094 * We must read in the whole chunk if the version number doesn't
2098 /* don't need data, just a unique dcache entry */
2099 ObtainWriteLock(&afs_xdcache, 608);
2100 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2101 hadd32(afs_indexCounter, 1);
2102 ReleaseWriteLock(&afs_xdcache);
2104 updateV2DC(setLocks, avc, tdc, 553);
2105 if (vType(avc) == VDIR)
2108 *aoffset = AFS_CHUNKOFFSET(abyte);
2109 if (tdc->validPos < abyte)
2110 *alen = (afs_size_t) 0;
2112 *alen = tdc->validPos - abyte;
2113 ReleaseSharedLock(&tdc->lock);
2116 ReleaseWriteLock(&avc->lock);
2118 ReleaseReadLock(&avc->lock);
2120 return tdc; /* check if we're done */
2125 * avc->lock(R) if setLocks && !slowPass
2126 * avc->lock(W) if !setLocks || slowPass
2129 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2131 setNewCallback = setVcacheStatus = 0;
2135 * avc->lock(R) if setLocks && !slowPass
2136 * avc->lock(W) if !setLocks || slowPass
2139 if (!hsame(avc->m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2141 * Version number mismatch.
2144 * If we are disconnected, then we can't do much of anything
2145 * because the data doesn't match the file.
2147 if (AFS_IS_DISCONNECTED) {
2148 ReleaseSharedLock(&tdc->lock);
2151 ReleaseWriteLock(&avc->lock);
2153 ReleaseReadLock(&avc->lock);
2155 /* Flush the Dcache */
2160 UpgradeSToWLock(&tdc->lock, 609);
2163 * If data ever existed for this vnode, and this is a text object,
2164 * do some clearing. Now, you'd think you need only do the flush
2165 * when VTEXT is on, but VTEXT is turned off when the text object
2166 * is freed, while pages are left lying around in memory marked
2167 * with this vnode. If we would reactivate (create a new text
2168 * object from) this vnode, we could easily stumble upon some of
2169 * these old pages in pagein. So, we always flush these guys.
2170 * Sun has a wonderful lack of useful invariants in this system.
2172 * avc->flushDV is the data version # of the file at the last text
2173 * flush. Clearly, at least, we don't have to flush the file more
2174 * often than it changes
2176 if (hcmp(avc->flushDV, avc->m.DataVersion) < 0) {
2178 * By here, the cache entry is always write-locked. We can
2179 * deadlock if we call osi_Flush with the cache entry locked...
2180 * Unlock the dcache too.
2182 ReleaseWriteLock(&tdc->lock);
2183 if (setLocks && !slowPass)
2184 ReleaseReadLock(&avc->lock);
2186 ReleaseWriteLock(&avc->lock);
2190 * Call osi_FlushPages in open, read/write, and map, since it
2191 * is too hard here to figure out if we should lock the
2194 if (setLocks && !slowPass)
2195 ObtainReadLock(&avc->lock);
2197 ObtainWriteLock(&avc->lock, 66);
2198 ObtainWriteLock(&tdc->lock, 610);
2203 * avc->lock(R) if setLocks && !slowPass
2204 * avc->lock(W) if !setLocks || slowPass
2208 /* Watch for standard race condition around osi_FlushText */
2209 if (hsame(avc->m.DataVersion, tdc->f.versionNo)) {
2210 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2211 afs_stats_cmperf.dcacheHits++;
2212 ConvertWToSLock(&tdc->lock);
2216 /* Sleep here when cache needs to be drained. */
2217 if (setLocks && !slowPass
2218 && (afs_blocksUsed >
2219 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2220 /* Make sure truncate daemon is running */
2221 afs_MaybeWakeupTruncateDaemon();
2222 ObtainWriteLock(&tdc->tlock, 614);
2223 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2224 ReleaseWriteLock(&tdc->tlock);
2225 ReleaseWriteLock(&tdc->lock);
2226 ReleaseReadLock(&avc->lock);
2227 while ((afs_blocksUsed - afs_blocksDiscarded) >
2228 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2229 afs_WaitForCacheDrain = 1;
2230 afs_osi_Sleep(&afs_WaitForCacheDrain);
2232 afs_MaybeFreeDiscardedDCache();
2233 /* need to check if someone else got the chunk first. */
2234 goto RetryGetDCache;
2237 /* Do not fetch data beyond truncPos. */
2238 maxGoodLength = avc->m.Length;
2239 if (avc->truncPos < maxGoodLength)
2240 maxGoodLength = avc->truncPos;
2241 Position = AFS_CHUNKBASE(abyte);
2242 if (vType(avc) == VDIR) {
2243 size = avc->m.Length;
2244 if (size > tdc->f.chunkBytes) {
2245 /* pre-reserve space for file */
2246 afs_AdjustSize(tdc, size);
2248 size = 999999999; /* max size for transfer */
2250 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2251 /* don't read past end of good data on server */
2252 if (Position + size > maxGoodLength)
2253 size = maxGoodLength - Position;
2255 size = 0; /* Handle random races */
2256 if (size > tdc->f.chunkBytes) {
2257 /* pre-reserve space for file */
2258 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2259 /* max size for transfer still in size */
2262 if (afs_mariner && !tdc->f.chunk)
2263 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2265 * Right now, we only have one tool, and it's a hammer. So, we
2266 * fetch the whole file.
2268 DZap(tdc); /* pages in cache may be old */
2269 file = afs_CFileOpen(tdc->f.inode);
2270 afs_RemoveVCB(&avc->fid);
2271 tdc->f.states |= DWriting;
2272 tdc->dflags |= DFFetching;
2273 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2274 if (tdc->mflags & DFFetchReq) {
2275 tdc->mflags &= ~DFFetchReq;
2276 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2277 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2278 __FILE__, ICL_TYPE_INT32, __LINE__,
2279 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2283 (struct tlocal1 *)osi_AllocLargeSpace(sizeof(struct tlocal1));
2284 setVcacheStatus = 0;
2287 * Remember if we are doing the reading from a replicated volume,
2288 * and how many times we've zipped around the fetch/analyze loop.
2290 fromReplica = (avc->states & CRO) ? 1 : 0;
2292 accP = &(afs_stats_cmfullperf.accessinf);
2294 (accP->replicatedRefs)++;
2296 (accP->unreplicatedRefs)++;
2297 #endif /* AFS_NOSTATS */
2298 /* this is a cache miss */
2299 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2300 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
2301 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2304 afs_stats_cmperf.dcacheMisses++;
2307 * Dynamic root support: fetch data from local memory.
2309 if (afs_IsDynroot(avc)) {
2313 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2315 dynrootDir += Position;
2316 dynrootLen -= Position;
2317 if (size > dynrootLen)
2321 code = afs_CFileWrite(file, 0, dynrootDir, size);
2329 tdc->validPos = Position + size;
2330 afs_CFileTruncate(file, size); /* prune it */
2331 } else if (afs_IsDynrootMount(avc)) {
2335 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2337 dynrootDir += Position;
2338 dynrootLen -= Position;
2339 if (size > dynrootLen)
2343 code = afs_CFileWrite(file, 0, dynrootDir, size);
2351 tdc->validPos = Position + size;
2352 afs_CFileTruncate(file, size); /* prune it */
2355 * Not a dynamic vnode: do the real fetch.
2360 * avc->lock(R) if setLocks && !slowPass
2361 * avc->lock(W) if !setLocks || slowPass
2365 tc = afs_Conn(&avc->fid, areq, SHARED_LOCK);
2367 afs_int32 length_hi, length, bytes;
2371 (accP->numReplicasAccessed)++;
2373 #endif /* AFS_NOSTATS */
2374 if (!setLocks || slowPass) {
2375 avc->callback = tc->srvr->server;
2377 newCallback = tc->srvr->server;
2382 tcall = rx_NewCall(tc->id);
2385 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
2386 #ifdef AFS_64BIT_CLIENT
2387 length_hi = code = 0;
2388 if (!afs_serverHasNo64Bit(tc)) {
2392 StartRXAFS_FetchData64(tcall,
2393 (struct AFSFid *)&avc->fid.
2394 Fid, Position, tsize);
2397 afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
2398 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32,
2402 rx_Read(tcall, (char *)&length_hi,
2405 if (bytes == sizeof(afs_int32)) {
2406 length_hi = ntohl(length_hi);
2409 code = rx_Error(tcall);
2411 code1 = rx_EndCall(tcall, code);
2413 tcall = (struct rx_call *)0;
2417 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
2418 if (Position > 0x7FFFFFFF) {
2425 tcall = rx_NewCall(tc->id);
2427 StartRXAFS_FetchData(tcall, (struct AFSFid *)
2432 afs_serverSetNo64Bit(tc);
2437 rx_Read(tcall, (char *)&length,
2440 if (bytes == sizeof(afs_int32)) {
2441 length = ntohl(length);
2443 code = rx_Error(tcall);
2446 FillInt64(lengthFound, length_hi, length);
2447 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64LENG,
2448 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
2450 ICL_HANDLE_OFFSET(lengthFound));
2451 #else /* AFS_64BIT_CLIENT */
2454 StartRXAFS_FetchData(tcall,
2455 (struct AFSFid *)&avc->fid.Fid,
2461 rx_Read(tcall, (char *)&length,
2464 if (bytes == sizeof(afs_int32)) {
2465 length = ntohl(length);
2467 code = rx_Error(tcall);
2470 #endif /* AFS_64BIT_CLIENT */
2475 &(afs_stats_cmfullperf.rpc.
2476 fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
2477 osi_GetuTime(&xferStartTime);
2480 afs_CacheFetchProc(tcall, file,
2481 (afs_size_t) Position, tdc,
2483 &bytesXferred, length);
2485 osi_GetuTime(&xferStopTime);
2486 (xferP->numXfers)++;
2488 (xferP->numSuccesses)++;
2489 afs_stats_XferSumBytes
2490 [AFS_STATS_FS_XFERIDX_FETCHDATA] +=
2492 (xferP->sumBytes) +=
2493 (afs_stats_XferSumBytes
2494 [AFS_STATS_FS_XFERIDX_FETCHDATA] >> 10);
2495 afs_stats_XferSumBytes
2496 [AFS_STATS_FS_XFERIDX_FETCHDATA] &= 0x3FF;
2497 if (bytesXferred < xferP->minBytes)
2498 xferP->minBytes = bytesXferred;
2499 if (bytesXferred > xferP->maxBytes)
2500 xferP->maxBytes = bytesXferred;
2503 * Tally the size of the object. Note: we tally the actual size,
2504 * NOT the number of bytes that made it out over the wire.
2506 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
2507 (xferP->count[0])++;
2508 else if (bytesToXfer <=
2509 AFS_STATS_MAXBYTES_BUCKET1)
2510 (xferP->count[1])++;
2511 else if (bytesToXfer <=
2512 AFS_STATS_MAXBYTES_BUCKET2)
2513 (xferP->count[2])++;
2514 else if (bytesToXfer <=
2515 AFS_STATS_MAXBYTES_BUCKET3)
2516 (xferP->count[3])++;
2517 else if (bytesToXfer <=
2518 AFS_STATS_MAXBYTES_BUCKET4)
2519 (xferP->count[4])++;
2520 else if (bytesToXfer <=
2521 AFS_STATS_MAXBYTES_BUCKET5)
2522 (xferP->count[5])++;
2523 else if (bytesToXfer <=
2524 AFS_STATS_MAXBYTES_BUCKET6)
2525 (xferP->count[6])++;
2526 else if (bytesToXfer <=
2527 AFS_STATS_MAXBYTES_BUCKET7)
2528 (xferP->count[7])++;
2530 (xferP->count[8])++;
2532 afs_stats_GetDiff(elapsedTime, xferStartTime,
2534 afs_stats_AddTo((xferP->sumTime), elapsedTime);
2535 afs_stats_SquareAddTo((xferP->sqrTime),
2537 if (afs_stats_TimeLessThan
2538 (elapsedTime, (xferP->minTime))) {
2539 afs_stats_TimeAssign((xferP->minTime),
2542 if (afs_stats_TimeGreaterThan
2543 (elapsedTime, (xferP->maxTime))) {
2544 afs_stats_TimeAssign((xferP->maxTime),
2550 afs_CacheFetchProc(tcall, file, Position, tdc,
2552 #endif /* AFS_NOSTATS */
2557 EndRXAFS_FetchData(tcall, &tsmall->OutStatus,
2565 code1 = rx_EndCall(tcall, code);
2574 /* callback could have been broken (or expired) in a race here,
2575 * but we return the data anyway. It's as good as we knew about
2576 * when we started. */
2578 * validPos is updated by CacheFetchProc, and can only be
2579 * modifed under a dcache write lock, which we've blocked out
2581 size = tdc->validPos - Position; /* actual segment size */
2584 afs_CFileTruncate(file, size); /* prune it */
2586 if (!setLocks || slowPass) {
2587 ObtainWriteLock(&afs_xcbhash, 453);
2588 afs_DequeueCallback(avc);
2589 avc->states &= ~(CStatd | CUnique);
2590 avc->callback = NULL;
2591 ReleaseWriteLock(&afs_xcbhash);
2592 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2593 osi_dnlc_purgedp(avc);
2595 /* Something lost. Forget about performance, and go
2596 * back with a vcache write lock.
2598 afs_CFileTruncate(file, 0);
2599 afs_AdjustSize(tdc, 0);
2600 afs_CFileClose(file);
2601 osi_FreeLargeSpace(tsmall);
2603 ReleaseWriteLock(&tdc->lock);
2606 ReleaseReadLock(&avc->lock);
2608 goto RetryGetDCache;
2612 } while (afs_Analyze
2613 (tc, code, &avc->fid, areq,
2614 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2618 * avc->lock(R) if setLocks && !slowPass
2619 * avc->lock(W) if !setLocks || slowPass
2625 * In the case of replicated access, jot down info on the number of
2626 * attempts it took before we got through or gave up.
2629 if (numFetchLoops <= 1)
2630 (accP->refFirstReplicaOK)++;
2631 if (numFetchLoops > accP->maxReplicasPerRef)
2632 accP->maxReplicasPerRef = numFetchLoops;
2634 #endif /* AFS_NOSTATS */
2636 tdc->dflags &= ~DFFetching;
2637 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2638 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2639 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2640 tdc, ICL_TYPE_INT32, tdc->dflags);
2641 if (avc->execsOrWriters == 0)
2642 tdc->f.states &= ~DWriting;
2644 /* now, if code != 0, we have an error and should punt.
2645 * note that we have the vcache write lock, either because
2646 * !setLocks or slowPass.
2649 afs_CFileTruncate(file, 0);
2650 afs_AdjustSize(tdc, 0);
2651 afs_CFileClose(file);
2652 ZapDCE(tdc); /* sets DFEntryMod */
2653 if (vType(avc) == VDIR) {
2656 tdc->f.states &= ~(DRO|DBackup|DRW);
2657 afs_DCMoveBucket(tdc, 0, 0);
2658 ReleaseWriteLock(&tdc->lock);
2660 if (!afs_IsDynroot(avc)) {
2661 ObtainWriteLock(&afs_xcbhash, 454);
2662 afs_DequeueCallback(avc);
2663 avc->states &= ~(CStatd | CUnique);
2664 ReleaseWriteLock(&afs_xcbhash);
2665 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2666 osi_dnlc_purgedp(avc);
2669 * avc->lock(W); assert(!setLocks || slowPass)
2671 osi_Assert(!setLocks || slowPass);
2673 tdc->f.states &= ~(DRO|DBackup|DRW);
2674 afs_DCMoveBucket(tdc, 0, 0);
2679 /* otherwise we copy in the just-fetched info */
2680 afs_CFileClose(file);
2681 afs_AdjustSize(tdc, size); /* new size */
2683 * Copy appropriate fields into vcache. Status is
2684 * copied later where we selectively acquire the
2685 * vcache write lock.
2688 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2690 setVcacheStatus = 1;
2691 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2692 tsmall->OutStatus.DataVersion);
2693 tdc->dflags |= DFEntryMod;
2694 afs_indexFlags[tdc->index] |= IFEverUsed;
2695 ConvertWToSLock(&tdc->lock);
2696 } /*Data version numbers don't match */
2699 * Data version numbers match.
2701 afs_stats_cmperf.dcacheHits++;
2702 } /*Data version numbers match */
2704 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2708 * avc->lock(R) if setLocks && !slowPass
2709 * avc->lock(W) if !setLocks || slowPass
2710 * tdc->lock(S) if tdc
2714 * See if this was a reference to a file in the local cell.
2716 if (afs_IsPrimaryCellNum(avc->fid.Cell))
2717 afs_stats_cmperf.dlocalAccesses++;
2719 afs_stats_cmperf.dremoteAccesses++;
2721 /* Fix up LRU info */
2724 MObtainWriteLock(&afs_xdcache, 602);
2725 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2726 hadd32(afs_indexCounter, 1);
2727 MReleaseWriteLock(&afs_xdcache);
2729 /* return the data */
2730 if (vType(avc) == VDIR)
2733 *aoffset = AFS_CHUNKOFFSET(abyte);
2734 *alen = (tdc->f.chunkBytes - *aoffset);
2735 ReleaseSharedLock(&tdc->lock);
2740 * avc->lock(R) if setLocks && !slowPass
2741 * avc->lock(W) if !setLocks || slowPass
2744 /* Fix up the callback and status values in the vcache */
2746 if (setLocks && !slowPass) {
2749 * This is our dirty little secret to parallel fetches.
2750 * We don't write-lock the vcache while doing the fetch,
2751 * but potentially we'll need to update the vcache after
2752 * the fetch is done.
2754 * Drop the read lock and try to re-obtain the write
2755 * lock. If the vcache still has the same DV, it's
2756 * ok to go ahead and install the new data.
2758 afs_hyper_t currentDV, statusDV;
2760 hset(currentDV, avc->m.DataVersion);
2762 if (setNewCallback && avc->callback != newCallback)
2766 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2767 tsmall->OutStatus.DataVersion);
2769 if (setVcacheStatus && avc->m.Length != tsmall->OutStatus.Length)
2771 if (setVcacheStatus && !hsame(currentDV, statusDV))
2775 ReleaseReadLock(&avc->lock);
2777 if (doVcacheUpdate) {
2778 ObtainWriteLock(&avc->lock, 615);
2779 if (!hsame(avc->m.DataVersion, currentDV)) {
2780 /* We lose. Someone will beat us to it. */
2782 ReleaseWriteLock(&avc->lock);
2787 /* With slow pass, we've already done all the updates */
2789 ReleaseWriteLock(&avc->lock);
2792 /* Check if we need to perform any last-minute fixes with a write-lock */
2793 if (!setLocks || doVcacheUpdate) {
2795 avc->callback = newCallback;
2796 if (tsmall && setVcacheStatus)
2797 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2799 ReleaseWriteLock(&avc->lock);
2803 osi_FreeLargeSpace(tsmall);
2806 } /*afs_GetDCache */
2810 * afs_WriteThroughDSlots
2813 * Sweep through the dcache slots and write out any modified
2814 * in-memory data back on to our caching store.
2820 * The afs_xdcache is write-locked through this whole affair.
2823 afs_WriteThroughDSlots(void)
2825 register struct dcache *tdc;
2826 register afs_int32 i, touchedit = 0;
2828 struct afs_q DirtyQ, *tq;
2830 AFS_STATCNT(afs_WriteThroughDSlots);
2833 * Because of lock ordering, we can't grab dcache locks while
2834 * holding afs_xdcache. So we enter xdcache, get a reference
2835 * for every dcache entry, and exit xdcache.
2837 MObtainWriteLock(&afs_xdcache, 283);
2839 for (i = 0; i < afs_cacheFiles; i++) {
2840 tdc = afs_indexTable[i];
2842 /* Grab tlock in case the existing refcount isn't zero */
2843 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2844 ObtainWriteLock(&tdc->tlock, 623);
2846 ReleaseWriteLock(&tdc->tlock);
2848 QAdd(&DirtyQ, &tdc->dirty);
2851 MReleaseWriteLock(&afs_xdcache);
2854 * Now, for each dcache entry we found, check if it's dirty.
2855 * If so, get write-lock, get afs_xdcache, which protects
2856 * afs_cacheInodep, and flush it. Don't forget to put back
2860 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2862 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2864 if (tdc->dflags & DFEntryMod) {
2867 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2869 /* Now that we have the write lock, double-check */
2870 if (wrLock && (tdc->dflags & DFEntryMod)) {
2871 tdc->dflags &= ~DFEntryMod;
2872 MObtainWriteLock(&afs_xdcache, 620);
2873 afs_WriteDCache(tdc, 1);
2874 MReleaseWriteLock(&afs_xdcache);
2878 ReleaseWriteLock(&tdc->lock);
2884 MObtainWriteLock(&afs_xdcache, 617);
2885 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2886 /* Touch the file to make sure that the mtime on the file is kept
2887 * up-to-date to avoid losing cached files on cold starts because
2888 * their mtime seems old...
2890 struct afs_fheader theader;
2892 theader.magic = AFS_FHMAGIC;
2893 theader.firstCSize = AFS_FIRSTCSIZE;
2894 theader.otherCSize = AFS_OTHERCSIZE;
2895 theader.version = AFS_CI_VERSION;
2896 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2898 MReleaseWriteLock(&afs_xdcache);
2905 * Return a pointer to an freshly initialized dcache entry using
2906 * a memory-based cache. The tlock will be read-locked.
2909 * aslot : Dcache slot to look at.
2910 * tmpdc : Ptr to dcache entry.
2913 * Must be called with afs_xdcache write-locked.
2917 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2919 register struct dcache *tdc;
2922 AFS_STATCNT(afs_MemGetDSlot);
2923 if (CheckLock(&afs_xdcache) != -1)
2924 osi_Panic("getdslot nolock");
2925 if (aslot < 0 || aslot >= afs_cacheFiles)
2926 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2927 tdc = afs_indexTable[aslot];
2929 QRemove(&tdc->lruq); /* move to queue head */
2930 QAdd(&afs_DLRU, &tdc->lruq);
2931 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2932 ObtainWriteLock(&tdc->tlock, 624);
2934 ConvertWToRLock(&tdc->tlock);
2937 if (tmpdc == NULL) {
2938 if (!afs_freeDSList)
2939 afs_GetDownDSlot(4);
2940 if (!afs_freeDSList) {
2941 /* none free, making one is better than a panic */
2942 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2943 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2944 #ifdef KERNEL_HAVE_PIN
2945 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2948 tdc = afs_freeDSList;
2949 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2952 tdc->dflags = 0; /* up-to-date, not in free q */
2954 QAdd(&afs_DLRU, &tdc->lruq);
2955 if (tdc->lruq.prev == &tdc->lruq)
2956 osi_Panic("lruq 3");
2962 /* initialize entry */
2963 tdc->f.fid.Cell = 0;
2964 tdc->f.fid.Fid.Volume = 0;
2966 hones(tdc->f.versionNo);
2967 tdc->f.inode = aslot;
2968 tdc->dflags |= DFEntryMod;
2971 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2974 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2975 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2976 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2979 RWLOCK_INIT(&tdc->lock, "dcache lock");
2980 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2981 RWLOCK_INIT(&tdc->mflock, "dcache flock");
2982 ObtainReadLock(&tdc->tlock);
2985 afs_indexTable[aslot] = tdc;
2988 } /*afs_MemGetDSlot */
2990 unsigned int last_error = 0, lasterrtime = 0;
2996 * Return a pointer to an freshly initialized dcache entry using
2997 * a UFS-based disk cache. The dcache tlock will be read-locked.
3000 * aslot : Dcache slot to look at.
3001 * tmpdc : Ptr to dcache entry.
3004 * afs_xdcache lock write-locked.
3007 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
3009 register afs_int32 code;
3010 register struct dcache *tdc;
3014 AFS_STATCNT(afs_UFSGetDSlot);
3015 if (CheckLock(&afs_xdcache) != -1)
3016 osi_Panic("getdslot nolock");
3017 if (aslot < 0 || aslot >= afs_cacheFiles)
3018 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
3019 tdc = afs_indexTable[aslot];
3021 QRemove(&tdc->lruq); /* move to queue head */
3022 QAdd(&afs_DLRU, &tdc->lruq);
3023 /* Grab tlock in case refCount != 0 */
3024 ObtainWriteLock(&tdc->tlock, 625);
3026 ConvertWToRLock(&tdc->tlock);
3029 /* otherwise we should read it in from the cache file */
3031 * If we weren't passed an in-memory region to place the file info,
3032 * we have to allocate one.
3034 if (tmpdc == NULL) {
3035 if (!afs_freeDSList)
3036 afs_GetDownDSlot(4);
3037 if (!afs_freeDSList) {
3038 /* none free, making one is better than a panic */
3039 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
3040 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
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");
3060 * Seek to the aslot'th entry and read it in.
3063 afs_osi_Read(afs_cacheInodep,
3064 sizeof(struct fcache) * aslot +
3065 sizeof(struct afs_fheader), (char *)(&tdc->f),
3066 sizeof(struct fcache));
3068 if (code != sizeof(struct fcache))
3070 if (!afs_CellNumValid(tdc->f.fid.Cell))
3074 tdc->f.fid.Cell = 0;
3075 tdc->f.fid.Fid.Volume = 0;
3077 hones(tdc->f.versionNo);
3078 tdc->dflags |= DFEntryMod;
3079 #if defined(KERNEL_HAVE_UERROR)
3080 last_error = getuerror();
3082 lasterrtime = osi_Time();
3083 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
3084 tdc->f.states &= ~(DRO|DBackup|DRW);
3085 afs_DCMoveBucket(tdc, 0, 0);
3088 if (tdc->f.states & DRO) {
3089 afs_DCMoveBucket(tdc, 0, 2);
3090 } else if (tdc->f.states & DBackup) {
3091 afs_DCMoveBucket(tdc, 0, 1);
3093 afs_DCMoveBucket(tdc, 0, 1);
3099 if (tdc->f.chunk >= 0)
3100 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
3105 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
3106 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
3107 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
3110 RWLOCK_INIT(&tdc->lock, "dcache lock");
3111 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
3112 RWLOCK_INIT(&tdc->mflock, "dcache flock");
3113 ObtainReadLock(&tdc->tlock);
3116 * If we didn't read into a temporary dcache region, update the
3117 * slot pointer table.
3120 afs_indexTable[aslot] = tdc;
3123 } /*afs_UFSGetDSlot */
3131 * write a particular dcache entry back to its home in the
3135 * adc : Pointer to the dcache entry to write.
3136 * atime : If true, set the modtime on the file to the current time.
3139 * Must be called with the afs_xdcache lock at least read-locked,
3140 * and dcache entry at least read-locked.
3141 * The reference count is not changed.
3145 afs_WriteDCache(register struct dcache *adc, int atime)
3147 register afs_int32 code;
3149 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
3151 AFS_STATCNT(afs_WriteDCache);
3152 osi_Assert(WriteLocked(&afs_xdcache));
3154 adc->f.modTime = osi_Time();
3156 * Seek to the right dcache slot and write the in-memory image out to disk.
3158 afs_cellname_write();
3160 afs_osi_Write(afs_cacheInodep,
3161 sizeof(struct fcache) * adc->index +
3162 sizeof(struct afs_fheader), (char *)(&adc->f),
3163 sizeof(struct fcache));
3164 if (code != sizeof(struct fcache))
3175 * Wake up users of a particular file waiting for stores to take
3179 * avc : Ptr to related vcache entry.
3182 * Nothing interesting.
3186 afs_wakeup(register struct vcache *avc)
3189 register struct brequest *tb;
3191 AFS_STATCNT(afs_wakeup);
3192 for (i = 0; i < NBRS; i++, tb++) {
3193 /* if request is valid and for this file, we've found it */
3194 if (tb->refCount > 0 && avc == tb->vc) {
3197 * If CSafeStore is on, then we don't awaken the guy
3198 * waiting for the store until the whole store has finished.
3199 * Otherwise, we do it now. Note that if CSafeStore is on,
3200 * the BStore routine actually wakes up the user, instead
3202 * I think this is redundant now because this sort of thing
3203 * is already being handled by the higher-level code.
3205 if ((avc->states & CSafeStore) == 0) {
3207 tb->flags |= BUVALID;
3208 if (tb->flags & BUWAIT) {
3209 tb->flags &= ~BUWAIT;
3224 * Given a file name and inode, set up that file to be an
3225 * active member in the AFS cache. This also involves checking
3226 * the usability of its data.
3229 * afile : Name of the cache file to initialize.
3230 * ainode : Inode of the file.
3233 * This function is called only during initialization.
3237 afs_InitCacheFile(char *afile, ino_t ainode)
3239 register afs_int32 code;
3240 #if defined(AFS_LINUX22_ENV)
3241 struct dentry *filevp;
3243 struct vnode *filevp;
3247 struct osi_file *tfile;
3248 struct osi_stat tstat;
3249 register struct dcache *tdc;
3251 AFS_STATCNT(afs_InitCacheFile);
3252 index = afs_stats_cmperf.cacheNumEntries;
3253 if (index >= afs_cacheFiles)
3256 MObtainWriteLock(&afs_xdcache, 282);
3257 tdc = afs_GetDSlot(index, NULL);
3258 ReleaseReadLock(&tdc->tlock);
3259 MReleaseWriteLock(&afs_xdcache);
3261 ObtainWriteLock(&tdc->lock, 621);
3262 MObtainWriteLock(&afs_xdcache, 622);
3264 code = gop_lookupname(afile, AFS_UIOSYS, 0, &filevp);
3266 ReleaseWriteLock(&afs_xdcache);
3267 ReleaseWriteLock(&tdc->lock);
3272 * We have a VN_HOLD on filevp. Get the useful info out and
3273 * return. We make use of the fact that the cache is in the
3274 * UFS file system, and just record the inode number.
3276 #ifdef AFS_LINUX22_ENV
3277 tdc->f.inode = VTOI(filevp->d_inode)->i_number;
3280 tdc->f.inode = afs_vnodeToInumber(filevp);
3282 #endif /* AFS_LINUX22_ENV */
3284 tdc->f.inode = ainode;
3287 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3289 tfile = osi_UFSOpen(tdc->f.inode);
3290 code = afs_osi_Stat(tfile, &tstat);
3292 osi_Panic("initcachefile stat");
3295 * If file size doesn't match the cache info file, it's probably bad.
3297 if (tdc->f.chunkBytes != tstat.size)
3299 tdc->f.chunkBytes = 0;
3302 * If file changed within T (120?) seconds of cache info file, it's
3303 * probably bad. In addition, if slot changed within last T seconds,
3304 * the cache info file may be incorrectly identified, and so slot
3307 if (cacheInfoModTime < tstat.mtime + 120)
3309 if (cacheInfoModTime < tdc->f.modTime + 120)
3311 /* In case write through is behind, make sure cache items entry is
3312 * at least as new as the chunk.
3314 if (tdc->f.modTime < tstat.mtime)
3317 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3318 if (tstat.size != 0)
3319 osi_UFSTruncate(tfile, 0);
3320 tdc->f.states &= ~(DRO|DBackup|DRW);
3321 afs_DCMoveBucket(tdc, 0, 0);
3322 /* put entry in free cache slot list */
3323 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3324 afs_freeDCList = index;
3326 afs_indexFlags[index] |= IFFree;
3327 afs_indexUnique[index] = 0;
3330 * We must put this entry in the appropriate hash tables.
3331 * Note that i is still set from the above DCHash call
3333 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3334 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3335 afs_dchashTbl[code] = tdc->index;
3336 code = DVHash(&tdc->f.fid);
3337 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3338 afs_dvhashTbl[code] = tdc->index;
3339 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3341 /* has nontrivial amt of data */
3342 afs_indexFlags[index] |= IFEverUsed;
3343 afs_stats_cmperf.cacheFilesReused++;
3345 * Initialize index times to file's mod times; init indexCounter
3348 hset32(afs_indexTimes[index], tstat.atime);
3349 if (hgetlo(afs_indexCounter) < tstat.atime) {
3350 hset32(afs_indexCounter, tstat.atime);
3352 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3353 } /*File is not bad */
3355 osi_UFSClose(tfile);
3356 tdc->f.states &= ~DWriting;
3357 tdc->dflags &= ~DFEntryMod;
3358 /* don't set f.modTime; we're just cleaning up */
3359 afs_WriteDCache(tdc, 0);
3360 ReleaseWriteLock(&afs_xdcache);
3361 ReleaseWriteLock(&tdc->lock);
3363 afs_stats_cmperf.cacheNumEntries++;
3368 /*Max # of struct dcache's resident at any time*/
3370 * If 'dchint' is enabled then in-memory dcache min is increased because of
3379 * Initialize dcache related variables.
3382 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3384 register struct dcache *tdp;
3388 afs_freeDCList = NULLIDX;
3389 afs_discardDCList = NULLIDX;
3390 afs_freeDCCount = 0;
3391 afs_freeDSList = NULL;
3392 hzero(afs_indexCounter);
3394 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3400 if (achunk < 0 || achunk > 30)
3401 achunk = 13; /* Use default */
3402 AFS_SETCHUNKSIZE(achunk);
3408 if (aflags & AFSCALL_INIT_MEMCACHE) {
3410 * Use a memory cache instead of a disk cache
3412 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3413 afs_cacheType = &afs_MemCacheOps;
3414 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3415 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3416 /* ablocks is reported in 1K blocks */
3417 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3419 printf("afsd: memory cache too large for available memory.\n");
3420 printf("afsd: AFS files cannot be accessed.\n\n");
3422 afiles = ablocks = 0;
3424 printf("Memory cache: Allocating %d dcache entries...",
3427 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3428 afs_cacheType = &afs_UfsCacheOps;
3431 if (aDentries > 512)
3432 afs_dhashsize = 2048;
3433 /* initialize hash tables */
3435 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3437 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3438 for (i = 0; i < afs_dhashsize; i++) {
3439 afs_dvhashTbl[i] = NULLIDX;
3440 afs_dchashTbl[i] = NULLIDX;
3442 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3443 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3444 for (i = 0; i < afiles; i++) {
3445 afs_dvnextTbl[i] = NULLIDX;
3446 afs_dcnextTbl[i] = NULLIDX;
3449 /* Allocate and zero the pointer array to the dcache entries */
3450 afs_indexTable = (struct dcache **)
3451 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3452 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3454 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3455 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3457 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3458 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3459 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3460 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3462 /* Allocate and thread the struct dcache entries themselves */
3463 tdp = afs_Initial_freeDSList =
3464 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3465 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3466 #ifdef KERNEL_HAVE_PIN
3467 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3468 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3469 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3470 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3471 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3472 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3473 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3474 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3475 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3478 afs_freeDSList = &tdp[0];
3479 for (i = 0; i < aDentries - 1; i++) {
3480 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3481 RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3482 RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3483 RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3485 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3486 RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3487 RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3488 RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3490 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3491 afs_cacheBlocks = ablocks;
3492 afs_ComputeCacheParms(); /* compute parms based on cache size */
3494 afs_dcentries = aDentries;
3496 afs_stats_cmperf.cacheBucket0_Discarded =
3497 afs_stats_cmperf.cacheBucket1_Discarded =
3498 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3508 shutdown_dcache(void)
3512 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3513 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3514 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3515 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3516 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3517 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3518 afs_osi_Free(afs_Initial_freeDSList,
3519 afs_dcentries * sizeof(struct dcache));
3520 #ifdef KERNEL_HAVE_PIN
3521 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3522 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3523 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3524 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3525 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3526 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3527 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3531 for (i = 0; i < afs_dhashsize; i++) {
3532 afs_dvhashTbl[i] = NULLIDX;
3533 afs_dchashTbl[i] = NULLIDX;
3536 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3537 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3539 afs_blocksUsed = afs_dcentries = 0;
3540 afs_stats_cmperf.cacheBucket0_Discarded =
3541 afs_stats_cmperf.cacheBucket1_Discarded =
3542 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3543 hzero(afs_indexCounter);
3545 afs_freeDCCount = 0;
3546 afs_freeDCList = NULLIDX;
3547 afs_discardDCList = NULLIDX;
3548 afs_freeDSList = afs_Initial_freeDSList = 0;
3550 LOCK_INIT(&afs_xdcache, "afs_xdcache");