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 */
519 AFS_STATCNT(afs_GetDownD);
520 if (CheckLock(&afs_xdcache) != -1)
521 osi_Panic("getdownd nolock");
522 /* decrement anumber first for all dudes in free list */
523 /* SHOULD always decrement anumber first, even if aneedSpace >0,
524 * because we should try to free space even if anumber <=0 */
525 if (!aneedSpace || *aneedSpace <= 0) {
526 anumber -= afs_freeDCCount;
528 return; /* enough already free */
530 /* bounds check parameter */
531 if (anumber > MAXATONCE)
532 anumber = MAXATONCE; /* all we can do */
534 /* rewrite so phases include a better eligiblity for gc test*/
536 * The phase variable manages reclaims. Set to 0, the first pass,
537 * we don't reclaim active entries, or other than target bucket.
538 * Set to 1, we reclaim even active ones in target bucket.
539 * Set to 2, we reclaim any inactive one.
540 * Set to 3, we reclaim even active ones.
548 for (i = 0; i < afs_cacheFiles; i++)
549 /* turn off all flags */
550 afs_indexFlags[i] &= ~IFFlag;
552 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
553 /* find oldest entries for reclamation */
554 maxVictimPtr = victimPtr = 0;
555 hzero(maxVictimTime);
556 curbucket = afs_DCWhichBucket(phase, buckethint);
557 /* select victims from access time array */
558 for (i = 0; i < afs_cacheFiles; i++) {
559 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
560 /* skip if dirty or already free */
563 tdc = afs_indexTable[i];
564 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
566 /* Wrong bucket; can't use it! */
569 if (tdc && (tdc->refCount != 0)) {
570 /* Referenced; can't use it! */
573 hset(vtime, afs_indexTimes[i]);
575 /* if we've already looked at this one, skip it */
576 if (afs_indexFlags[i] & IFFlag)
579 if (victimPtr < MAXATONCE) {
580 /* if there's at least one free victim slot left */
581 victims[victimPtr] = i;
582 hset(victimTimes[victimPtr], vtime);
583 if (hcmp(vtime, maxVictimTime) > 0) {
584 hset(maxVictimTime, vtime);
585 maxVictimPtr = victimPtr;
588 } else if (hcmp(vtime, maxVictimTime) < 0) {
590 * We're older than youngest victim, so we replace at
593 /* find youngest (largest LRU) victim */
596 osi_Panic("getdownd local");
598 hset(victimTimes[j], vtime);
599 /* recompute maxVictimTime */
600 hset(maxVictimTime, vtime);
601 for (j = 0; j < victimPtr; j++)
602 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
603 hset(maxVictimTime, victimTimes[j]);
609 /* now really reclaim the victims */
610 j = 0; /* flag to track if we actually got any of the victims */
611 /* first, hold all the victims, since we're going to release the lock
612 * during the truncate operation.
614 for (i = 0; i < victimPtr; i++) {
615 tdc = afs_GetDSlot(victims[i], 0);
616 /* We got tdc->tlock(R) here */
617 if (tdc->refCount == 1)
621 ReleaseReadLock(&tdc->tlock);
625 for (i = 0; i < victimPtr; i++) {
626 /* q is first elt in dcache entry */
628 /* now, since we're dropping the afs_xdcache lock below, we
629 * have to verify, before proceeding, that there are no other
630 * references to this dcache entry, even now. Note that we
631 * compare with 1, since we bumped it above when we called
632 * afs_GetDSlot to preserve the entry's identity.
634 if (tdc && tdc->refCount == 1) {
635 unsigned char chunkFlags;
636 afs_size_t tchunkoffset = 0;
638 /* xdcache is lower than the xvcache lock */
639 MReleaseWriteLock(&afs_xdcache);
640 MObtainReadLock(&afs_xvcache);
641 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
642 MReleaseReadLock(&afs_xvcache);
643 MObtainWriteLock(&afs_xdcache, 527);
645 if (tdc->refCount > 1)
648 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
649 chunkFlags = afs_indexFlags[tdc->index];
650 if (((phase & 1) == 0) && osi_Active(tvc))
652 if (((phase & 1) == 1) && osi_Active(tvc)
653 && (tvc->states & CDCLock)
654 && (chunkFlags & IFAnyPages))
656 if (chunkFlags & IFDataMod)
658 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
659 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
660 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
661 ICL_HANDLE_OFFSET(tchunkoffset));
663 #if defined(AFS_SUN5_ENV)
665 * Now we try to invalidate pages. We do this only for
666 * Solaris. For other platforms, it's OK to recycle a
667 * dcache entry out from under a page, because the strategy
668 * function can call afs_GetDCache().
670 if (!skip && (chunkFlags & IFAnyPages)) {
673 MReleaseWriteLock(&afs_xdcache);
674 MObtainWriteLock(&tvc->vlock, 543);
675 if (tvc->multiPage) {
679 /* block locking pages */
680 tvc->vstates |= VPageCleaning;
681 /* block getting new pages */
683 MReleaseWriteLock(&tvc->vlock);
684 /* One last recheck */
685 MObtainWriteLock(&afs_xdcache, 333);
686 chunkFlags = afs_indexFlags[tdc->index];
687 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
688 || (osi_Active(tvc) && (tvc->states & CDCLock)
689 && (chunkFlags & IFAnyPages))) {
691 MReleaseWriteLock(&afs_xdcache);
694 MReleaseWriteLock(&afs_xdcache);
696 code = osi_VM_GetDownD(tvc, tdc);
698 MObtainWriteLock(&afs_xdcache, 269);
699 /* we actually removed all pages, clean and dirty */
701 afs_indexFlags[tdc->index] &=
702 ~(IFDirtyPages | IFAnyPages);
705 MReleaseWriteLock(&afs_xdcache);
707 MObtainWriteLock(&tvc->vlock, 544);
708 if (--tvc->activeV == 0
709 && (tvc->vstates & VRevokeWait)) {
710 tvc->vstates &= ~VRevokeWait;
711 afs_osi_Wakeup((char *)&tvc->vstates);
714 if (tvc->vstates & VPageCleaning) {
715 tvc->vstates &= ~VPageCleaning;
716 afs_osi_Wakeup((char *)&tvc->vstates);
719 MReleaseWriteLock(&tvc->vlock);
721 #endif /* AFS_SUN5_ENV */
723 MReleaseWriteLock(&afs_xdcache);
726 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
727 MObtainWriteLock(&afs_xdcache, 528);
728 if (afs_indexFlags[tdc->index] &
729 (IFDataMod | IFDirtyPages | IFAnyPages))
731 if (tdc->refCount > 1)
734 #if defined(AFS_SUN5_ENV)
736 /* no vnode, so IFDirtyPages is spurious (we don't
737 * sweep dcaches on vnode recycling, so we can have
738 * DIRTYPAGES set even when all pages are gone). Just
740 * Hold vcache lock to prevent vnode from being
741 * created while we're clearing IFDirtyPages.
743 afs_indexFlags[tdc->index] &=
744 ~(IFDirtyPages | IFAnyPages);
748 /* skip this guy and mark him as recently used */
749 afs_indexFlags[tdc->index] |= IFFlag;
750 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
751 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
752 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
753 ICL_HANDLE_OFFSET(tchunkoffset));
755 /* flush this dude from the data cache and reclaim;
756 * first, make sure no one will care that we damage
757 * it, by removing it from all hash tables. Then,
758 * melt it down for parts. Note that any concurrent
759 * (new possibility!) calls to GetDownD won't touch
760 * this guy because his reference count is > 0. */
761 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
762 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
763 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
764 ICL_HANDLE_OFFSET(tchunkoffset));
765 AFS_STATCNT(afs_gget);
766 afs_HashOutDCache(tdc);
767 if (tdc->f.chunkBytes != 0) {
771 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
776 afs_DiscardDCache(tdc);
781 j = 1; /* we reclaimed at least one victim */
788 /* Phase is 0 and no one was found, so try phase 1 (ignore
789 * osi_Active flag) */
792 for (i = 0; i < afs_cacheFiles; i++)
793 /* turn off all flags */
794 afs_indexFlags[i] &= ~IFFlag;
797 /* found no one in phases 0-5, we're hosed */
801 } /* big while loop */
808 * Description: remove adc from any hash tables that would allow it to be located
809 * again by afs_FindDCache or afs_GetDCache.
811 * Parameters: adc -- pointer to dcache entry to remove from hash tables.
813 * Locks: Must have the afs_xdcache lock write-locked to call this function.
816 afs_HashOutDCache(struct dcache *adc)
820 AFS_STATCNT(afs_glink);
821 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
823 /* if this guy is in the hash table, pull him out */
824 if (adc->f.fid.Fid.Volume != 0) {
825 /* remove entry from first hash chains */
826 i = DCHash(&adc->f.fid, adc->f.chunk);
827 us = afs_dchashTbl[i];
828 if (us == adc->index) {
829 /* first dude in the list */
830 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
832 /* somewhere on the chain */
833 while (us != NULLIDX) {
834 if (afs_dcnextTbl[us] == adc->index) {
835 /* found item pointing at the one to delete */
836 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
839 us = afs_dcnextTbl[us];
842 osi_Panic("dcache hc");
844 /* remove entry from *other* hash chain */
845 i = DVHash(&adc->f.fid);
846 us = afs_dvhashTbl[i];
847 if (us == adc->index) {
848 /* first dude in the list */
849 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
851 /* somewhere on the chain */
852 while (us != NULLIDX) {
853 if (afs_dvnextTbl[us] == adc->index) {
854 /* found item pointing at the one to delete */
855 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
858 us = afs_dvnextTbl[us];
861 osi_Panic("dcache hv");
865 /* prevent entry from being found on a reboot (it is already out of
866 * the hash table, but after a crash, we just look at fid fields of
867 * stable (old) entries).
869 adc->f.fid.Fid.Volume = 0; /* invalid */
871 /* mark entry as modified */
872 adc->dflags |= DFEntryMod;
876 } /*afs_HashOutDCache */
883 * Flush the given dcache entry, pulling it from hash chains
884 * and truncating the associated cache file.
887 * adc: Ptr to dcache entry to flush.
890 * This routine must be called with the afs_xdcache lock held
895 afs_FlushDCache(register struct dcache *adc)
897 AFS_STATCNT(afs_FlushDCache);
899 * Bump the number of cache files flushed.
901 afs_stats_cmperf.cacheFlushes++;
903 /* remove from all hash tables */
904 afs_HashOutDCache(adc);
906 /* Free its space; special case null operation, since truncate operation
907 * in UFS is slow even in this case, and this allows us to pre-truncate
908 * these files at more convenient times with fewer locks set
909 * (see afs_GetDownD).
911 if (adc->f.chunkBytes != 0) {
912 afs_DiscardDCache(adc);
913 afs_MaybeWakeupTruncateDaemon();
918 if (afs_WaitForCacheDrain) {
919 if (afs_blocksUsed <=
920 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
921 afs_WaitForCacheDrain = 0;
922 afs_osi_Wakeup(&afs_WaitForCacheDrain);
925 } /*afs_FlushDCache */
931 * Description: put a dcache entry on the free dcache entry list.
933 * Parameters: adc -- dcache entry to free
935 * Environment: called with afs_xdcache lock write-locked.
938 afs_FreeDCache(register struct dcache *adc)
940 /* Thread on free list, update free list count and mark entry as
941 * freed in its indexFlags element. Also, ensure DCache entry gets
942 * written out (set DFEntryMod).
945 afs_dvnextTbl[adc->index] = afs_freeDCList;
946 afs_freeDCList = adc->index;
948 afs_indexFlags[adc->index] |= IFFree;
949 adc->dflags |= DFEntryMod;
951 if (afs_WaitForCacheDrain) {
952 if ((afs_blocksUsed - afs_blocksDiscarded) <=
953 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
954 afs_WaitForCacheDrain = 0;
955 afs_osi_Wakeup(&afs_WaitForCacheDrain);
964 * Discard the cache element by moving it to the discardDCList.
965 * This puts the cache element into a quasi-freed state, where
966 * the space may be reused, but the file has not been truncated.
968 * Major Assumptions Here:
969 * Assumes that frag size is an integral power of two, less one,
970 * and that this is a two's complement machine. I don't
971 * know of any filesystems which violate this assumption...
974 * adc : Ptr to dcache entry.
977 * Must be called with afs_xdcache write-locked.
981 afs_DiscardDCache(register struct dcache *adc)
983 register afs_int32 size;
985 AFS_STATCNT(afs_DiscardDCache);
987 osi_Assert(adc->refCount == 1);
989 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
990 afs_blocksDiscarded += size;
991 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
993 afs_dvnextTbl[adc->index] = afs_discardDCList;
994 afs_discardDCList = adc->index;
995 afs_discardDCCount++;
997 adc->f.fid.Fid.Volume = 0;
998 adc->dflags |= DFEntryMod;
999 afs_indexFlags[adc->index] |= IFDiscarded;
1001 if (afs_WaitForCacheDrain) {
1002 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1003 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1004 afs_WaitForCacheDrain = 0;
1005 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1009 } /*afs_DiscardDCache */
1012 * afs_FreeDiscardedDCache
1015 * Free the next element on the list of discarded cache elements.
1018 afs_FreeDiscardedDCache(void)
1020 register struct dcache *tdc;
1021 register struct osi_file *tfile;
1022 register afs_int32 size;
1024 AFS_STATCNT(afs_FreeDiscardedDCache);
1026 MObtainWriteLock(&afs_xdcache, 510);
1027 if (!afs_blocksDiscarded) {
1028 MReleaseWriteLock(&afs_xdcache);
1033 * Get an entry from the list of discarded cache elements
1035 tdc = afs_GetDSlot(afs_discardDCList, 0);
1036 osi_Assert(tdc->refCount == 1);
1037 ReleaseReadLock(&tdc->tlock);
1039 afs_discardDCList = afs_dvnextTbl[tdc->index];
1040 afs_dvnextTbl[tdc->index] = NULLIDX;
1041 afs_discardDCCount--;
1042 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1043 afs_blocksDiscarded -= size;
1044 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1045 /* We can lock because we just took it off the free list */
1046 ObtainWriteLock(&tdc->lock, 626);
1047 MReleaseWriteLock(&afs_xdcache);
1050 * Truncate the element to reclaim its space
1052 tfile = afs_CFileOpen(tdc->f.inode);
1053 afs_CFileTruncate(tfile, 0);
1054 afs_CFileClose(tfile);
1055 afs_AdjustSize(tdc, 0);
1056 afs_DCMoveBucket(tdc, 0, 0);
1059 * Free the element we just truncated
1061 MObtainWriteLock(&afs_xdcache, 511);
1062 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1063 afs_FreeDCache(tdc);
1064 tdc->f.states &= ~(DRO|DBackup|DRW);
1065 ReleaseWriteLock(&tdc->lock);
1067 MReleaseWriteLock(&afs_xdcache);
1071 * afs_MaybeFreeDiscardedDCache
1074 * Free as many entries from the list of discarded cache elements
1075 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1081 afs_MaybeFreeDiscardedDCache(void)
1084 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1086 while (afs_blocksDiscarded
1087 && (afs_blocksUsed >
1088 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1089 afs_FreeDiscardedDCache();
1098 * Try to free up a certain number of disk slots.
1101 * anumber : Targeted number of disk slots to free up.
1104 * Must be called with afs_xdcache write-locked.
1107 afs_GetDownDSlot(int anumber)
1109 struct afs_q *tq, *nq;
1114 AFS_STATCNT(afs_GetDownDSlot);
1115 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1116 osi_Panic("diskless getdowndslot");
1118 if (CheckLock(&afs_xdcache) != -1)
1119 osi_Panic("getdowndslot nolock");
1121 /* decrement anumber first for all dudes in free list */
1122 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1125 return; /* enough already free */
1127 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1129 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1130 nq = QPrev(tq); /* in case we remove it */
1131 if (tdc->refCount == 0) {
1132 if ((ix = tdc->index) == NULLIDX)
1133 osi_Panic("getdowndslot");
1134 /* pull the entry out of the lruq and put it on the free list */
1135 QRemove(&tdc->lruq);
1137 /* write-through if modified */
1138 if (tdc->dflags & DFEntryMod) {
1139 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1141 * ask proxy to do this for us - we don't have the stack space
1143 while (tdc->dflags & DFEntryMod) {
1146 s = SPLOCK(afs_sgibklock);
1147 if (afs_sgibklist == NULL) {
1148 /* if slot is free, grab it. */
1149 afs_sgibklist = tdc;
1150 SV_SIGNAL(&afs_sgibksync);
1152 /* wait for daemon to (start, then) finish. */
1153 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1157 tdc->dflags &= ~DFEntryMod;
1158 afs_WriteDCache(tdc, 1);
1162 /* finally put the entry in the free list */
1163 afs_indexTable[ix] = NULL;
1164 afs_indexFlags[ix] &= ~IFEverUsed;
1165 tdc->index = NULLIDX;
1166 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1167 afs_freeDSList = tdc;
1171 } /*afs_GetDownDSlot */
1178 * Increment the reference count on a disk cache entry,
1179 * which already has a non-zero refcount. In order to
1180 * increment the refcount of a zero-reference entry, you
1181 * have to hold afs_xdcache.
1184 * adc : Pointer to the dcache entry to increment.
1187 * Nothing interesting.
1190 afs_RefDCache(struct dcache *adc)
1192 ObtainWriteLock(&adc->tlock, 627);
1193 if (adc->refCount < 0)
1194 osi_Panic("RefDCache: negative refcount");
1196 ReleaseWriteLock(&adc->tlock);
1205 * Decrement the reference count on a disk cache entry.
1208 * ad : Ptr to the dcache entry to decrement.
1211 * Nothing interesting.
1214 afs_PutDCache(register struct dcache *adc)
1216 AFS_STATCNT(afs_PutDCache);
1217 ObtainWriteLock(&adc->tlock, 276);
1218 if (adc->refCount <= 0)
1219 osi_Panic("putdcache");
1221 ReleaseWriteLock(&adc->tlock);
1230 * Try to discard all data associated with this file from the
1234 * avc : Pointer to the cache info for the file.
1237 * Both pvnLock and lock are write held.
1240 afs_TryToSmush(register struct vcache *avc, struct AFS_UCRED *acred, int sync)
1242 register struct dcache *tdc;
1245 AFS_STATCNT(afs_TryToSmush);
1246 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1247 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length));
1248 sync = 1; /* XX Temp testing XX */
1250 #if defined(AFS_SUN5_ENV)
1251 ObtainWriteLock(&avc->vlock, 573);
1252 avc->activeV++; /* block new getpages */
1253 ReleaseWriteLock(&avc->vlock);
1256 /* Flush VM pages */
1257 osi_VM_TryToSmush(avc, acred, sync);
1260 * Get the hash chain containing all dce's for this fid
1262 i = DVHash(&avc->fid);
1263 MObtainWriteLock(&afs_xdcache, 277);
1264 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1265 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1266 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1267 int releaseTlock = 1;
1268 tdc = afs_GetDSlot(index, NULL);
1269 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1271 if ((afs_indexFlags[index] & IFDataMod) == 0
1272 && tdc->refCount == 1) {
1273 ReleaseReadLock(&tdc->tlock);
1275 afs_FlushDCache(tdc);
1278 afs_indexTable[index] = 0;
1281 ReleaseReadLock(&tdc->tlock);
1285 #if defined(AFS_SUN5_ENV)
1286 ObtainWriteLock(&avc->vlock, 545);
1287 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1288 avc->vstates &= ~VRevokeWait;
1289 afs_osi_Wakeup((char *)&avc->vstates);
1291 ReleaseWriteLock(&avc->vlock);
1293 MReleaseWriteLock(&afs_xdcache);
1295 * It's treated like a callback so that when we do lookups we'll
1296 * invalidate the unique bit if any
1297 * trytoSmush occured during the lookup call
1303 * afs_DCacheMissingChunks
1306 * Given the cached info for a file, return the number of chunks that
1307 * are not available from the dcache.
1310 * avc: Pointer to the (held) vcache entry to look in.
1313 * The number of chunks which are not currently cached.
1316 * The vcache entry is held upon entry.
1320 afs_DCacheMissingChunks(struct vcache *avc)
1323 afs_size_t totalLength;
1324 afs_uint32 totalChunks;
1327 totalLength = avc->m.Length;
1328 if (avc->truncPos < totalLength)
1329 totalLength = avc->truncPos;
1331 totalChunks = AFS_CHUNK(totalLength) + 1;
1333 /*printf("Should have %d chunks for %d bytes\n", totalChunks, totalLength);*/
1335 i = DVHash(&avc->fid);
1336 MObtainWriteLock(&afs_xdcache, 1001);
1337 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1338 i = afs_dvnextTbl[index];
1339 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1340 tdc = afs_GetDSlot(index, NULL);
1341 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1344 ReleaseReadLock(&tdc->tlock);
1348 MReleaseWriteLock(&afs_xdcache);
1350 /*printf("Missing %d chunks\n", totalChunks);*/
1352 return (totalChunks);
1359 * Given the cached info for a file and a byte offset into the
1360 * file, make sure the dcache entry for that file and containing
1361 * the given byte is available, returning it to our caller.
1364 * avc : Pointer to the (held) vcache entry to look in.
1365 * abyte : Which byte we want to get to.
1368 * Pointer to the dcache entry covering the file & desired byte,
1369 * or NULL if not found.
1372 * The vcache entry is held upon entry.
1376 afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1379 register afs_int32 i, index;
1380 register struct dcache *tdc = NULL;
1382 AFS_STATCNT(afs_FindDCache);
1383 chunk = AFS_CHUNK(abyte);
1386 * Hash on the [fid, chunk] and get the corresponding dcache index
1387 * after write-locking the dcache.
1389 i = DCHash(&avc->fid, chunk);
1390 MObtainWriteLock(&afs_xdcache, 278);
1391 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1392 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1393 tdc = afs_GetDSlot(index, NULL);
1394 ReleaseReadLock(&tdc->tlock);
1395 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1396 break; /* leaving refCount high for caller */
1400 index = afs_dcnextTbl[index];
1402 if (index != NULLIDX) {
1403 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1404 hadd32(afs_indexCounter, 1);
1405 MReleaseWriteLock(&afs_xdcache);
1408 MReleaseWriteLock(&afs_xdcache);
1410 } /*afs_FindDCache */
1414 * afs_UFSCacheStoreProc
1417 * Called upon store.
1420 * acall : Ptr to the Rx call structure involved.
1421 * afile : Ptr to the related file descriptor.
1422 * alen : Size of the file in bytes.
1423 * avc : Ptr to the vcache entry.
1424 * shouldWake : is it "safe" to return early from close() ?
1425 * abytesToXferP : Set to the number of bytes to xfer.
1426 * NOTE: This parameter is only used if AFS_NOSTATS
1428 * abytesXferredP : Set to the number of bytes actually xferred.
1429 * NOTE: This parameter is only used if AFS_NOSTATS
1433 * Nothing interesting.
1436 afs_UFSCacheStoreProc(register struct rx_call *acall, struct osi_file *afile,
1437 register afs_int32 alen, struct vcache *avc,
1438 int *shouldWake, afs_size_t * abytesToXferP,
1439 afs_size_t * abytesXferredP)
1441 afs_int32 code, got;
1442 register char *tbuffer;
1445 AFS_STATCNT(UFS_CacheStoreProc);
1449 * In this case, alen is *always* the amount of data we'll be trying
1452 (*abytesToXferP) = alen;
1453 (*abytesXferredP) = 0;
1454 #endif /* AFS_NOSTATS */
1456 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1457 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1458 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1459 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1461 tlen = (alen > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : alen);
1462 got = afs_osi_Read(afile, -1, tbuffer, tlen);
1464 #if defined(KERNEL_HAVE_UERROR)
1465 || (got != tlen && getuerror())
1468 osi_FreeLargeSpace(tbuffer);
1471 afs_Trace2(afs_iclSetp, CM_TRACE_STOREPROC2, ICL_TYPE_OFFSET,
1472 ICL_HANDLE_OFFSET(*tbuffer), ICL_TYPE_INT32, got);
1474 code = rx_Write(acall, tbuffer, got); /* writing 0 bytes will
1475 * push a short packet. Is that really what we want, just because the
1476 * data didn't come back from the disk yet? Let's try it and see. */
1479 (*abytesXferredP) += code;
1480 #endif /* AFS_NOSTATS */
1482 code = rx_Error(acall);
1483 osi_FreeLargeSpace(tbuffer);
1484 return code ? code : -33;
1488 * If file has been locked on server, we can allow the store
1491 if (shouldWake && *shouldWake && (rx_GetRemoteStatus(acall) & 1)) {
1492 *shouldWake = 0; /* only do this once */
1496 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1497 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1498 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1499 osi_FreeLargeSpace(tbuffer);
1502 } /* afs_UFSCacheStoreProc */
1506 * afs_UFSCacheFetchProc
1509 * Routine called on fetch; also tells people waiting for data
1510 * that more has arrived.
1513 * acall : Ptr to the Rx call structure.
1514 * afile : File descriptor for the cache file.
1515 * abase : Base offset to fetch.
1516 * adc : Ptr to the dcache entry for the file, write-locked.
1517 * avc : Ptr to the vcache entry for the file.
1518 * abytesToXferP : Set to the number of bytes to xfer.
1519 * NOTE: This parameter is only used if AFS_NOSTATS
1521 * abytesXferredP : Set to the number of bytes actually xferred.
1522 * NOTE: This parameter is only used if AFS_NOSTATS
1526 * Nothing interesting.
1530 afs_UFSCacheFetchProc(register struct rx_call *acall, struct osi_file *afile,
1531 afs_size_t abase, struct dcache *adc,
1532 struct vcache *avc, afs_size_t * abytesToXferP,
1533 afs_size_t * abytesXferredP, afs_int32 lengthFound)
1536 register afs_int32 code;
1537 register char *tbuffer;
1541 AFS_STATCNT(UFS_CacheFetchProc);
1542 osi_Assert(WriteLocked(&adc->lock));
1543 afile->offset = 0; /* Each time start from the beginning */
1544 length = lengthFound;
1546 (*abytesToXferP) = 0;
1547 (*abytesXferredP) = 0;
1548 #endif /* AFS_NOSTATS */
1549 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1550 adc->validPos = abase;
1554 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
1556 length = ntohl(length);
1557 if (code != sizeof(afs_int32)) {
1558 osi_FreeLargeSpace(tbuffer);
1559 code = rx_Error(acall);
1560 return (code ? code : -1); /* try to return code, not -1 */
1564 * The fetch protocol is extended for the AFS/DFS translator
1565 * to allow multiple blocks of data, each with its own length,
1566 * to be returned. As long as the top bit is set, there are more
1569 * We do not do this for AFS file servers because they sometimes
1570 * return large negative numbers as the transfer size.
1572 if (avc->states & CForeign) {
1573 moredata = length & 0x80000000;
1574 length &= ~0x80000000;
1579 (*abytesToXferP) += length;
1580 #endif /* AFS_NOSTATS */
1581 while (length > 0) {
1582 tlen = (length > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : length);
1583 #ifdef RX_KERNEL_TRACE
1584 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1588 code = rx_Read(acall, tbuffer, tlen);
1590 #ifdef RX_KERNEL_TRACE
1591 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1595 (*abytesXferredP) += code;
1596 #endif /* AFS_NOSTATS */
1598 osi_FreeLargeSpace(tbuffer);
1599 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64READ,
1600 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
1601 ICL_TYPE_INT32, length);
1604 code = afs_osi_Write(afile, -1, tbuffer, tlen);
1606 osi_FreeLargeSpace(tbuffer);
1611 adc->validPos = abase;
1612 if (afs_osi_Wakeup(&adc->validPos) == 0)
1613 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
1614 __FILE__, ICL_TYPE_INT32, __LINE__,
1615 ICL_TYPE_POINTER, adc, ICL_TYPE_INT32,
1619 osi_FreeLargeSpace(tbuffer);
1622 } /* afs_UFSCacheFetchProc */
1628 * This function is called to obtain a reference to data stored in
1629 * the disk cache, locating a chunk of data containing the desired
1630 * byte and returning a reference to the disk cache entry, with its
1631 * reference count incremented.
1635 * avc : Ptr to a vcache entry (unlocked)
1636 * abyte : Byte position in the file desired
1637 * areq : Request structure identifying the requesting user.
1638 * aflags : Settings as follows:
1640 * 2 : Return after creating entry.
1641 * 4 : called from afs_vnop_write.c
1642 * *alen contains length of data to be written.
1644 * aoffset : Set to the offset within the chunk where the resident
1646 * alen : Set to the number of bytes of data after the desired
1647 * byte (including the byte itself) which can be read
1651 * The vcache entry pointed to by avc is unlocked upon entry.
1655 struct AFSVolSync tsync;
1656 struct AFSFetchStatus OutStatus;
1657 struct AFSCallBack CallBack;
1661 * Update the vnode-to-dcache hint if we can get the vnode lock
1662 * right away. Assumes dcache entry is at least read-locked.
1665 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1667 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1668 if (hsame(v->m.DataVersion, d->f.versionNo) && v->callback)
1671 ReleaseWriteLock(&v->lock);
1675 /* avc - Write-locked unless aflags & 1 */
1677 afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1678 register struct vrequest *areq, afs_size_t * aoffset,
1679 afs_size_t * alen, int aflags)
1681 register afs_int32 i, code, code1 = 0, shortcut;
1682 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1683 register afs_int32 adjustsize = 0;
1689 afs_size_t maxGoodLength; /* amount of good data at server */
1690 struct rx_call *tcall;
1691 afs_size_t Position = 0;
1692 #ifdef AFS_64BIT_CLIENT
1694 afs_size_t lengthFound; /* as returned from server */
1695 #endif /* AFS_64BIT_CLIENT */
1696 afs_int32 size, tlen; /* size of segment to transfer */
1697 struct tlocal1 *tsmall = 0;
1698 register struct dcache *tdc;
1699 register struct osi_file *file;
1700 register struct conn *tc;
1702 struct server *newCallback = NULL;
1703 char setNewCallback;
1704 char setVcacheStatus;
1705 char doVcacheUpdate;
1707 int doAdjustSize = 0;
1708 int doReallyAdjustSize = 0;
1709 int overWriteWholeChunk = 0;
1713 struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
1714 osi_timeval_t xferStartTime, /*FS xfer start time */
1715 xferStopTime; /*FS xfer stop time */
1716 afs_size_t bytesToXfer; /* # bytes to xfer */
1717 afs_size_t bytesXferred; /* # bytes actually xferred */
1718 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1719 int fromReplica; /*Are we reading from a replica? */
1720 int numFetchLoops; /*# times around the fetch/analyze loop */
1721 #endif /* AFS_NOSTATS */
1723 AFS_STATCNT(afs_GetDCache);
1728 setLocks = aflags & 1;
1731 * Determine the chunk number and offset within the chunk corresponding
1732 * to the desired byte.
1734 if (avc->fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1737 chunk = AFS_CHUNK(abyte);
1740 /* come back to here if we waited for the cache to drain. */
1743 setNewCallback = setVcacheStatus = 0;
1747 ObtainWriteLock(&avc->lock, 616);
1749 ObtainReadLock(&avc->lock);
1754 * avc->lock(R) if setLocks && !slowPass
1755 * avc->lock(W) if !setLocks || slowPass
1760 /* check hints first! (might could use bcmp or some such...) */
1761 if ((tdc = avc->dchint)) {
1765 * The locking order between afs_xdcache and dcache lock matters.
1766 * The hint dcache entry could be anywhere, even on the free list.
1767 * Locking afs_xdcache ensures that noone is trying to pull dcache
1768 * entries from the free list, and thereby assuming them to be not
1769 * referenced and not locked.
1771 MObtainReadLock(&afs_xdcache);
1772 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1774 if (dcLocked && (tdc->index != NULLIDX)
1775 && !FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk
1776 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1777 /* got the right one. It might not be the right version, and it
1778 * might be fetching, but it's the right dcache entry.
1780 /* All this code should be integrated better with what follows:
1781 * I can save a good bit more time under a write lock if I do..
1783 ObtainWriteLock(&tdc->tlock, 603);
1785 ReleaseWriteLock(&tdc->tlock);
1787 MReleaseReadLock(&afs_xdcache);
1790 if (hsame(tdc->f.versionNo, avc->m.DataVersion)
1791 && !(tdc->dflags & DFFetching)) {
1793 afs_stats_cmperf.dcacheHits++;
1794 MObtainWriteLock(&afs_xdcache, 559);
1795 QRemove(&tdc->lruq);
1796 QAdd(&afs_DLRU, &tdc->lruq);
1797 MReleaseWriteLock(&afs_xdcache);
1800 * avc->lock(R) if setLocks && !slowPass
1801 * avc->lock(W) if !setLocks || slowPass
1808 ReleaseSharedLock(&tdc->lock);
1809 MReleaseReadLock(&afs_xdcache);
1817 * avc->lock(R) if setLocks && !slowPass
1818 * avc->lock(W) if !setLocks || slowPass
1819 * tdc->lock(S) if tdc
1822 if (!tdc) { /* If the hint wasn't the right dcache entry */
1824 * Hash on the [fid, chunk] and get the corresponding dcache index
1825 * after write-locking the dcache.
1830 * avc->lock(R) if setLocks && !slowPass
1831 * avc->lock(W) if !setLocks || slowPass
1834 i = DCHash(&avc->fid, chunk);
1835 /* check to make sure our space is fine */
1836 afs_MaybeWakeupTruncateDaemon();
1838 MObtainWriteLock(&afs_xdcache, 280);
1840 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1841 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1842 tdc = afs_GetDSlot(index, NULL);
1843 ReleaseReadLock(&tdc->tlock);
1846 * avc->lock(R) if setLocks && !slowPass
1847 * avc->lock(W) if !setLocks || slowPass
1850 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1851 /* Move it up in the beginning of the list */
1852 if (afs_dchashTbl[i] != index) {
1853 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1854 afs_dcnextTbl[index] = afs_dchashTbl[i];
1855 afs_dchashTbl[i] = index;
1857 MReleaseWriteLock(&afs_xdcache);
1858 ObtainSharedLock(&tdc->lock, 606);
1859 break; /* leaving refCount high for caller */
1865 index = afs_dcnextTbl[index];
1869 * If we didn't find the entry, we'll create one.
1871 if (index == NULLIDX) {
1872 /* If we're disconnected, we can't do anything */
1873 if (AFS_IS_DISCONNECTED) {
1874 MReleaseWriteLock(&afs_xdcache);
1877 ReleaseWriteLock(&avc->lock);
1879 ReleaseReadLock(&avc->lock);
1885 * avc->lock(R) if setLocks
1886 * avc->lock(W) if !setLocks
1889 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1890 avc, ICL_TYPE_INT32, chunk);
1892 /* Make sure there is a free dcache entry for us to use */
1893 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1896 avc->states |= CDCLock;
1897 /* just need slots */
1898 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1900 avc->states &= ~CDCLock;
1901 if (afs_discardDCList != NULLIDX
1902 || afs_freeDCList != NULLIDX)
1904 /* If we can't get space for 5 mins we give up and panic */
1905 if (++downDCount > 300)
1906 osi_Panic("getdcache");
1907 MReleaseWriteLock(&afs_xdcache);
1910 * avc->lock(R) if setLocks
1911 * avc->lock(W) if !setLocks
1913 afs_osi_Wait(1000, 0, 0);
1918 if (afs_discardDCList == NULLIDX
1919 || ((aflags & 2) && afs_freeDCList != NULLIDX)) {
1921 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1922 tdc = afs_GetDSlot(afs_freeDCList, 0);
1923 osi_Assert(tdc->refCount == 1);
1924 ReleaseReadLock(&tdc->tlock);
1925 ObtainWriteLock(&tdc->lock, 604);
1926 afs_freeDCList = afs_dvnextTbl[tdc->index];
1929 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1930 tdc = afs_GetDSlot(afs_discardDCList, 0);
1931 osi_Assert(tdc->refCount == 1);
1932 ReleaseReadLock(&tdc->tlock);
1933 ObtainWriteLock(&tdc->lock, 605);
1934 afs_discardDCList = afs_dvnextTbl[tdc->index];
1935 afs_discardDCCount--;
1937 ((tdc->f.chunkBytes +
1938 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1939 tdc->f.states &= ~(DRO|DBackup|DRW);
1940 afs_DCMoveBucket(tdc, size, 0);
1941 afs_blocksDiscarded -= size;
1942 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1944 /* Truncate the chunk so zeroes get filled properly */
1945 file = afs_CFileOpen(tdc->f.inode);
1946 afs_CFileTruncate(file, 0);
1947 afs_CFileClose(file);
1948 afs_AdjustSize(tdc, 0);
1954 * avc->lock(R) if setLocks
1955 * avc->lock(W) if !setLocks
1961 * Fill in the newly-allocated dcache record.
1963 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1964 tdc->f.fid = avc->fid;
1965 if (avc->states & CRO)
1966 tdc->f.states = DRO;
1967 else if (avc->states & CBackup)
1968 tdc->f.states = DBackup;
1970 tdc->f.states = DRW;
1971 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1972 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1973 hones(tdc->f.versionNo); /* invalid value */
1974 tdc->f.chunk = chunk;
1975 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1977 if (tdc->lruq.prev == &tdc->lruq)
1978 osi_Panic("lruq 1");
1981 * Now add to the two hash chains - note that i is still set
1982 * from the above DCHash call.
1984 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1985 afs_dchashTbl[i] = tdc->index;
1986 i = DVHash(&avc->fid);
1987 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1988 afs_dvhashTbl[i] = tdc->index;
1989 tdc->dflags = DFEntryMod;
1991 afs_MaybeWakeupTruncateDaemon();
1992 MReleaseWriteLock(&afs_xdcache);
1993 ConvertWToSLock(&tdc->lock);
1998 /* vcache->dcache hint failed */
2001 * avc->lock(R) if setLocks && !slowPass
2002 * avc->lock(W) if !setLocks || slowPass
2005 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
2006 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2007 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
2008 hgetlo(avc->m.DataVersion));
2010 * Here we have the entry in tdc, with its refCount incremented.
2011 * Note: we don't use the S-lock on avc; it costs concurrency when
2012 * storing a file back to the server.
2016 * Not a newly created file so we need to check the file's length and
2017 * compare data versions since someone could have changed the data or we're
2018 * reading a file written elsewhere. We only want to bypass doing no-op
2019 * read rpcs on newly created files (dv of 0) since only then we guarantee
2020 * that this chunk's data hasn't been filled by another client.
2022 size = AFS_CHUNKSIZE(abyte);
2023 if (aflags & 4) /* called from write */
2025 else /* called from read */
2026 tlen = tdc->validPos - abyte;
2027 Position = AFS_CHUNKTOBASE(chunk);
2028 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
2029 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
2030 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
2031 ICL_HANDLE_OFFSET(Position));
2032 if ((aflags & 4) && (hiszero(avc->m.DataVersion)))
2034 if ((AFS_CHUNKTOBASE(chunk) >= avc->m.Length) ||
2035 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
2036 overWriteWholeChunk = 1;
2037 if (doAdjustSize || overWriteWholeChunk) {
2038 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
2040 #ifdef AFS_SGI64_ENV
2043 #else /* AFS_SGI64_ENV */
2046 #endif /* AFS_SGI64_ENV */
2047 #else /* AFS_SGI_ENV */
2050 #endif /* AFS_SGI_ENV */
2051 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->m.Length &&
2052 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2053 #if defined(AFS_SUN5_ENV) || defined(AFS_OSF_ENV)
2054 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
2056 if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
2058 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2059 !hsame(avc->m.DataVersion, tdc->f.versionNo))
2060 doReallyAdjustSize = 1;
2062 if (doReallyAdjustSize || overWriteWholeChunk) {
2063 /* no data in file to read at this position */
2064 UpgradeSToWLock(&tdc->lock, 607);
2066 file = afs_CFileOpen(tdc->f.inode);
2067 afs_CFileTruncate(file, 0);
2068 afs_CFileClose(file);
2069 afs_AdjustSize(tdc, 0);
2070 hset(tdc->f.versionNo, avc->m.DataVersion);
2071 tdc->dflags |= DFEntryMod;
2073 ConvertWToSLock(&tdc->lock);
2078 * We must read in the whole chunk if the version number doesn't
2082 /* don't need data, just a unique dcache entry */
2083 ObtainWriteLock(&afs_xdcache, 608);
2084 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2085 hadd32(afs_indexCounter, 1);
2086 ReleaseWriteLock(&afs_xdcache);
2088 updateV2DC(setLocks, avc, tdc, 553);
2089 if (vType(avc) == VDIR)
2092 *aoffset = AFS_CHUNKOFFSET(abyte);
2093 if (tdc->validPos < abyte)
2094 *alen = (afs_size_t) 0;
2096 *alen = tdc->validPos - abyte;
2097 ReleaseSharedLock(&tdc->lock);
2100 ReleaseWriteLock(&avc->lock);
2102 ReleaseReadLock(&avc->lock);
2104 return tdc; /* check if we're done */
2109 * avc->lock(R) if setLocks && !slowPass
2110 * avc->lock(W) if !setLocks || slowPass
2113 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2115 setNewCallback = setVcacheStatus = 0;
2119 * avc->lock(R) if setLocks && !slowPass
2120 * avc->lock(W) if !setLocks || slowPass
2123 if (!hsame(avc->m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2125 * Version number mismatch.
2128 * If we are disconnected, then we can't do much of anything
2129 * because the data doesn't match the file.
2131 if (AFS_IS_DISCONNECTED) {
2132 ReleaseSharedLock(&tdc->lock);
2135 ReleaseWriteLock(&avc->lock);
2137 ReleaseReadLock(&avc->lock);
2139 /* Flush the Dcache */
2144 UpgradeSToWLock(&tdc->lock, 609);
2147 * If data ever existed for this vnode, and this is a text object,
2148 * do some clearing. Now, you'd think you need only do the flush
2149 * when VTEXT is on, but VTEXT is turned off when the text object
2150 * is freed, while pages are left lying around in memory marked
2151 * with this vnode. If we would reactivate (create a new text
2152 * object from) this vnode, we could easily stumble upon some of
2153 * these old pages in pagein. So, we always flush these guys.
2154 * Sun has a wonderful lack of useful invariants in this system.
2156 * avc->flushDV is the data version # of the file at the last text
2157 * flush. Clearly, at least, we don't have to flush the file more
2158 * often than it changes
2160 if (hcmp(avc->flushDV, avc->m.DataVersion) < 0) {
2162 * By here, the cache entry is always write-locked. We can
2163 * deadlock if we call osi_Flush with the cache entry locked...
2164 * Unlock the dcache too.
2166 ReleaseWriteLock(&tdc->lock);
2167 if (setLocks && !slowPass)
2168 ReleaseReadLock(&avc->lock);
2170 ReleaseWriteLock(&avc->lock);
2174 * Call osi_FlushPages in open, read/write, and map, since it
2175 * is too hard here to figure out if we should lock the
2178 if (setLocks && !slowPass)
2179 ObtainReadLock(&avc->lock);
2181 ObtainWriteLock(&avc->lock, 66);
2182 ObtainWriteLock(&tdc->lock, 610);
2187 * avc->lock(R) if setLocks && !slowPass
2188 * avc->lock(W) if !setLocks || slowPass
2192 /* Watch for standard race condition around osi_FlushText */
2193 if (hsame(avc->m.DataVersion, tdc->f.versionNo)) {
2194 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2195 afs_stats_cmperf.dcacheHits++;
2196 ConvertWToSLock(&tdc->lock);
2200 /* Sleep here when cache needs to be drained. */
2201 if (setLocks && !slowPass
2202 && (afs_blocksUsed >
2203 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2204 /* Make sure truncate daemon is running */
2205 afs_MaybeWakeupTruncateDaemon();
2206 ObtainWriteLock(&tdc->tlock, 614);
2207 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2208 ReleaseWriteLock(&tdc->tlock);
2209 ReleaseWriteLock(&tdc->lock);
2210 ReleaseReadLock(&avc->lock);
2211 while ((afs_blocksUsed - afs_blocksDiscarded) >
2212 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2213 afs_WaitForCacheDrain = 1;
2214 afs_osi_Sleep(&afs_WaitForCacheDrain);
2216 afs_MaybeFreeDiscardedDCache();
2217 /* need to check if someone else got the chunk first. */
2218 goto RetryGetDCache;
2221 /* Do not fetch data beyond truncPos. */
2222 maxGoodLength = avc->m.Length;
2223 if (avc->truncPos < maxGoodLength)
2224 maxGoodLength = avc->truncPos;
2225 Position = AFS_CHUNKBASE(abyte);
2226 if (vType(avc) == VDIR) {
2227 size = avc->m.Length;
2228 if (size > tdc->f.chunkBytes) {
2229 /* pre-reserve space for file */
2230 afs_AdjustSize(tdc, size);
2232 size = 999999999; /* max size for transfer */
2234 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2235 /* don't read past end of good data on server */
2236 if (Position + size > maxGoodLength)
2237 size = maxGoodLength - Position;
2239 size = 0; /* Handle random races */
2240 if (size > tdc->f.chunkBytes) {
2241 /* pre-reserve space for file */
2242 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2243 /* max size for transfer still in size */
2246 if (afs_mariner && !tdc->f.chunk)
2247 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2249 * Right now, we only have one tool, and it's a hammer. So, we
2250 * fetch the whole file.
2252 DZap(tdc); /* pages in cache may be old */
2253 file = afs_CFileOpen(tdc->f.inode);
2254 afs_RemoveVCB(&avc->fid);
2255 tdc->f.states |= DWriting;
2256 tdc->dflags |= DFFetching;
2257 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2258 if (tdc->mflags & DFFetchReq) {
2259 tdc->mflags &= ~DFFetchReq;
2260 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2261 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2262 __FILE__, ICL_TYPE_INT32, __LINE__,
2263 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2267 (struct tlocal1 *)osi_AllocLargeSpace(sizeof(struct tlocal1));
2268 setVcacheStatus = 0;
2271 * Remember if we are doing the reading from a replicated volume,
2272 * and how many times we've zipped around the fetch/analyze loop.
2274 fromReplica = (avc->states & CRO) ? 1 : 0;
2276 accP = &(afs_stats_cmfullperf.accessinf);
2278 (accP->replicatedRefs)++;
2280 (accP->unreplicatedRefs)++;
2281 #endif /* AFS_NOSTATS */
2282 /* this is a cache miss */
2283 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2284 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
2285 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2288 afs_stats_cmperf.dcacheMisses++;
2291 * Dynamic root support: fetch data from local memory.
2293 if (afs_IsDynroot(avc)) {
2297 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2299 dynrootDir += Position;
2300 dynrootLen -= Position;
2301 if (size > dynrootLen)
2305 code = afs_CFileWrite(file, 0, dynrootDir, size);
2313 tdc->validPos = Position + size;
2314 afs_CFileTruncate(file, size); /* prune it */
2315 } else if (afs_IsDynrootMount(avc)) {
2319 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2321 dynrootDir += Position;
2322 dynrootLen -= Position;
2323 if (size > dynrootLen)
2327 code = afs_CFileWrite(file, 0, dynrootDir, size);
2335 tdc->validPos = Position + size;
2336 afs_CFileTruncate(file, size); /* prune it */
2339 * Not a dynamic vnode: do the real fetch.
2344 * avc->lock(R) if setLocks && !slowPass
2345 * avc->lock(W) if !setLocks || slowPass
2349 tc = afs_Conn(&avc->fid, areq, SHARED_LOCK);
2351 afs_int32 length_hi, length, bytes;
2355 (accP->numReplicasAccessed)++;
2357 #endif /* AFS_NOSTATS */
2358 if (!setLocks || slowPass) {
2359 avc->callback = tc->srvr->server;
2361 newCallback = tc->srvr->server;
2366 tcall = rx_NewCall(tc->id);
2369 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
2370 #ifdef AFS_64BIT_CLIENT
2371 length_hi = code = 0;
2372 if (!afs_serverHasNo64Bit(tc)) {
2376 StartRXAFS_FetchData64(tcall,
2377 (struct AFSFid *)&avc->fid.
2378 Fid, Position, tsize);
2381 afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
2382 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32,
2386 rx_Read(tcall, (char *)&length_hi,
2389 if (bytes == sizeof(afs_int32)) {
2390 length_hi = ntohl(length_hi);
2393 code = rx_Error(tcall);
2395 code1 = rx_EndCall(tcall, code);
2397 tcall = (struct rx_call *)0;
2401 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
2402 if (Position > 0x7FFFFFFF) {
2409 tcall = rx_NewCall(tc->id);
2411 StartRXAFS_FetchData(tcall, (struct AFSFid *)
2416 afs_serverSetNo64Bit(tc);
2421 rx_Read(tcall, (char *)&length,
2424 if (bytes == sizeof(afs_int32)) {
2425 length = ntohl(length);
2427 code = rx_Error(tcall);
2430 FillInt64(lengthFound, length_hi, length);
2431 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64LENG,
2432 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
2434 ICL_HANDLE_OFFSET(lengthFound));
2435 #else /* AFS_64BIT_CLIENT */
2438 StartRXAFS_FetchData(tcall,
2439 (struct AFSFid *)&avc->fid.Fid,
2445 rx_Read(tcall, (char *)&length,
2448 if (bytes == sizeof(afs_int32)) {
2449 length = ntohl(length);
2451 code = rx_Error(tcall);
2454 #endif /* AFS_64BIT_CLIENT */
2459 &(afs_stats_cmfullperf.rpc.
2460 fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
2461 osi_GetuTime(&xferStartTime);
2464 afs_CacheFetchProc(tcall, file,
2465 (afs_size_t) Position, tdc,
2467 &bytesXferred, length);
2469 osi_GetuTime(&xferStopTime);
2470 (xferP->numXfers)++;
2472 (xferP->numSuccesses)++;
2473 afs_stats_XferSumBytes
2474 [AFS_STATS_FS_XFERIDX_FETCHDATA] +=
2476 (xferP->sumBytes) +=
2477 (afs_stats_XferSumBytes
2478 [AFS_STATS_FS_XFERIDX_FETCHDATA] >> 10);
2479 afs_stats_XferSumBytes
2480 [AFS_STATS_FS_XFERIDX_FETCHDATA] &= 0x3FF;
2481 if (bytesXferred < xferP->minBytes)
2482 xferP->minBytes = bytesXferred;
2483 if (bytesXferred > xferP->maxBytes)
2484 xferP->maxBytes = bytesXferred;
2487 * Tally the size of the object. Note: we tally the actual size,
2488 * NOT the number of bytes that made it out over the wire.
2490 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
2491 (xferP->count[0])++;
2492 else if (bytesToXfer <=
2493 AFS_STATS_MAXBYTES_BUCKET1)
2494 (xferP->count[1])++;
2495 else if (bytesToXfer <=
2496 AFS_STATS_MAXBYTES_BUCKET2)
2497 (xferP->count[2])++;
2498 else if (bytesToXfer <=
2499 AFS_STATS_MAXBYTES_BUCKET3)
2500 (xferP->count[3])++;
2501 else if (bytesToXfer <=
2502 AFS_STATS_MAXBYTES_BUCKET4)
2503 (xferP->count[4])++;
2504 else if (bytesToXfer <=
2505 AFS_STATS_MAXBYTES_BUCKET5)
2506 (xferP->count[5])++;
2507 else if (bytesToXfer <=
2508 AFS_STATS_MAXBYTES_BUCKET6)
2509 (xferP->count[6])++;
2510 else if (bytesToXfer <=
2511 AFS_STATS_MAXBYTES_BUCKET7)
2512 (xferP->count[7])++;
2514 (xferP->count[8])++;
2516 afs_stats_GetDiff(elapsedTime, xferStartTime,
2518 afs_stats_AddTo((xferP->sumTime), elapsedTime);
2519 afs_stats_SquareAddTo((xferP->sqrTime),
2521 if (afs_stats_TimeLessThan
2522 (elapsedTime, (xferP->minTime))) {
2523 afs_stats_TimeAssign((xferP->minTime),
2526 if (afs_stats_TimeGreaterThan
2527 (elapsedTime, (xferP->maxTime))) {
2528 afs_stats_TimeAssign((xferP->maxTime),
2534 afs_CacheFetchProc(tcall, file, Position, tdc,
2536 #endif /* AFS_NOSTATS */
2541 EndRXAFS_FetchData(tcall, &tsmall->OutStatus,
2549 code1 = rx_EndCall(tcall, code);
2558 /* callback could have been broken (or expired) in a race here,
2559 * but we return the data anyway. It's as good as we knew about
2560 * when we started. */
2562 * validPos is updated by CacheFetchProc, and can only be
2563 * modifed under a dcache write lock, which we've blocked out
2565 size = tdc->validPos - Position; /* actual segment size */
2568 afs_CFileTruncate(file, size); /* prune it */
2570 if (!setLocks || slowPass) {
2571 ObtainWriteLock(&afs_xcbhash, 453);
2572 afs_DequeueCallback(avc);
2573 avc->states &= ~(CStatd | CUnique);
2574 avc->callback = NULL;
2575 ReleaseWriteLock(&afs_xcbhash);
2576 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2577 osi_dnlc_purgedp(avc);
2579 /* Something lost. Forget about performance, and go
2580 * back with a vcache write lock.
2582 afs_CFileTruncate(file, 0);
2583 afs_AdjustSize(tdc, 0);
2584 afs_CFileClose(file);
2585 osi_FreeLargeSpace(tsmall);
2587 ReleaseWriteLock(&tdc->lock);
2590 ReleaseReadLock(&avc->lock);
2592 goto RetryGetDCache;
2596 } while (afs_Analyze
2597 (tc, code, &avc->fid, areq,
2598 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2602 * avc->lock(R) if setLocks && !slowPass
2603 * avc->lock(W) if !setLocks || slowPass
2609 * In the case of replicated access, jot down info on the number of
2610 * attempts it took before we got through or gave up.
2613 if (numFetchLoops <= 1)
2614 (accP->refFirstReplicaOK)++;
2615 if (numFetchLoops > accP->maxReplicasPerRef)
2616 accP->maxReplicasPerRef = numFetchLoops;
2618 #endif /* AFS_NOSTATS */
2620 tdc->dflags &= ~DFFetching;
2621 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2622 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2623 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2624 tdc, ICL_TYPE_INT32, tdc->dflags);
2625 if (avc->execsOrWriters == 0)
2626 tdc->f.states &= ~DWriting;
2628 /* now, if code != 0, we have an error and should punt.
2629 * note that we have the vcache write lock, either because
2630 * !setLocks or slowPass.
2633 afs_CFileTruncate(file, 0);
2634 afs_AdjustSize(tdc, 0);
2635 afs_CFileClose(file);
2636 ZapDCE(tdc); /* sets DFEntryMod */
2637 if (vType(avc) == VDIR) {
2640 tdc->f.states &= ~(DRO|DBackup|DRW);
2641 afs_DCMoveBucket(tdc, 0, 0);
2642 ReleaseWriteLock(&tdc->lock);
2644 if (!afs_IsDynroot(avc)) {
2645 ObtainWriteLock(&afs_xcbhash, 454);
2646 afs_DequeueCallback(avc);
2647 avc->states &= ~(CStatd | CUnique);
2648 ReleaseWriteLock(&afs_xcbhash);
2649 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2650 osi_dnlc_purgedp(avc);
2653 * avc->lock(W); assert(!setLocks || slowPass)
2655 osi_Assert(!setLocks || slowPass);
2657 tdc->f.states &= ~(DRO|DBackup|DRW);
2658 afs_DCMoveBucket(tdc, 0, 0);
2663 /* otherwise we copy in the just-fetched info */
2664 afs_CFileClose(file);
2665 afs_AdjustSize(tdc, size); /* new size */
2667 * Copy appropriate fields into vcache. Status is
2668 * copied later where we selectively acquire the
2669 * vcache write lock.
2672 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2674 setVcacheStatus = 1;
2675 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2676 tsmall->OutStatus.DataVersion);
2677 tdc->dflags |= DFEntryMod;
2678 afs_indexFlags[tdc->index] |= IFEverUsed;
2679 ConvertWToSLock(&tdc->lock);
2680 } /*Data version numbers don't match */
2683 * Data version numbers match.
2685 afs_stats_cmperf.dcacheHits++;
2686 } /*Data version numbers match */
2688 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2692 * avc->lock(R) if setLocks && !slowPass
2693 * avc->lock(W) if !setLocks || slowPass
2694 * tdc->lock(S) if tdc
2698 * See if this was a reference to a file in the local cell.
2700 if (afs_IsPrimaryCellNum(avc->fid.Cell))
2701 afs_stats_cmperf.dlocalAccesses++;
2703 afs_stats_cmperf.dremoteAccesses++;
2705 /* Fix up LRU info */
2708 MObtainWriteLock(&afs_xdcache, 602);
2709 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2710 hadd32(afs_indexCounter, 1);
2711 MReleaseWriteLock(&afs_xdcache);
2713 /* return the data */
2714 if (vType(avc) == VDIR)
2717 *aoffset = AFS_CHUNKOFFSET(abyte);
2718 *alen = (tdc->f.chunkBytes - *aoffset);
2719 ReleaseSharedLock(&tdc->lock);
2724 * avc->lock(R) if setLocks && !slowPass
2725 * avc->lock(W) if !setLocks || slowPass
2728 /* Fix up the callback and status values in the vcache */
2730 if (setLocks && !slowPass) {
2733 * This is our dirty little secret to parallel fetches.
2734 * We don't write-lock the vcache while doing the fetch,
2735 * but potentially we'll need to update the vcache after
2736 * the fetch is done.
2738 * Drop the read lock and try to re-obtain the write
2739 * lock. If the vcache still has the same DV, it's
2740 * ok to go ahead and install the new data.
2742 afs_hyper_t currentDV, statusDV;
2744 hset(currentDV, avc->m.DataVersion);
2746 if (setNewCallback && avc->callback != newCallback)
2750 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2751 tsmall->OutStatus.DataVersion);
2753 if (setVcacheStatus && avc->m.Length != tsmall->OutStatus.Length)
2755 if (setVcacheStatus && !hsame(currentDV, statusDV))
2759 ReleaseReadLock(&avc->lock);
2761 if (doVcacheUpdate) {
2762 ObtainWriteLock(&avc->lock, 615);
2763 if (!hsame(avc->m.DataVersion, currentDV)) {
2764 /* We lose. Someone will beat us to it. */
2766 ReleaseWriteLock(&avc->lock);
2771 /* With slow pass, we've already done all the updates */
2773 ReleaseWriteLock(&avc->lock);
2776 /* Check if we need to perform any last-minute fixes with a write-lock */
2777 if (!setLocks || doVcacheUpdate) {
2779 avc->callback = newCallback;
2780 if (tsmall && setVcacheStatus)
2781 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2783 ReleaseWriteLock(&avc->lock);
2787 osi_FreeLargeSpace(tsmall);
2790 } /*afs_GetDCache */
2794 * afs_WriteThroughDSlots
2797 * Sweep through the dcache slots and write out any modified
2798 * in-memory data back on to our caching store.
2804 * The afs_xdcache is write-locked through this whole affair.
2807 afs_WriteThroughDSlots(void)
2809 register struct dcache *tdc;
2810 register afs_int32 i, touchedit = 0;
2812 struct afs_q DirtyQ, *tq;
2814 AFS_STATCNT(afs_WriteThroughDSlots);
2817 * Because of lock ordering, we can't grab dcache locks while
2818 * holding afs_xdcache. So we enter xdcache, get a reference
2819 * for every dcache entry, and exit xdcache.
2821 MObtainWriteLock(&afs_xdcache, 283);
2823 for (i = 0; i < afs_cacheFiles; i++) {
2824 tdc = afs_indexTable[i];
2826 /* Grab tlock in case the existing refcount isn't zero */
2827 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2828 ObtainWriteLock(&tdc->tlock, 623);
2830 ReleaseWriteLock(&tdc->tlock);
2832 QAdd(&DirtyQ, &tdc->dirty);
2835 MReleaseWriteLock(&afs_xdcache);
2838 * Now, for each dcache entry we found, check if it's dirty.
2839 * If so, get write-lock, get afs_xdcache, which protects
2840 * afs_cacheInodep, and flush it. Don't forget to put back
2844 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2846 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2848 if (tdc->dflags & DFEntryMod) {
2851 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2853 /* Now that we have the write lock, double-check */
2854 if (wrLock && (tdc->dflags & DFEntryMod)) {
2855 tdc->dflags &= ~DFEntryMod;
2856 MObtainWriteLock(&afs_xdcache, 620);
2857 afs_WriteDCache(tdc, 1);
2858 MReleaseWriteLock(&afs_xdcache);
2862 ReleaseWriteLock(&tdc->lock);
2868 MObtainWriteLock(&afs_xdcache, 617);
2869 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2870 /* Touch the file to make sure that the mtime on the file is kept
2871 * up-to-date to avoid losing cached files on cold starts because
2872 * their mtime seems old...
2874 struct afs_fheader theader;
2876 theader.magic = AFS_FHMAGIC;
2877 theader.firstCSize = AFS_FIRSTCSIZE;
2878 theader.otherCSize = AFS_OTHERCSIZE;
2879 theader.version = AFS_CI_VERSION;
2880 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2882 MReleaseWriteLock(&afs_xdcache);
2889 * Return a pointer to an freshly initialized dcache entry using
2890 * a memory-based cache. The tlock will be read-locked.
2893 * aslot : Dcache slot to look at.
2894 * tmpdc : Ptr to dcache entry.
2897 * Must be called with afs_xdcache write-locked.
2901 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2903 register struct dcache *tdc;
2906 AFS_STATCNT(afs_MemGetDSlot);
2907 if (CheckLock(&afs_xdcache) != -1)
2908 osi_Panic("getdslot nolock");
2909 if (aslot < 0 || aslot >= afs_cacheFiles)
2910 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2911 tdc = afs_indexTable[aslot];
2913 QRemove(&tdc->lruq); /* move to queue head */
2914 QAdd(&afs_DLRU, &tdc->lruq);
2915 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2916 ObtainWriteLock(&tdc->tlock, 624);
2918 ConvertWToRLock(&tdc->tlock);
2921 if (tmpdc == NULL) {
2922 if (!afs_freeDSList)
2923 afs_GetDownDSlot(4);
2924 if (!afs_freeDSList) {
2925 /* none free, making one is better than a panic */
2926 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2927 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
2928 #ifdef KERNEL_HAVE_PIN
2929 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2932 tdc = afs_freeDSList;
2933 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2936 tdc->dflags = 0; /* up-to-date, not in free q */
2938 QAdd(&afs_DLRU, &tdc->lruq);
2939 if (tdc->lruq.prev == &tdc->lruq)
2940 osi_Panic("lruq 3");
2946 /* initialize entry */
2947 tdc->f.fid.Cell = 0;
2948 tdc->f.fid.Fid.Volume = 0;
2950 hones(tdc->f.versionNo);
2951 tdc->f.inode = aslot;
2952 tdc->dflags |= DFEntryMod;
2955 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2958 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2959 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2960 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2963 RWLOCK_INIT(&tdc->lock, "dcache lock");
2964 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2965 RWLOCK_INIT(&tdc->mflock, "dcache flock");
2966 ObtainReadLock(&tdc->tlock);
2969 afs_indexTable[aslot] = tdc;
2972 } /*afs_MemGetDSlot */
2974 unsigned int last_error = 0, lasterrtime = 0;
2980 * Return a pointer to an freshly initialized dcache entry using
2981 * a UFS-based disk cache. The dcache tlock will be read-locked.
2984 * aslot : Dcache slot to look at.
2985 * tmpdc : Ptr to dcache entry.
2988 * afs_xdcache lock write-locked.
2991 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
2993 register afs_int32 code;
2994 register struct dcache *tdc;
2998 AFS_STATCNT(afs_UFSGetDSlot);
2999 if (CheckLock(&afs_xdcache) != -1)
3000 osi_Panic("getdslot nolock");
3001 if (aslot < 0 || aslot >= afs_cacheFiles)
3002 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
3003 tdc = afs_indexTable[aslot];
3005 QRemove(&tdc->lruq); /* move to queue head */
3006 QAdd(&afs_DLRU, &tdc->lruq);
3007 /* Grab tlock in case refCount != 0 */
3008 ObtainWriteLock(&tdc->tlock, 625);
3010 ConvertWToRLock(&tdc->tlock);
3013 /* otherwise we should read it in from the cache file */
3015 * If we weren't passed an in-memory region to place the file info,
3016 * we have to allocate one.
3018 if (tmpdc == NULL) {
3019 if (!afs_freeDSList)
3020 afs_GetDownDSlot(4);
3021 if (!afs_freeDSList) {
3022 /* none free, making one is better than a panic */
3023 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
3024 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
3025 #ifdef KERNEL_HAVE_PIN
3026 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
3029 tdc = afs_freeDSList;
3030 afs_freeDSList = (struct dcache *)tdc->lruq.next;
3033 tdc->dflags = 0; /* up-to-date, not in free q */
3035 QAdd(&afs_DLRU, &tdc->lruq);
3036 if (tdc->lruq.prev == &tdc->lruq)
3037 osi_Panic("lruq 3");
3044 * Seek to the aslot'th entry and read it in.
3047 afs_osi_Read(afs_cacheInodep,
3048 sizeof(struct fcache) * aslot +
3049 sizeof(struct afs_fheader), (char *)(&tdc->f),
3050 sizeof(struct fcache));
3052 if (code != sizeof(struct fcache))
3054 if (!afs_CellNumValid(tdc->f.fid.Cell))
3058 tdc->f.fid.Cell = 0;
3059 tdc->f.fid.Fid.Volume = 0;
3061 hones(tdc->f.versionNo);
3062 tdc->dflags |= DFEntryMod;
3063 #if defined(KERNEL_HAVE_UERROR)
3064 last_error = getuerror();
3066 lasterrtime = osi_Time();
3067 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
3068 tdc->f.states &= ~(DRO|DBackup|DRW);
3069 afs_DCMoveBucket(tdc, 0, 0);
3072 if (tdc->f.states & DRO) {
3073 afs_DCMoveBucket(tdc, 0, 2);
3074 } else if (tdc->f.states & DBackup) {
3075 afs_DCMoveBucket(tdc, 0, 1);
3077 afs_DCMoveBucket(tdc, 0, 1);
3083 if (tdc->f.chunk >= 0)
3084 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
3089 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
3090 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
3091 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
3094 RWLOCK_INIT(&tdc->lock, "dcache lock");
3095 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
3096 RWLOCK_INIT(&tdc->mflock, "dcache flock");
3097 ObtainReadLock(&tdc->tlock);
3100 * If we didn't read into a temporary dcache region, update the
3101 * slot pointer table.
3104 afs_indexTable[aslot] = tdc;
3107 } /*afs_UFSGetDSlot */
3115 * write a particular dcache entry back to its home in the
3119 * adc : Pointer to the dcache entry to write.
3120 * atime : If true, set the modtime on the file to the current time.
3123 * Must be called with the afs_xdcache lock at least read-locked,
3124 * and dcache entry at least read-locked.
3125 * The reference count is not changed.
3129 afs_WriteDCache(register struct dcache *adc, int atime)
3131 register afs_int32 code;
3133 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
3135 AFS_STATCNT(afs_WriteDCache);
3136 osi_Assert(WriteLocked(&afs_xdcache));
3138 adc->f.modTime = osi_Time();
3140 * Seek to the right dcache slot and write the in-memory image out to disk.
3142 afs_cellname_write();
3144 afs_osi_Write(afs_cacheInodep,
3145 sizeof(struct fcache) * adc->index +
3146 sizeof(struct afs_fheader), (char *)(&adc->f),
3147 sizeof(struct fcache));
3148 if (code != sizeof(struct fcache))
3159 * Wake up users of a particular file waiting for stores to take
3163 * avc : Ptr to related vcache entry.
3166 * Nothing interesting.
3170 afs_wakeup(register struct vcache *avc)
3173 register struct brequest *tb;
3175 AFS_STATCNT(afs_wakeup);
3176 for (i = 0; i < NBRS; i++, tb++) {
3177 /* if request is valid and for this file, we've found it */
3178 if (tb->refCount > 0 && avc == tb->vc) {
3181 * If CSafeStore is on, then we don't awaken the guy
3182 * waiting for the store until the whole store has finished.
3183 * Otherwise, we do it now. Note that if CSafeStore is on,
3184 * the BStore routine actually wakes up the user, instead
3186 * I think this is redundant now because this sort of thing
3187 * is already being handled by the higher-level code.
3189 if ((avc->states & CSafeStore) == 0) {
3191 tb->flags |= BUVALID;
3192 if (tb->flags & BUWAIT) {
3193 tb->flags &= ~BUWAIT;
3208 * Given a file name and inode, set up that file to be an
3209 * active member in the AFS cache. This also involves checking
3210 * the usability of its data.
3213 * afile : Name of the cache file to initialize.
3214 * ainode : Inode of the file.
3217 * This function is called only during initialization.
3221 afs_InitCacheFile(char *afile, ino_t ainode)
3223 register afs_int32 code;
3224 #if defined(AFS_LINUX22_ENV)
3225 struct dentry *filevp;
3227 struct vnode *filevp;
3231 struct osi_file *tfile;
3232 struct osi_stat tstat;
3233 register struct dcache *tdc;
3235 AFS_STATCNT(afs_InitCacheFile);
3236 index = afs_stats_cmperf.cacheNumEntries;
3237 if (index >= afs_cacheFiles)
3240 MObtainWriteLock(&afs_xdcache, 282);
3241 tdc = afs_GetDSlot(index, NULL);
3242 ReleaseReadLock(&tdc->tlock);
3243 MReleaseWriteLock(&afs_xdcache);
3245 ObtainWriteLock(&tdc->lock, 621);
3246 MObtainWriteLock(&afs_xdcache, 622);
3248 code = gop_lookupname(afile, AFS_UIOSYS, 0, &filevp);
3250 ReleaseWriteLock(&afs_xdcache);
3251 ReleaseWriteLock(&tdc->lock);
3256 * We have a VN_HOLD on filevp. Get the useful info out and
3257 * return. We make use of the fact that the cache is in the
3258 * UFS file system, and just record the inode number.
3260 #ifdef AFS_LINUX22_ENV
3261 tdc->f.inode = VTOI(filevp->d_inode)->i_number;
3264 tdc->f.inode = afs_vnodeToInumber(filevp);
3266 #endif /* AFS_LINUX22_ENV */
3268 tdc->f.inode = ainode;
3271 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3273 tfile = osi_UFSOpen(tdc->f.inode);
3274 code = afs_osi_Stat(tfile, &tstat);
3276 osi_Panic("initcachefile stat");
3279 * If file size doesn't match the cache info file, it's probably bad.
3281 if (tdc->f.chunkBytes != tstat.size)
3283 tdc->f.chunkBytes = 0;
3286 * If file changed within T (120?) seconds of cache info file, it's
3287 * probably bad. In addition, if slot changed within last T seconds,
3288 * the cache info file may be incorrectly identified, and so slot
3291 if (cacheInfoModTime < tstat.mtime + 120)
3293 if (cacheInfoModTime < tdc->f.modTime + 120)
3295 /* In case write through is behind, make sure cache items entry is
3296 * at least as new as the chunk.
3298 if (tdc->f.modTime < tstat.mtime)
3301 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3302 if (tstat.size != 0)
3303 osi_UFSTruncate(tfile, 0);
3304 tdc->f.states &= ~(DRO|DBackup|DRW);
3305 afs_DCMoveBucket(tdc, 0, 0);
3306 /* put entry in free cache slot list */
3307 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3308 afs_freeDCList = index;
3310 afs_indexFlags[index] |= IFFree;
3311 afs_indexUnique[index] = 0;
3314 * We must put this entry in the appropriate hash tables.
3315 * Note that i is still set from the above DCHash call
3317 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3318 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3319 afs_dchashTbl[code] = tdc->index;
3320 code = DVHash(&tdc->f.fid);
3321 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3322 afs_dvhashTbl[code] = tdc->index;
3323 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3325 /* has nontrivial amt of data */
3326 afs_indexFlags[index] |= IFEverUsed;
3327 afs_stats_cmperf.cacheFilesReused++;
3329 * Initialize index times to file's mod times; init indexCounter
3332 hset32(afs_indexTimes[index], tstat.atime);
3333 if (hgetlo(afs_indexCounter) < tstat.atime) {
3334 hset32(afs_indexCounter, tstat.atime);
3336 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3337 } /*File is not bad */
3339 osi_UFSClose(tfile);
3340 tdc->f.states &= ~DWriting;
3341 tdc->dflags &= ~DFEntryMod;
3342 /* don't set f.modTime; we're just cleaning up */
3343 afs_WriteDCache(tdc, 0);
3344 ReleaseWriteLock(&afs_xdcache);
3345 ReleaseWriteLock(&tdc->lock);
3347 afs_stats_cmperf.cacheNumEntries++;
3352 /*Max # of struct dcache's resident at any time*/
3354 * If 'dchint' is enabled then in-memory dcache min is increased because of
3363 * Initialize dcache related variables.
3366 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3368 register struct dcache *tdp;
3372 afs_freeDCList = NULLIDX;
3373 afs_discardDCList = NULLIDX;
3374 afs_freeDCCount = 0;
3375 afs_freeDSList = NULL;
3376 hzero(afs_indexCounter);
3378 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3384 if (achunk < 0 || achunk > 30)
3385 achunk = 13; /* Use default */
3386 AFS_SETCHUNKSIZE(achunk);
3392 if (aflags & AFSCALL_INIT_MEMCACHE) {
3394 * Use a memory cache instead of a disk cache
3396 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3397 afs_cacheType = &afs_MemCacheOps;
3398 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3399 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3400 /* ablocks is reported in 1K blocks */
3401 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3403 printf("afsd: memory cache too large for available memory.\n");
3404 printf("afsd: AFS files cannot be accessed.\n\n");
3406 afiles = ablocks = 0;
3408 printf("Memory cache: Allocating %d dcache entries...",
3411 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3412 afs_cacheType = &afs_UfsCacheOps;
3415 if (aDentries > 512)
3416 afs_dhashsize = 2048;
3417 /* initialize hash tables */
3419 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3421 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3422 for (i = 0; i < afs_dhashsize; i++) {
3423 afs_dvhashTbl[i] = NULLIDX;
3424 afs_dchashTbl[i] = NULLIDX;
3426 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3427 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3428 for (i = 0; i < afiles; i++) {
3429 afs_dvnextTbl[i] = NULLIDX;
3430 afs_dcnextTbl[i] = NULLIDX;
3433 /* Allocate and zero the pointer array to the dcache entries */
3434 afs_indexTable = (struct dcache **)
3435 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3436 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3438 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3439 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3441 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3442 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3443 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3444 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3446 /* Allocate and thread the struct dcache entries themselves */
3447 tdp = afs_Initial_freeDSList =
3448 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3449 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3450 #ifdef KERNEL_HAVE_PIN
3451 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3452 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3453 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3454 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3455 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3456 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3457 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3458 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3459 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3462 afs_freeDSList = &tdp[0];
3463 for (i = 0; i < aDentries - 1; i++) {
3464 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3465 RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3466 RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3467 RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3469 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3470 RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3471 RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3472 RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3474 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3475 afs_cacheBlocks = ablocks;
3476 afs_ComputeCacheParms(); /* compute parms based on cache size */
3478 afs_dcentries = aDentries;
3480 afs_stats_cmperf.cacheBucket0_Discarded =
3481 afs_stats_cmperf.cacheBucket1_Discarded =
3482 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3492 shutdown_dcache(void)
3496 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3497 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3498 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3499 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3500 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3501 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3502 afs_osi_Free(afs_Initial_freeDSList,
3503 afs_dcentries * sizeof(struct dcache));
3504 #ifdef KERNEL_HAVE_PIN
3505 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3506 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3507 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3508 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3509 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3510 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3511 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3515 for (i = 0; i < afs_dhashsize; i++) {
3516 afs_dvhashTbl[i] = NULLIDX;
3517 afs_dchashTbl[i] = NULLIDX;
3520 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3521 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3523 afs_blocksUsed = afs_dcentries = 0;
3524 afs_stats_cmperf.cacheBucket0_Discarded =
3525 afs_stats_cmperf.cacheBucket1_Discarded =
3526 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3527 hzero(afs_indexCounter);
3529 afs_freeDCCount = 0;
3530 afs_freeDCList = NULLIDX;
3531 afs_discardDCList = NULLIDX;
3532 afs_freeDSList = afs_Initial_freeDSList = 0;
3534 LOCK_INIT(&afs_xdcache, "afs_xdcache");