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);
38 * --------------------- Exported definitions ---------------------
41 afs_int32 afs_blocksUsed_0; /*!< 1K blocks in cache - in theory is zero */
42 afs_int32 afs_blocksUsed_1; /*!< 1K blocks in cache */
43 afs_int32 afs_blocksUsed_2; /*!< 1K blocks in cache */
44 afs_int32 afs_pct1 = -1;
45 afs_int32 afs_pct2 = -1;
46 afs_uint32 afs_tpct1 = 0;
47 afs_uint32 afs_tpct2 = 0;
48 afs_uint32 splitdcache = 0;
50 afs_lock_t afs_xdcache; /*!< Lock: alloc new disk cache entries */
51 afs_int32 afs_freeDCList; /*!< Free list for disk cache entries */
52 afs_int32 afs_freeDCCount; /*!< Count of elts in freeDCList */
53 afs_int32 afs_discardDCList; /*!< Discarded disk cache entries */
54 afs_int32 afs_discardDCCount; /*!< Count of elts in discardDCList */
55 struct dcache *afs_freeDSList; /*!< Free list for disk slots */
56 struct dcache *afs_Initial_freeDSList; /*!< Initial list for above */
57 #if defined(LINUX_USE_FH)
58 struct fid cacheitems_fh;
59 int cacheitems_fh_type;
61 ino_t cacheInode; /*!< Inode for CacheItems file */
63 struct osi_file *afs_cacheInodep = 0; /*!< file for CacheItems inode */
64 struct afs_q afs_DLRU; /*!< dcache LRU */
65 afs_int32 afs_dhashsize = 1024;
66 afs_int32 *afs_dvhashTbl; /*!< Data cache hash table: hashed by FID + chunk number. */
67 afs_int32 *afs_dchashTbl; /*!< Data cache hash table: hashed by FID. */
68 afs_int32 *afs_dvnextTbl; /*!< Dcache hash table links */
69 afs_int32 *afs_dcnextTbl; /*!< Dcache hash table links */
70 struct dcache **afs_indexTable; /*!< Pointers to dcache entries */
71 afs_hyper_t *afs_indexTimes; /*!< Dcache entry Access times */
72 afs_int32 *afs_indexUnique; /*!< dcache entry Fid.Unique */
73 unsigned char *afs_indexFlags; /*!< (only one) Is there data there? */
74 afs_hyper_t afs_indexCounter; /*!< Fake time for marking index
76 afs_int32 afs_cacheFiles = 0; /*!< Size of afs_indexTable */
77 afs_int32 afs_cacheBlocks; /*!< 1K blocks in cache */
78 afs_int32 afs_cacheStats; /*!< Stat entries in cache */
79 afs_int32 afs_blocksUsed; /*!< Number of blocks in use */
80 afs_int32 afs_blocksDiscarded; /*!<Blocks freed but not truncated */
81 afs_int32 afs_fsfragsize = 1023; /*!< Underlying Filesystem minimum unit
82 *of disk allocation usually 1K
83 *this value is (truefrag -1 ) to
84 *save a bunch of subtracts... */
85 #ifdef AFS_64BIT_CLIENT
86 #ifdef AFS_VM_RDWR_ENV
87 afs_size_t afs_vmMappingEnd; /* !< For large files (>= 2GB) the VM
88 * mapping an 32bit addressing machines
89 * can only be used below the 2 GB
90 * line. From this point upwards we
91 * must do direct I/O into the cache
92 * files. The value should be on a
94 #endif /* AFS_VM_RDWR_ENV */
95 #endif /* AFS_64BIT_CLIENT */
97 /* The following is used to ensure that new dcache's aren't obtained when
98 * the cache is nearly full.
100 int afs_WaitForCacheDrain = 0;
101 int afs_TruncateDaemonRunning = 0;
102 int afs_CacheTooFull = 0;
104 afs_int32 afs_dcentries; /*!< In-memory dcache entries */
107 int dcacheDisabled = 0;
109 static int afs_UFSCacheFetchProc(), afs_UFSCacheStoreProc();
110 struct afs_cacheOps afs_UfsCacheOps = {
111 #if defined(LINUX_USE_FH)
122 afs_UFSCacheFetchProc,
123 afs_UFSCacheStoreProc,
129 struct afs_cacheOps afs_MemCacheOps = {
131 afs_MemCacheTruncate,
137 afs_MemCacheFetchProc,
138 afs_MemCacheStoreProc,
144 int cacheDiskType; /*Type of backing disk for cache */
145 struct afs_cacheOps *afs_cacheType;
148 * Where is this vcache's entry associated dcache located/
149 * \param avc The vcache entry.
150 * \return Bucket index:
155 afs_DCGetBucket(struct vcache *avc)
160 /* This should be replaced with some sort of user configurable function */
161 if (avc->states & CRO) {
163 } else if (avc->states & CBackup) {
173 * Readjust a dcache's size.
175 * \param adc The dcache to be adjusted.
176 * \param oldSize Old size for the dcache.
177 * \param newSize The new size to be adjusted to.
181 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
183 afs_int32 adjustSize = newSize - oldSize;
191 afs_blocksUsed_0 += adjustSize;
192 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
195 afs_blocksUsed_1 += adjustSize;
196 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
199 afs_blocksUsed_2 += adjustSize;
200 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
208 * Move a dcache from one bucket to another.
210 * \param adc Operate on this dcache.
211 * \param size Size in bucket (?).
212 * \param newBucket Destination bucket.
216 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
221 /* Substract size from old bucket. */
225 afs_blocksUsed_0 -= size;
228 afs_blocksUsed_1 -= size;
231 afs_blocksUsed_2 -= size;
235 /* Set new bucket and increase destination bucket size. */
236 adc->bucket = newBucket;
241 afs_blocksUsed_0 += size;
244 afs_blocksUsed_1 += size;
247 afs_blocksUsed_2 += size;
255 * Init split caches size.
260 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
269 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
274 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
275 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
277 /* Short cut: if we don't know about it, try to kill it */
278 if (phase < 2 && afs_blocksUsed_0)
281 if (afs_pct1 > afs_tpct1)
283 if (afs_pct2 > afs_tpct2)
285 return 0; /* unlikely */
290 * Warn about failing to store a file.
292 * \param acode Associated error code.
293 * \param avolume Volume involved.
294 * \param aflags How to handle the output:
295 * aflags & 1: Print out on console
296 * aflags & 2: Print out on controlling tty
298 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
302 afs_StoreWarn(register afs_int32 acode, afs_int32 avolume,
303 register afs_int32 aflags)
305 static char problem_fmt[] =
306 "afs: failed to store file in volume %d (%s)\n";
307 static char problem_fmt_w_error[] =
308 "afs: failed to store file in volume %d (error %d)\n";
309 static char netproblems[] = "network problems";
310 static char partfull[] = "partition full";
311 static char overquota[] = "over quota";
313 AFS_STATCNT(afs_StoreWarn);
319 afs_warn(problem_fmt, avolume, netproblems);
321 afs_warnuser(problem_fmt, avolume, netproblems);
322 } else if (acode == ENOSPC) {
327 afs_warn(problem_fmt, avolume, partfull);
329 afs_warnuser(problem_fmt, avolume, partfull);
332 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
333 * Instead ENOSPC will be sent...
335 if (acode == EDQUOT) {
340 afs_warn(problem_fmt, avolume, overquota);
342 afs_warnuser(problem_fmt, avolume, overquota);
350 afs_warn(problem_fmt_w_error, avolume, acode);
352 afs_warnuser(problem_fmt_w_error, avolume, acode);
357 * Try waking up truncation daemon, if it's worth it.
360 afs_MaybeWakeupTruncateDaemon(void)
362 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
363 afs_CacheTooFull = 1;
364 if (!afs_TruncateDaemonRunning)
365 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
366 } else if (!afs_TruncateDaemonRunning
367 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
368 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
375 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
376 * struct so we need only export one symbol for AIX.
378 static struct CTD_stats {
379 osi_timeval_t CTD_beforeSleep;
380 osi_timeval_t CTD_afterSleep;
381 osi_timeval_t CTD_sleepTime;
382 osi_timeval_t CTD_runTime;
386 u_int afs_min_cache = 0;
389 * Keeps the cache clean and free by truncating uneeded files, when used.
394 afs_CacheTruncateDaemon(void)
396 osi_timeval_t CTD_tmpTime;
400 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
402 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
404 osi_GetuTime(&CTD_stats.CTD_afterSleep);
405 afs_TruncateDaemonRunning = 1;
407 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
408 MObtainWriteLock(&afs_xdcache, 266);
409 if (afs_CacheTooFull) {
410 int space_needed, slots_needed;
411 /* if we get woken up, we should try to clean something out */
412 for (counter = 0; counter < 10; counter++) {
414 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
416 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
417 afs_GetDownD(slots_needed, &space_needed, 0);
418 if ((space_needed <= 0) && (slots_needed <= 0)) {
421 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
424 if (!afs_CacheIsTooFull())
425 afs_CacheTooFull = 0;
426 } /* end of cache cleanup */
427 MReleaseWriteLock(&afs_xdcache);
430 * This is a defensive check to try to avoid starving threads
431 * that may need the global lock so thay can help free some
432 * cache space. If this thread won't be sleeping or truncating
433 * any cache files then give up the global lock so other
434 * threads get a chance to run.
436 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
437 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
438 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
442 * This is where we free the discarded cache elements.
444 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
445 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
446 afs_FreeDiscardedDCache();
449 /* See if we need to continue to run. Someone may have
450 * signalled us while we were executing.
452 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
453 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
454 /* Collect statistics on truncate daemon. */
455 CTD_stats.CTD_nSleeps++;
456 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
457 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
458 CTD_stats.CTD_beforeSleep);
459 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
461 afs_TruncateDaemonRunning = 0;
462 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
463 afs_TruncateDaemonRunning = 1;
465 osi_GetuTime(&CTD_stats.CTD_afterSleep);
466 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
467 CTD_stats.CTD_afterSleep);
468 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
470 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
472 afs_termState = AFSOP_STOP_AFSDB;
474 afs_termState = AFSOP_STOP_RXEVENT;
476 afs_osi_Wakeup(&afs_termState);
484 * Make adjustment for the new size in the disk cache entry
486 * \note Major Assumptions Here:
487 * Assumes that frag size is an integral power of two, less one,
488 * and that this is a two's complement machine. I don't
489 * know of any filesystems which violate this assumption...
491 * \param adc Ptr to dcache entry.
492 * \param anewsize New size desired.
497 afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
499 register afs_int32 oldSize;
501 AFS_STATCNT(afs_AdjustSize);
503 adc->dflags |= DFEntryMod;
504 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
505 adc->f.chunkBytes = newSize;
508 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
509 afs_DCAdjustSize(adc, oldSize, newSize);
510 if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {
512 /* We're growing the file, wakeup the daemon */
513 afs_MaybeWakeupTruncateDaemon();
515 afs_blocksUsed += (newSize - oldSize);
516 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
521 * This routine is responsible for moving at least one entry (but up
522 * to some number of them) from the LRU queue to the free queue.
524 * \param anumber Number of entries that should ideally be moved.
525 * \param aneedSpace How much space we need (1K blocks);
528 * The anumber parameter is just a hint; at least one entry MUST be
529 * moved, or we'll panic. We must be called with afs_xdcache
530 * write-locked. We should try to satisfy both anumber and aneedspace,
531 * whichever is more demanding - need to do several things:
532 * 1. only grab up to anumber victims if aneedSpace <= 0, not
533 * the whole set of MAXATONCE.
534 * 2. dynamically choose MAXATONCE to reflect severity of
535 * demand: something like (*aneedSpace >> (logChunk - 9))
537 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
538 * indicates that the cache is not properly configured/tuned or
539 * something. We should be able to automatically correct that problem.
542 #define MAXATONCE 16 /* max we can obtain at once */
544 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
548 struct VenusFid *afid;
552 register struct vcache *tvc;
553 afs_uint32 victims[MAXATONCE];
554 struct dcache *victimDCs[MAXATONCE];
555 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
556 afs_uint32 victimPtr; /* next free item in victim arrays */
557 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
558 afs_uint32 maxVictimPtr; /* where it is */
563 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
564 vfslocked = VFS_LOCK_GIANT(afs_globalVFS);
567 AFS_STATCNT(afs_GetDownD);
569 if (CheckLock(&afs_xdcache) != -1)
570 osi_Panic("getdownd nolock");
571 /* decrement anumber first for all dudes in free list */
572 /* SHOULD always decrement anumber first, even if aneedSpace >0,
573 * because we should try to free space even if anumber <=0 */
574 if (!aneedSpace || *aneedSpace <= 0) {
575 anumber -= afs_freeDCCount;
577 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
578 VFS_UNLOCK_GIANT(vfslocked);
580 return; /* enough already free */
584 /* bounds check parameter */
585 if (anumber > MAXATONCE)
586 anumber = MAXATONCE; /* all we can do */
588 /* rewrite so phases include a better eligiblity for gc test*/
590 * The phase variable manages reclaims. Set to 0, the first pass,
591 * we don't reclaim active entries, or other than target bucket.
592 * Set to 1, we reclaim even active ones in target bucket.
593 * Set to 2, we reclaim any inactive one.
594 * Set to 3, we reclaim even active ones.
602 for (i = 0; i < afs_cacheFiles; i++)
603 /* turn off all flags */
604 afs_indexFlags[i] &= ~IFFlag;
606 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
607 /* find oldest entries for reclamation */
608 maxVictimPtr = victimPtr = 0;
609 hzero(maxVictimTime);
610 curbucket = afs_DCWhichBucket(phase, buckethint);
611 /* select victims from access time array */
612 for (i = 0; i < afs_cacheFiles; i++) {
613 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
614 /* skip if dirty or already free */
617 tdc = afs_indexTable[i];
618 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
620 /* Wrong bucket; can't use it! */
623 if (tdc && (tdc->refCount != 0)) {
624 /* Referenced; can't use it! */
627 hset(vtime, afs_indexTimes[i]);
629 /* if we've already looked at this one, skip it */
630 if (afs_indexFlags[i] & IFFlag)
633 if (victimPtr < MAXATONCE) {
634 /* if there's at least one free victim slot left */
635 victims[victimPtr] = i;
636 hset(victimTimes[victimPtr], vtime);
637 if (hcmp(vtime, maxVictimTime) > 0) {
638 hset(maxVictimTime, vtime);
639 maxVictimPtr = victimPtr;
642 } else if (hcmp(vtime, maxVictimTime) < 0) {
644 * We're older than youngest victim, so we replace at
647 /* find youngest (largest LRU) victim */
650 osi_Panic("getdownd local");
652 hset(victimTimes[j], vtime);
653 /* recompute maxVictimTime */
654 hset(maxVictimTime, vtime);
655 for (j = 0; j < victimPtr; j++)
656 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
657 hset(maxVictimTime, victimTimes[j]);
663 /* now really reclaim the victims */
664 j = 0; /* flag to track if we actually got any of the victims */
665 /* first, hold all the victims, since we're going to release the lock
666 * during the truncate operation.
668 for (i = 0; i < victimPtr; i++) {
669 tdc = afs_GetDSlot(victims[i], 0);
670 /* We got tdc->tlock(R) here */
671 if (tdc->refCount == 1)
675 ReleaseReadLock(&tdc->tlock);
679 for (i = 0; i < victimPtr; i++) {
680 /* q is first elt in dcache entry */
682 /* now, since we're dropping the afs_xdcache lock below, we
683 * have to verify, before proceeding, that there are no other
684 * references to this dcache entry, even now. Note that we
685 * compare with 1, since we bumped it above when we called
686 * afs_GetDSlot to preserve the entry's identity.
688 if (tdc && tdc->refCount == 1) {
689 unsigned char chunkFlags;
690 afs_size_t tchunkoffset = 0;
692 /* xdcache is lower than the xvcache lock */
693 MReleaseWriteLock(&afs_xdcache);
694 MObtainReadLock(&afs_xvcache);
695 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
696 MReleaseReadLock(&afs_xvcache);
697 MObtainWriteLock(&afs_xdcache, 527);
699 if (tdc->refCount > 1)
702 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
703 chunkFlags = afs_indexFlags[tdc->index];
704 if (((phase & 1) == 0) && osi_Active(tvc))
706 if (((phase & 1) == 1) && osi_Active(tvc)
707 && (tvc->states & CDCLock)
708 && (chunkFlags & IFAnyPages))
710 if (chunkFlags & IFDataMod)
712 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
713 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
714 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
715 ICL_HANDLE_OFFSET(tchunkoffset));
717 #if defined(AFS_SUN5_ENV)
719 * Now we try to invalidate pages. We do this only for
720 * Solaris. For other platforms, it's OK to recycle a
721 * dcache entry out from under a page, because the strategy
722 * function can call afs_GetDCache().
724 if (!skip && (chunkFlags & IFAnyPages)) {
727 MReleaseWriteLock(&afs_xdcache);
728 MObtainWriteLock(&tvc->vlock, 543);
729 if (tvc->multiPage) {
733 /* block locking pages */
734 tvc->vstates |= VPageCleaning;
735 /* block getting new pages */
737 MReleaseWriteLock(&tvc->vlock);
738 /* One last recheck */
739 MObtainWriteLock(&afs_xdcache, 333);
740 chunkFlags = afs_indexFlags[tdc->index];
741 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
742 || (osi_Active(tvc) && (tvc->states & CDCLock)
743 && (chunkFlags & IFAnyPages))) {
745 MReleaseWriteLock(&afs_xdcache);
748 MReleaseWriteLock(&afs_xdcache);
750 code = osi_VM_GetDownD(tvc, tdc);
752 MObtainWriteLock(&afs_xdcache, 269);
753 /* we actually removed all pages, clean and dirty */
755 afs_indexFlags[tdc->index] &=
756 ~(IFDirtyPages | IFAnyPages);
759 MReleaseWriteLock(&afs_xdcache);
761 MObtainWriteLock(&tvc->vlock, 544);
762 if (--tvc->activeV == 0
763 && (tvc->vstates & VRevokeWait)) {
764 tvc->vstates &= ~VRevokeWait;
765 afs_osi_Wakeup((char *)&tvc->vstates);
768 if (tvc->vstates & VPageCleaning) {
769 tvc->vstates &= ~VPageCleaning;
770 afs_osi_Wakeup((char *)&tvc->vstates);
773 MReleaseWriteLock(&tvc->vlock);
775 #endif /* AFS_SUN5_ENV */
777 MReleaseWriteLock(&afs_xdcache);
780 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
781 MObtainWriteLock(&afs_xdcache, 528);
782 if (afs_indexFlags[tdc->index] &
783 (IFDataMod | IFDirtyPages | IFAnyPages))
785 if (tdc->refCount > 1)
788 #if defined(AFS_SUN5_ENV)
790 /* no vnode, so IFDirtyPages is spurious (we don't
791 * sweep dcaches on vnode recycling, so we can have
792 * DIRTYPAGES set even when all pages are gone). Just
794 * Hold vcache lock to prevent vnode from being
795 * created while we're clearing IFDirtyPages.
797 afs_indexFlags[tdc->index] &=
798 ~(IFDirtyPages | IFAnyPages);
802 /* skip this guy and mark him as recently used */
803 afs_indexFlags[tdc->index] |= IFFlag;
804 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
805 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
806 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
807 ICL_HANDLE_OFFSET(tchunkoffset));
809 /* flush this dude from the data cache and reclaim;
810 * first, make sure no one will care that we damage
811 * it, by removing it from all hash tables. Then,
812 * melt it down for parts. Note that any concurrent
813 * (new possibility!) calls to GetDownD won't touch
814 * this guy because his reference count is > 0. */
815 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
816 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
817 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
818 ICL_HANDLE_OFFSET(tchunkoffset));
819 AFS_STATCNT(afs_gget);
820 afs_HashOutDCache(tdc, 1);
821 if (tdc->f.chunkBytes != 0) {
825 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
830 afs_DiscardDCache(tdc);
835 j = 1; /* we reclaimed at least one victim */
839 } /* end of for victims loop */
842 /* Phase is 0 and no one was found, so try phase 1 (ignore
843 * osi_Active flag) */
846 for (i = 0; i < afs_cacheFiles; i++)
847 /* turn off all flags */
848 afs_indexFlags[i] &= ~IFFlag;
851 /* found no one in phases 0-5, we're hosed */
855 } /* big while loop */
857 #if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
858 VFS_UNLOCK_GIANT(vfslocked);
867 * Remove adc from any hash tables that would allow it to be located
868 * again by afs_FindDCache or afs_GetDCache.
870 * \param adc Pointer to dcache entry to remove from hash tables.
872 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
876 afs_HashOutDCache(struct dcache *adc, int zap)
880 AFS_STATCNT(afs_glink);
882 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
884 /* if this guy is in the hash table, pull him out */
885 if (adc->f.fid.Fid.Volume != 0) {
886 /* remove entry from first hash chains */
887 i = DCHash(&adc->f.fid, adc->f.chunk);
888 us = afs_dchashTbl[i];
889 if (us == adc->index) {
890 /* first dude in the list */
891 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
893 /* somewhere on the chain */
894 while (us != NULLIDX) {
895 if (afs_dcnextTbl[us] == adc->index) {
896 /* found item pointing at the one to delete */
897 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
900 us = afs_dcnextTbl[us];
903 osi_Panic("dcache hc");
905 /* remove entry from *other* hash chain */
906 i = DVHash(&adc->f.fid);
907 us = afs_dvhashTbl[i];
908 if (us == adc->index) {
909 /* first dude in the list */
910 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
912 /* somewhere on the chain */
913 while (us != NULLIDX) {
914 if (afs_dvnextTbl[us] == adc->index) {
915 /* found item pointing at the one to delete */
916 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
919 us = afs_dvnextTbl[us];
922 osi_Panic("dcache hv");
927 /* prevent entry from being found on a reboot (it is already out of
928 * the hash table, but after a crash, we just look at fid fields of
929 * stable (old) entries).
931 adc->f.fid.Fid.Volume = 0; /* invalid */
933 /* mark entry as modified */
934 adc->dflags |= DFEntryMod;
939 } /*afs_HashOutDCache */
942 * Flush the given dcache entry, pulling it from hash chains
943 * and truncating the associated cache file.
945 * \param adc Ptr to dcache entry to flush.
948 * This routine must be called with the afs_xdcache lock held
952 afs_FlushDCache(register struct dcache *adc)
954 AFS_STATCNT(afs_FlushDCache);
956 * Bump the number of cache files flushed.
958 afs_stats_cmperf.cacheFlushes++;
960 /* remove from all hash tables */
961 afs_HashOutDCache(adc, 1);
963 /* Free its space; special case null operation, since truncate operation
964 * in UFS is slow even in this case, and this allows us to pre-truncate
965 * these files at more convenient times with fewer locks set
966 * (see afs_GetDownD).
968 if (adc->f.chunkBytes != 0) {
969 afs_DiscardDCache(adc);
970 afs_MaybeWakeupTruncateDaemon();
975 if (afs_WaitForCacheDrain) {
976 if (afs_blocksUsed <=
977 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
978 afs_WaitForCacheDrain = 0;
979 afs_osi_Wakeup(&afs_WaitForCacheDrain);
982 } /*afs_FlushDCache */
986 * Put a dcache entry on the free dcache entry list.
988 * \param adc dcache entry to free.
990 * \note Environment: called with afs_xdcache lock write-locked.
993 afs_FreeDCache(register struct dcache *adc)
995 /* Thread on free list, update free list count and mark entry as
996 * freed in its indexFlags element. Also, ensure DCache entry gets
997 * written out (set DFEntryMod).
1000 afs_dvnextTbl[adc->index] = afs_freeDCList;
1001 afs_freeDCList = adc->index;
1003 afs_indexFlags[adc->index] |= IFFree;
1004 adc->dflags |= DFEntryMod;
1006 if (afs_WaitForCacheDrain) {
1007 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1008 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1009 afs_WaitForCacheDrain = 0;
1010 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1013 } /* afs_FreeDCache */
1016 * Discard the cache element by moving it to the discardDCList.
1017 * This puts the cache element into a quasi-freed state, where
1018 * the space may be reused, but the file has not been truncated.
1020 * \note Major Assumptions Here:
1021 * Assumes that frag size is an integral power of two, less one,
1022 * and that this is a two's complement machine. I don't
1023 * know of any filesystems which violate this assumption...
1025 * \param adr Ptr to dcache entry.
1027 * \note Environment:
1028 * Must be called with afs_xdcache write-locked.
1032 afs_DiscardDCache(register struct dcache *adc)
1034 register afs_int32 size;
1036 AFS_STATCNT(afs_DiscardDCache);
1038 osi_Assert(adc->refCount == 1);
1040 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1041 afs_blocksDiscarded += size;
1042 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1044 afs_dvnextTbl[adc->index] = afs_discardDCList;
1045 afs_discardDCList = adc->index;
1046 afs_discardDCCount++;
1048 adc->f.fid.Fid.Volume = 0;
1049 adc->dflags |= DFEntryMod;
1050 afs_indexFlags[adc->index] |= IFDiscarded;
1052 if (afs_WaitForCacheDrain) {
1053 if ((afs_blocksUsed - afs_blocksDiscarded) <=
1054 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
1055 afs_WaitForCacheDrain = 0;
1056 afs_osi_Wakeup(&afs_WaitForCacheDrain);
1060 } /*afs_DiscardDCache */
1063 * Free the next element on the list of discarded cache elements.
1066 afs_FreeDiscardedDCache(void)
1068 register struct dcache *tdc;
1069 register struct osi_file *tfile;
1070 register afs_int32 size;
1072 AFS_STATCNT(afs_FreeDiscardedDCache);
1074 MObtainWriteLock(&afs_xdcache, 510);
1075 if (!afs_blocksDiscarded) {
1076 MReleaseWriteLock(&afs_xdcache);
1081 * Get an entry from the list of discarded cache elements
1083 tdc = afs_GetDSlot(afs_discardDCList, 0);
1084 osi_Assert(tdc->refCount == 1);
1085 ReleaseReadLock(&tdc->tlock);
1087 afs_discardDCList = afs_dvnextTbl[tdc->index];
1088 afs_dvnextTbl[tdc->index] = NULLIDX;
1089 afs_discardDCCount--;
1090 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1091 afs_blocksDiscarded -= size;
1092 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1093 /* We can lock because we just took it off the free list */
1094 ObtainWriteLock(&tdc->lock, 626);
1095 MReleaseWriteLock(&afs_xdcache);
1098 * Truncate the element to reclaim its space
1100 #if defined(LINUX_USE_FH)
1101 tfile = afs_CFileOpen(&tdc->f.fh, tdc->f.fh_type);
1103 tfile = afs_CFileOpen(tdc->f.inode);
1105 afs_CFileTruncate(tfile, 0);
1106 afs_CFileClose(tfile);
1107 afs_AdjustSize(tdc, 0);
1108 afs_DCMoveBucket(tdc, 0, 0);
1111 * Free the element we just truncated
1113 MObtainWriteLock(&afs_xdcache, 511);
1114 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1115 afs_FreeDCache(tdc);
1116 tdc->f.states &= ~(DRO|DBackup|DRW);
1117 ReleaseWriteLock(&tdc->lock);
1119 MReleaseWriteLock(&afs_xdcache);
1123 * Free as many entries from the list of discarded cache elements
1124 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1129 afs_MaybeFreeDiscardedDCache(void)
1132 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1134 while (afs_blocksDiscarded
1135 && (afs_blocksUsed >
1136 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1137 afs_FreeDiscardedDCache();
1143 * Try to free up a certain number of disk slots.
1145 * \param anumber Targeted number of disk slots to free up.
1147 * \note Environment:
1148 * Must be called with afs_xdcache write-locked.
1152 afs_GetDownDSlot(int anumber)
1154 struct afs_q *tq, *nq;
1159 AFS_STATCNT(afs_GetDownDSlot);
1160 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1161 osi_Panic("diskless getdowndslot");
1163 if (CheckLock(&afs_xdcache) != -1)
1164 osi_Panic("getdowndslot nolock");
1166 /* decrement anumber first for all dudes in free list */
1167 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1170 return; /* enough already free */
1172 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1174 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1175 nq = QPrev(tq); /* in case we remove it */
1176 if (tdc->refCount == 0) {
1177 if ((ix = tdc->index) == NULLIDX)
1178 osi_Panic("getdowndslot");
1179 /* pull the entry out of the lruq and put it on the free list */
1180 QRemove(&tdc->lruq);
1182 /* write-through if modified */
1183 if (tdc->dflags & DFEntryMod) {
1184 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1186 * ask proxy to do this for us - we don't have the stack space
1188 while (tdc->dflags & DFEntryMod) {
1191 s = SPLOCK(afs_sgibklock);
1192 if (afs_sgibklist == NULL) {
1193 /* if slot is free, grab it. */
1194 afs_sgibklist = tdc;
1195 SV_SIGNAL(&afs_sgibksync);
1197 /* wait for daemon to (start, then) finish. */
1198 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1202 tdc->dflags &= ~DFEntryMod;
1203 afs_WriteDCache(tdc, 1);
1207 /* finally put the entry in the free list */
1208 afs_indexTable[ix] = NULL;
1209 afs_indexFlags[ix] &= ~IFEverUsed;
1210 tdc->index = NULLIDX;
1211 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1212 afs_freeDSList = tdc;
1216 } /*afs_GetDownDSlot */
1223 * Increment the reference count on a disk cache entry,
1224 * which already has a non-zero refcount. In order to
1225 * increment the refcount of a zero-reference entry, you
1226 * have to hold afs_xdcache.
1229 * adc : Pointer to the dcache entry to increment.
1232 * Nothing interesting.
1235 afs_RefDCache(struct dcache *adc)
1237 ObtainWriteLock(&adc->tlock, 627);
1238 if (adc->refCount < 0)
1239 osi_Panic("RefDCache: negative refcount");
1241 ReleaseWriteLock(&adc->tlock);
1250 * Decrement the reference count on a disk cache entry.
1253 * ad : Ptr to the dcache entry to decrement.
1256 * Nothing interesting.
1259 afs_PutDCache(register struct dcache *adc)
1261 AFS_STATCNT(afs_PutDCache);
1262 ObtainWriteLock(&adc->tlock, 276);
1263 if (adc->refCount <= 0)
1264 osi_Panic("putdcache");
1266 ReleaseWriteLock(&adc->tlock);
1275 * Try to discard all data associated with this file from the
1279 * avc : Pointer to the cache info for the file.
1282 * Both pvnLock and lock are write held.
1285 afs_TryToSmush(register struct vcache *avc, struct AFS_UCRED *acred, int sync)
1287 register struct dcache *tdc;
1290 AFS_STATCNT(afs_TryToSmush);
1291 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1292 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length));
1293 sync = 1; /* XX Temp testing XX */
1295 #if defined(AFS_SUN5_ENV)
1296 ObtainWriteLock(&avc->vlock, 573);
1297 avc->activeV++; /* block new getpages */
1298 ReleaseWriteLock(&avc->vlock);
1301 /* Flush VM pages */
1302 osi_VM_TryToSmush(avc, acred, sync);
1305 * Get the hash chain containing all dce's for this fid
1307 i = DVHash(&avc->fid);
1308 MObtainWriteLock(&afs_xdcache, 277);
1309 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1310 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1311 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1312 int releaseTlock = 1;
1313 tdc = afs_GetDSlot(index, NULL);
1314 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1316 if ((afs_indexFlags[index] & IFDataMod) == 0
1317 && tdc->refCount == 1) {
1318 ReleaseReadLock(&tdc->tlock);
1320 afs_FlushDCache(tdc);
1323 afs_indexTable[index] = 0;
1326 ReleaseReadLock(&tdc->tlock);
1330 #if defined(AFS_SUN5_ENV)
1331 ObtainWriteLock(&avc->vlock, 545);
1332 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1333 avc->vstates &= ~VRevokeWait;
1334 afs_osi_Wakeup((char *)&avc->vstates);
1336 ReleaseWriteLock(&avc->vlock);
1338 MReleaseWriteLock(&afs_xdcache);
1340 * It's treated like a callback so that when we do lookups we'll
1341 * invalidate the unique bit if any
1342 * trytoSmush occured during the lookup call
1348 * afs_DCacheMissingChunks
1351 * Given the cached info for a file, return the number of chunks that
1352 * are not available from the dcache.
1355 * avc: Pointer to the (held) vcache entry to look in.
1358 * The number of chunks which are not currently cached.
1361 * The vcache entry is held upon entry.
1365 afs_DCacheMissingChunks(struct vcache *avc)
1368 afs_size_t totalLength = 0;
1369 afs_uint32 totalChunks = 0;
1372 totalLength = avc->m.Length;
1373 if (avc->truncPos < totalLength)
1374 totalLength = avc->truncPos;
1376 /* Length is 0, no chunk missing. */
1377 if (totalLength == 0)
1380 /* If totalLength is a multiple of chunksize, the last byte appears
1381 * as being part of the next chunk, which does not exist.
1382 * Decrementing totalLength by one fixes that.
1385 totalChunks = (AFS_CHUNK(totalLength) + 1);
1388 printf("Should have %d chunks for %u bytes\n",
1389 totalChunks, (totalLength + 1));
1391 i = DVHash(&avc->fid);
1392 MObtainWriteLock(&afs_xdcache, 1001);
1393 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1394 i = afs_dvnextTbl[index];
1395 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1396 tdc = afs_GetDSlot(index, NULL);
1397 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
1400 ReleaseReadLock(&tdc->tlock);
1404 MReleaseWriteLock(&afs_xdcache);
1406 /*printf("Missing %d chunks\n", totalChunks);*/
1408 return (totalChunks);
1415 * Given the cached info for a file and a byte offset into the
1416 * file, make sure the dcache entry for that file and containing
1417 * the given byte is available, returning it to our caller.
1420 * avc : Pointer to the (held) vcache entry to look in.
1421 * abyte : Which byte we want to get to.
1424 * Pointer to the dcache entry covering the file & desired byte,
1425 * or NULL if not found.
1428 * The vcache entry is held upon entry.
1432 afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
1435 register afs_int32 i, index;
1436 register struct dcache *tdc = NULL;
1438 AFS_STATCNT(afs_FindDCache);
1439 chunk = AFS_CHUNK(abyte);
1442 * Hash on the [fid, chunk] and get the corresponding dcache index
1443 * after write-locking the dcache.
1445 i = DCHash(&avc->fid, chunk);
1446 MObtainWriteLock(&afs_xdcache, 278);
1447 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1448 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1449 tdc = afs_GetDSlot(index, NULL);
1450 ReleaseReadLock(&tdc->tlock);
1451 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
1452 break; /* leaving refCount high for caller */
1456 index = afs_dcnextTbl[index];
1458 if (index != NULLIDX) {
1459 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1460 hadd32(afs_indexCounter, 1);
1461 MReleaseWriteLock(&afs_xdcache);
1464 MReleaseWriteLock(&afs_xdcache);
1466 } /*afs_FindDCache */
1470 * afs_UFSCacheStoreProc
1473 * Called upon store.
1476 * acall : Ptr to the Rx call structure involved.
1477 * afile : Ptr to the related file descriptor.
1478 * alen : Size of the file in bytes.
1479 * avc : Ptr to the vcache entry.
1480 * shouldWake : is it "safe" to return early from close() ?
1481 * abytesToXferP : Set to the number of bytes to xfer.
1482 * NOTE: This parameter is only used if AFS_NOSTATS
1484 * abytesXferredP : Set to the number of bytes actually xferred.
1485 * NOTE: This parameter is only used if AFS_NOSTATS
1489 * Nothing interesting.
1492 afs_UFSCacheStoreProc(register struct rx_call *acall, struct osi_file *afile,
1493 register afs_int32 alen, struct vcache *avc,
1494 int *shouldWake, afs_size_t * abytesToXferP,
1495 afs_size_t * abytesXferredP)
1497 afs_int32 code, got;
1498 register char *tbuffer;
1501 AFS_STATCNT(UFS_CacheStoreProc);
1505 * In this case, alen is *always* the amount of data we'll be trying
1508 (*abytesToXferP) = alen;
1509 (*abytesXferredP) = 0;
1510 #endif /* AFS_NOSTATS */
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 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1517 tlen = (alen > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : alen);
1518 got = afs_osi_Read(afile, -1, tbuffer, tlen);
1520 #if defined(KERNEL_HAVE_UERROR)
1521 || (got != tlen && getuerror())
1524 osi_FreeLargeSpace(tbuffer);
1527 afs_Trace2(afs_iclSetp, CM_TRACE_STOREPROC2, ICL_TYPE_OFFSET,
1528 ICL_HANDLE_OFFSET(*tbuffer), ICL_TYPE_INT32, got);
1530 code = rx_Write(acall, tbuffer, got); /* writing 0 bytes will
1531 * push a short packet. Is that really what we want, just because the
1532 * data didn't come back from the disk yet? Let's try it and see. */
1535 (*abytesXferredP) += code;
1536 #endif /* AFS_NOSTATS */
1538 code = rx_Error(acall);
1539 osi_FreeLargeSpace(tbuffer);
1540 return code ? code : -33;
1544 * If file has been locked on server, we can allow the store
1547 if (shouldWake && *shouldWake && (rx_GetRemoteStatus(acall) & 1)) {
1548 *shouldWake = 0; /* only do this once */
1552 afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
1553 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
1554 ICL_HANDLE_OFFSET(avc->m.Length), ICL_TYPE_INT32, alen);
1555 osi_FreeLargeSpace(tbuffer);
1558 } /* afs_UFSCacheStoreProc */
1562 * afs_UFSCacheFetchProc
1565 * Routine called on fetch; also tells people waiting for data
1566 * that more has arrived.
1569 * acall : Ptr to the Rx call structure.
1570 * afile : File descriptor for the cache file.
1571 * abase : Base offset to fetch.
1572 * adc : Ptr to the dcache entry for the file, write-locked.
1573 * avc : Ptr to the vcache entry for the file.
1574 * abytesToXferP : Set to the number of bytes to xfer.
1575 * NOTE: This parameter is only used if AFS_NOSTATS
1577 * abytesXferredP : Set to the number of bytes actually xferred.
1578 * NOTE: This parameter is only used if AFS_NOSTATS
1582 * Nothing interesting.
1586 afs_UFSCacheFetchProc(register struct rx_call *acall, struct osi_file *afile,
1587 afs_size_t abase, struct dcache *adc,
1588 struct vcache *avc, afs_size_t * abytesToXferP,
1589 afs_size_t * abytesXferredP, afs_int32 lengthFound)
1592 register afs_int32 code;
1593 register char *tbuffer;
1597 AFS_STATCNT(UFS_CacheFetchProc);
1598 osi_Assert(WriteLocked(&adc->lock));
1599 afile->offset = 0; /* Each time start from the beginning */
1600 length = lengthFound;
1602 (*abytesToXferP) = 0;
1603 (*abytesXferredP) = 0;
1604 #endif /* AFS_NOSTATS */
1605 tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
1606 adc->validPos = abase;
1610 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
1612 length = ntohl(length);
1613 if (code != sizeof(afs_int32)) {
1614 osi_FreeLargeSpace(tbuffer);
1615 code = rx_Error(acall);
1616 return (code ? code : -1); /* try to return code, not -1 */
1620 * The fetch protocol is extended for the AFS/DFS translator
1621 * to allow multiple blocks of data, each with its own length,
1622 * to be returned. As long as the top bit is set, there are more
1625 * We do not do this for AFS file servers because they sometimes
1626 * return large negative numbers as the transfer size.
1628 if (avc->states & CForeign) {
1629 moredata = length & 0x80000000;
1630 length &= ~0x80000000;
1635 (*abytesToXferP) += length;
1636 #endif /* AFS_NOSTATS */
1637 while (length > 0) {
1638 tlen = (length > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : length);
1639 #ifdef RX_KERNEL_TRACE
1640 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1644 code = rx_Read(acall, tbuffer, tlen);
1646 #ifdef RX_KERNEL_TRACE
1647 afs_Trace1(afs_iclSetp, CM_TRACE_TIMESTAMP, ICL_TYPE_STRING,
1651 (*abytesXferredP) += code;
1652 #endif /* AFS_NOSTATS */
1654 osi_FreeLargeSpace(tbuffer);
1655 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64READ,
1656 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
1657 ICL_TYPE_INT32, length);
1660 code = afs_osi_Write(afile, -1, tbuffer, tlen);
1662 osi_FreeLargeSpace(tbuffer);
1667 adc->validPos = abase;
1668 if (afs_osi_Wakeup(&adc->validPos) == 0)
1669 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
1670 __FILE__, ICL_TYPE_INT32, __LINE__,
1671 ICL_TYPE_POINTER, adc, ICL_TYPE_INT32,
1675 osi_FreeLargeSpace(tbuffer);
1678 } /* afs_UFSCacheFetchProc */
1681 * Get a fresh dcache from the free or discarded list.
1683 * \param avc Who's dcache is this going to be?
1684 * \param chunk The position where it will be placed in.
1685 * \param lock How are locks held.
1686 * \param ashFid If this dcache going to be used for a shadow dir,
1689 * \note Required locks:
1691 * - avc (R if (lock & 1) set and W otherwise)
1692 * \note It write locks the new dcache. The caller must unlock it.
1694 * \return The new dcache.
1696 struct dcache *afs_AllocDCache(struct vcache *avc,
1699 struct VenusFid *ashFid)
1701 struct dcache *tdc = NULL;
1702 afs_uint32 size = 0;
1703 struct osi_file *file;
1705 if (afs_discardDCList == NULLIDX
1706 || ((lock & 2) && afs_freeDCList != NULLIDX)) {
1708 afs_indexFlags[afs_freeDCList] &= ~IFFree;
1709 tdc = afs_GetDSlot(afs_freeDCList, 0);
1710 osi_Assert(tdc->refCount == 1);
1711 ReleaseReadLock(&tdc->tlock);
1712 ObtainWriteLock(&tdc->lock, 604);
1713 afs_freeDCList = afs_dvnextTbl[tdc->index];
1716 afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
1717 tdc = afs_GetDSlot(afs_discardDCList, 0);
1718 osi_Assert(tdc->refCount == 1);
1719 ReleaseReadLock(&tdc->tlock);
1720 ObtainWriteLock(&tdc->lock, 605);
1721 afs_discardDCList = afs_dvnextTbl[tdc->index];
1722 afs_discardDCCount--;
1724 ((tdc->f.chunkBytes +
1725 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1726 tdc->f.states &= ~(DRO|DBackup|DRW);
1727 afs_DCMoveBucket(tdc, size, 0);
1728 afs_blocksDiscarded -= size;
1729 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1731 /* Truncate the chunk so zeroes get filled properly */
1732 #if defined(LINUX_USE_FH)
1733 file = afs_CFileOpen(&tdc->f.fh, tdc->f.fh_type);
1735 file = afs_CFileOpen(tdc->f.inode);
1737 afs_CFileTruncate(file, 0);
1738 afs_CFileClose(file);
1739 afs_AdjustSize(tdc, 0);
1745 * avc->lock(R) if setLocks
1746 * avc->lock(W) if !setLocks
1752 * Fill in the newly-allocated dcache record.
1754 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1756 /* Use shadow fid if provided. */
1757 tdc->f.fid = *ashFid;
1759 /* Use normal vcache's fid otherwise. */
1760 tdc->f.fid = avc->fid;
1761 if (avc->states & CRO)
1762 tdc->f.states = DRO;
1763 else if (avc->states & CBackup)
1764 tdc->f.states = DBackup;
1766 tdc->f.states = DRW;
1767 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1768 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1770 hones(tdc->f.versionNo); /* invalid value */
1771 tdc->f.chunk = chunk;
1772 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1774 if (tdc->lruq.prev == &tdc->lruq)
1775 osi_Panic("lruq 1");
1784 * This function is called to obtain a reference to data stored in
1785 * the disk cache, locating a chunk of data containing the desired
1786 * byte and returning a reference to the disk cache entry, with its
1787 * reference count incremented.
1791 * avc : Ptr to a vcache entry (unlocked)
1792 * abyte : Byte position in the file desired
1793 * areq : Request structure identifying the requesting user.
1794 * aflags : Settings as follows:
1796 * 2 : Return after creating entry.
1797 * 4 : called from afs_vnop_write.c
1798 * *alen contains length of data to be written.
1800 * aoffset : Set to the offset within the chunk where the resident
1802 * alen : Set to the number of bytes of data after the desired
1803 * byte (including the byte itself) which can be read
1807 * The vcache entry pointed to by avc is unlocked upon entry.
1811 struct AFSVolSync tsync;
1812 struct AFSFetchStatus OutStatus;
1813 struct AFSCallBack CallBack;
1817 * Update the vnode-to-dcache hint if we can get the vnode lock
1818 * right away. Assumes dcache entry is at least read-locked.
1821 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1823 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1824 if (hsame(v->m.DataVersion, d->f.versionNo) && v->callback)
1827 ReleaseWriteLock(&v->lock);
1831 /* avc - Write-locked unless aflags & 1 */
1833 afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
1834 register struct vrequest *areq, afs_size_t * aoffset,
1835 afs_size_t * alen, int aflags)
1837 register afs_int32 i, code, code1 = 0, shortcut;
1838 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1839 register afs_int32 adjustsize = 0;
1845 afs_size_t maxGoodLength; /* amount of good data at server */
1846 struct rx_call *tcall;
1847 afs_size_t Position = 0;
1848 #ifdef AFS_64BIT_CLIENT
1850 afs_size_t lengthFound; /* as returned from server */
1851 #endif /* AFS_64BIT_CLIENT */
1852 afs_int32 size, tlen; /* size of segment to transfer */
1853 struct tlocal1 *tsmall = 0;
1854 register struct dcache *tdc;
1855 register struct osi_file *file;
1856 register struct conn *tc;
1858 struct server *newCallback = NULL;
1859 char setNewCallback;
1860 char setVcacheStatus;
1861 char doVcacheUpdate;
1863 int doAdjustSize = 0;
1864 int doReallyAdjustSize = 0;
1865 int overWriteWholeChunk = 0;
1869 struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
1870 osi_timeval_t xferStartTime, /*FS xfer start time */
1871 xferStopTime; /*FS xfer stop time */
1872 afs_size_t bytesToXfer; /* # bytes to xfer */
1873 afs_size_t bytesXferred; /* # bytes actually xferred */
1874 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1875 int fromReplica; /*Are we reading from a replica? */
1876 int numFetchLoops; /*# times around the fetch/analyze loop */
1877 #endif /* AFS_NOSTATS */
1879 AFS_STATCNT(afs_GetDCache);
1883 setLocks = aflags & 1;
1886 * Determine the chunk number and offset within the chunk corresponding
1887 * to the desired byte.
1889 if (avc->fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1892 chunk = AFS_CHUNK(abyte);
1895 /* come back to here if we waited for the cache to drain. */
1898 setNewCallback = setVcacheStatus = 0;
1902 ObtainWriteLock(&avc->lock, 616);
1904 ObtainReadLock(&avc->lock);
1909 * avc->lock(R) if setLocks && !slowPass
1910 * avc->lock(W) if !setLocks || slowPass
1915 /* check hints first! (might could use bcmp or some such...) */
1916 if ((tdc = avc->dchint)) {
1920 * The locking order between afs_xdcache and dcache lock matters.
1921 * The hint dcache entry could be anywhere, even on the free list.
1922 * Locking afs_xdcache ensures that noone is trying to pull dcache
1923 * entries from the free list, and thereby assuming them to be not
1924 * referenced and not locked.
1926 MObtainReadLock(&afs_xdcache);
1927 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1929 if (dcLocked && (tdc->index != NULLIDX)
1930 && !FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk
1931 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1932 /* got the right one. It might not be the right version, and it
1933 * might be fetching, but it's the right dcache entry.
1935 /* All this code should be integrated better with what follows:
1936 * I can save a good bit more time under a write lock if I do..
1938 ObtainWriteLock(&tdc->tlock, 603);
1940 ReleaseWriteLock(&tdc->tlock);
1942 MReleaseReadLock(&afs_xdcache);
1945 if (hsame(tdc->f.versionNo, avc->m.DataVersion)
1946 && !(tdc->dflags & DFFetching)) {
1948 afs_stats_cmperf.dcacheHits++;
1949 MObtainWriteLock(&afs_xdcache, 559);
1950 QRemove(&tdc->lruq);
1951 QAdd(&afs_DLRU, &tdc->lruq);
1952 MReleaseWriteLock(&afs_xdcache);
1955 * avc->lock(R) if setLocks && !slowPass
1956 * avc->lock(W) if !setLocks || slowPass
1963 ReleaseSharedLock(&tdc->lock);
1964 MReleaseReadLock(&afs_xdcache);
1972 * avc->lock(R) if setLocks && !slowPass
1973 * avc->lock(W) if !setLocks || slowPass
1974 * tdc->lock(S) if tdc
1977 if (!tdc) { /* If the hint wasn't the right dcache entry */
1979 * Hash on the [fid, chunk] and get the corresponding dcache index
1980 * after write-locking the dcache.
1985 * avc->lock(R) if setLocks && !slowPass
1986 * avc->lock(W) if !setLocks || slowPass
1989 i = DCHash(&avc->fid, chunk);
1990 /* check to make sure our space is fine */
1991 afs_MaybeWakeupTruncateDaemon();
1993 MObtainWriteLock(&afs_xdcache, 280);
1995 for (index = afs_dchashTbl[i]; index != NULLIDX;) {
1996 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
1997 tdc = afs_GetDSlot(index, NULL);
1998 ReleaseReadLock(&tdc->tlock);
2001 * avc->lock(R) if setLocks && !slowPass
2002 * avc->lock(W) if !setLocks || slowPass
2005 if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
2006 /* Move it up in the beginning of the list */
2007 if (afs_dchashTbl[i] != index) {
2008 afs_dcnextTbl[us] = afs_dcnextTbl[index];
2009 afs_dcnextTbl[index] = afs_dchashTbl[i];
2010 afs_dchashTbl[i] = index;
2012 MReleaseWriteLock(&afs_xdcache);
2013 ObtainSharedLock(&tdc->lock, 606);
2014 break; /* leaving refCount high for caller */
2020 index = afs_dcnextTbl[index];
2024 * If we didn't find the entry, we'll create one.
2026 if (index == NULLIDX) {
2029 * avc->lock(R) if setLocks
2030 * avc->lock(W) if !setLocks
2033 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
2034 avc, ICL_TYPE_INT32, chunk);
2036 /* Make sure there is a free dcache entry for us to use */
2037 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
2040 avc->states |= CDCLock;
2041 /* just need slots */
2042 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
2044 avc->states &= ~CDCLock;
2045 if (afs_discardDCList != NULLIDX
2046 || afs_freeDCList != NULLIDX)
2048 /* If we can't get space for 5 mins we give up and panic */
2049 if (++downDCount > 300) {
2050 #if defined(AFS_CACHE_BYPASS)
2051 afs_warn("GetDCache calling osi_Panic: No space in five minutes.\n downDCount: %d\n aoffset: %d alen: %d\n", downDCount, aoffset, alen);
2053 osi_Panic("getdcache");
2055 MReleaseWriteLock(&afs_xdcache);
2058 * avc->lock(R) if setLocks
2059 * avc->lock(W) if !setLocks
2061 afs_osi_Wait(1000, 0, 0);
2066 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
2069 * Now add to the two hash chains - note that i is still set
2070 * from the above DCHash call.
2072 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
2073 afs_dchashTbl[i] = tdc->index;
2074 i = DVHash(&avc->fid);
2075 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
2076 afs_dvhashTbl[i] = tdc->index;
2077 tdc->dflags = DFEntryMod;
2079 afs_MaybeWakeupTruncateDaemon();
2080 MReleaseWriteLock(&afs_xdcache);
2081 ConvertWToSLock(&tdc->lock);
2086 /* vcache->dcache hint failed */
2089 * avc->lock(R) if setLocks && !slowPass
2090 * avc->lock(W) if !setLocks || slowPass
2093 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
2094 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2095 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
2096 hgetlo(avc->m.DataVersion));
2098 * Here we have the entry in tdc, with its refCount incremented.
2099 * Note: we don't use the S-lock on avc; it costs concurrency when
2100 * storing a file back to the server.
2104 * Not a newly created file so we need to check the file's length and
2105 * compare data versions since someone could have changed the data or we're
2106 * reading a file written elsewhere. We only want to bypass doing no-op
2107 * read rpcs on newly created files (dv of 0) since only then we guarantee
2108 * that this chunk's data hasn't been filled by another client.
2110 size = AFS_CHUNKSIZE(abyte);
2111 if (aflags & 4) /* called from write */
2113 else /* called from read */
2114 tlen = tdc->validPos - abyte;
2115 Position = AFS_CHUNKTOBASE(chunk);
2116 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
2117 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
2118 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
2119 ICL_HANDLE_OFFSET(Position));
2120 if ((aflags & 4) && (hiszero(avc->m.DataVersion)))
2122 if ((AFS_CHUNKTOBASE(chunk) >= avc->m.Length) ||
2123 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
2124 overWriteWholeChunk = 1;
2125 if (doAdjustSize || overWriteWholeChunk) {
2126 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
2128 #ifdef AFS_SGI64_ENV
2131 #else /* AFS_SGI64_ENV */
2134 #endif /* AFS_SGI64_ENV */
2135 #else /* AFS_SGI_ENV */
2138 #endif /* AFS_SGI_ENV */
2139 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->m.Length &&
2140 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2141 #if defined(AFS_SUN5_ENV) || defined(AFS_OSF_ENV)
2142 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
2144 if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
2146 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2147 !hsame(avc->m.DataVersion, tdc->f.versionNo))
2148 doReallyAdjustSize = 1;
2150 if (doReallyAdjustSize || overWriteWholeChunk) {
2151 /* no data in file to read at this position */
2152 UpgradeSToWLock(&tdc->lock, 607);
2154 #if defined(LINUX_USE_FH)
2155 file = afs_CFileOpen(&tdc->f.fh, tdc->f.fh_type);
2157 file = afs_CFileOpen(tdc->f.inode);
2159 afs_CFileTruncate(file, 0);
2160 afs_CFileClose(file);
2161 afs_AdjustSize(tdc, 0);
2162 hset(tdc->f.versionNo, avc->m.DataVersion);
2163 tdc->dflags |= DFEntryMod;
2165 ConvertWToSLock(&tdc->lock);
2170 * We must read in the whole chunk if the version number doesn't
2174 /* don't need data, just a unique dcache entry */
2175 ObtainWriteLock(&afs_xdcache, 608);
2176 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2177 hadd32(afs_indexCounter, 1);
2178 ReleaseWriteLock(&afs_xdcache);
2180 updateV2DC(setLocks, avc, tdc, 553);
2181 if (vType(avc) == VDIR)
2184 *aoffset = AFS_CHUNKOFFSET(abyte);
2185 if (tdc->validPos < abyte)
2186 *alen = (afs_size_t) 0;
2188 *alen = tdc->validPos - abyte;
2189 ReleaseSharedLock(&tdc->lock);
2192 ReleaseWriteLock(&avc->lock);
2194 ReleaseReadLock(&avc->lock);
2196 return tdc; /* check if we're done */
2201 * avc->lock(R) if setLocks && !slowPass
2202 * avc->lock(W) if !setLocks || slowPass
2205 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2207 setNewCallback = setVcacheStatus = 0;
2211 * avc->lock(R) if setLocks && !slowPass
2212 * avc->lock(W) if !setLocks || slowPass
2215 if (!hsame(avc->m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2217 * Version number mismatch.
2220 * If we are disconnected, then we can't do much of anything
2221 * because the data doesn't match the file.
2223 if (AFS_IS_DISCONNECTED) {
2224 ReleaseSharedLock(&tdc->lock);
2227 ReleaseWriteLock(&avc->lock);
2229 ReleaseReadLock(&avc->lock);
2231 /* Flush the Dcache */
2236 UpgradeSToWLock(&tdc->lock, 609);
2239 * If data ever existed for this vnode, and this is a text object,
2240 * do some clearing. Now, you'd think you need only do the flush
2241 * when VTEXT is on, but VTEXT is turned off when the text object
2242 * is freed, while pages are left lying around in memory marked
2243 * with this vnode. If we would reactivate (create a new text
2244 * object from) this vnode, we could easily stumble upon some of
2245 * these old pages in pagein. So, we always flush these guys.
2246 * Sun has a wonderful lack of useful invariants in this system.
2248 * avc->flushDV is the data version # of the file at the last text
2249 * flush. Clearly, at least, we don't have to flush the file more
2250 * often than it changes
2252 if (hcmp(avc->flushDV, avc->m.DataVersion) < 0) {
2254 * By here, the cache entry is always write-locked. We can
2255 * deadlock if we call osi_Flush with the cache entry locked...
2256 * Unlock the dcache too.
2258 ReleaseWriteLock(&tdc->lock);
2259 if (setLocks && !slowPass)
2260 ReleaseReadLock(&avc->lock);
2262 ReleaseWriteLock(&avc->lock);
2266 * Call osi_FlushPages in open, read/write, and map, since it
2267 * is too hard here to figure out if we should lock the
2270 if (setLocks && !slowPass)
2271 ObtainReadLock(&avc->lock);
2273 ObtainWriteLock(&avc->lock, 66);
2274 ObtainWriteLock(&tdc->lock, 610);
2279 * avc->lock(R) if setLocks && !slowPass
2280 * avc->lock(W) if !setLocks || slowPass
2284 /* Watch for standard race condition around osi_FlushText */
2285 if (hsame(avc->m.DataVersion, tdc->f.versionNo)) {
2286 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2287 afs_stats_cmperf.dcacheHits++;
2288 ConvertWToSLock(&tdc->lock);
2292 /* Sleep here when cache needs to be drained. */
2293 if (setLocks && !slowPass
2294 && (afs_blocksUsed >
2295 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2296 /* Make sure truncate daemon is running */
2297 afs_MaybeWakeupTruncateDaemon();
2298 ObtainWriteLock(&tdc->tlock, 614);
2299 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2300 ReleaseWriteLock(&tdc->tlock);
2301 ReleaseWriteLock(&tdc->lock);
2302 ReleaseReadLock(&avc->lock);
2303 while ((afs_blocksUsed - afs_blocksDiscarded) >
2304 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2305 afs_WaitForCacheDrain = 1;
2306 afs_osi_Sleep(&afs_WaitForCacheDrain);
2308 afs_MaybeFreeDiscardedDCache();
2309 /* need to check if someone else got the chunk first. */
2310 goto RetryGetDCache;
2313 /* Do not fetch data beyond truncPos. */
2314 maxGoodLength = avc->m.Length;
2315 if (avc->truncPos < maxGoodLength)
2316 maxGoodLength = avc->truncPos;
2317 Position = AFS_CHUNKBASE(abyte);
2318 if (vType(avc) == VDIR) {
2319 size = avc->m.Length;
2320 if (size > tdc->f.chunkBytes) {
2321 /* pre-reserve space for file */
2322 afs_AdjustSize(tdc, size);
2324 size = 999999999; /* max size for transfer */
2326 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2327 /* don't read past end of good data on server */
2328 if (Position + size > maxGoodLength)
2329 size = maxGoodLength - Position;
2331 size = 0; /* Handle random races */
2332 if (size > tdc->f.chunkBytes) {
2333 /* pre-reserve space for file */
2334 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2335 /* max size for transfer still in size */
2338 if (afs_mariner && !tdc->f.chunk)
2339 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2341 * Right now, we only have one tool, and it's a hammer. So, we
2342 * fetch the whole file.
2344 DZap(tdc); /* pages in cache may be old */
2345 #if defined(LINUX_USE_FH)
2346 file = afs_CFileOpen(&tdc->f.fh, tdc->f.fh_type);
2348 file = afs_CFileOpen(tdc->f.inode);
2350 afs_RemoveVCB(&avc->fid);
2351 tdc->f.states |= DWriting;
2352 tdc->dflags |= DFFetching;
2353 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2354 if (tdc->mflags & DFFetchReq) {
2355 tdc->mflags &= ~DFFetchReq;
2356 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2357 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2358 __FILE__, ICL_TYPE_INT32, __LINE__,
2359 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2363 (struct tlocal1 *)osi_AllocLargeSpace(sizeof(struct tlocal1));
2364 setVcacheStatus = 0;
2367 * Remember if we are doing the reading from a replicated volume,
2368 * and how many times we've zipped around the fetch/analyze loop.
2370 fromReplica = (avc->states & CRO) ? 1 : 0;
2372 accP = &(afs_stats_cmfullperf.accessinf);
2374 (accP->replicatedRefs)++;
2376 (accP->unreplicatedRefs)++;
2377 #endif /* AFS_NOSTATS */
2378 /* this is a cache miss */
2379 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2380 ICL_TYPE_FID, &(avc->fid), ICL_TYPE_OFFSET,
2381 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2384 afs_stats_cmperf.dcacheMisses++;
2387 * Dynamic root support: fetch data from local memory.
2389 if (afs_IsDynroot(avc)) {
2393 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2395 dynrootDir += Position;
2396 dynrootLen -= Position;
2397 if (size > dynrootLen)
2401 code = afs_CFileWrite(file, 0, dynrootDir, size);
2409 tdc->validPos = Position + size;
2410 afs_CFileTruncate(file, size); /* prune it */
2411 } else if (afs_IsDynrootMount(avc)) {
2415 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2417 dynrootDir += Position;
2418 dynrootLen -= Position;
2419 if (size > dynrootLen)
2423 code = afs_CFileWrite(file, 0, dynrootDir, size);
2431 tdc->validPos = Position + size;
2432 afs_CFileTruncate(file, size); /* prune it */
2433 } else if (afs_IsDynrootMount(avc)) {
2437 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2439 dynrootDir += Position;
2440 dynrootLen -= Position;
2441 if (size > dynrootLen)
2445 code = afs_CFileWrite(file, 0, dynrootDir, size);
2453 tdc->validPos = Position + size;
2454 afs_CFileTruncate(file, size); /* prune it */
2457 * Not a dynamic vnode: do the real fetch.
2462 * avc->lock(R) if setLocks && !slowPass
2463 * avc->lock(W) if !setLocks || slowPass
2467 tc = afs_Conn(&avc->fid, areq, SHARED_LOCK);
2469 afs_int32 length_hi, length, bytes;
2473 (accP->numReplicasAccessed)++;
2475 #endif /* AFS_NOSTATS */
2476 if (!setLocks || slowPass) {
2477 avc->callback = tc->srvr->server;
2479 newCallback = tc->srvr->server;
2484 tcall = rx_NewCall(tc->id);
2487 XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
2488 #ifdef AFS_64BIT_CLIENT
2489 length_hi = code = 0;
2490 if (!afs_serverHasNo64Bit(tc)) {
2494 StartRXAFS_FetchData64(tcall,
2495 (struct AFSFid *)&avc->fid.
2496 Fid, Position, tsize);
2499 afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
2500 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32,
2504 rx_Read(tcall, (char *)&length_hi,
2507 if (bytes == sizeof(afs_int32)) {
2508 length_hi = ntohl(length_hi);
2511 code = rx_Error(tcall);
2513 code1 = rx_EndCall(tcall, code);
2515 tcall = (struct rx_call *)0;
2519 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
2520 if (Position > 0x7FFFFFFF) {
2527 tcall = rx_NewCall(tc->id);
2529 StartRXAFS_FetchData(tcall, (struct AFSFid *)
2534 afs_serverSetNo64Bit(tc);
2539 rx_Read(tcall, (char *)&length,
2542 if (bytes == sizeof(afs_int32)) {
2543 length = ntohl(length);
2545 code = rx_Error(tcall);
2548 FillInt64(lengthFound, length_hi, length);
2549 afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64LENG,
2550 ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
2552 ICL_HANDLE_OFFSET(lengthFound));
2553 #else /* AFS_64BIT_CLIENT */
2556 StartRXAFS_FetchData(tcall,
2557 (struct AFSFid *)&avc->fid.Fid,
2563 rx_Read(tcall, (char *)&length,
2566 if (bytes == sizeof(afs_int32)) {
2567 length = ntohl(length);
2569 code = rx_Error(tcall);
2572 #endif /* AFS_64BIT_CLIENT */
2577 &(afs_stats_cmfullperf.rpc.
2578 fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
2579 osi_GetuTime(&xferStartTime);
2582 afs_CacheFetchProc(tcall, file,
2583 (afs_size_t) Position, tdc,
2585 &bytesXferred, length);
2587 osi_GetuTime(&xferStopTime);
2588 (xferP->numXfers)++;
2590 (xferP->numSuccesses)++;
2591 afs_stats_XferSumBytes
2592 [AFS_STATS_FS_XFERIDX_FETCHDATA] +=
2594 (xferP->sumBytes) +=
2595 (afs_stats_XferSumBytes
2596 [AFS_STATS_FS_XFERIDX_FETCHDATA] >> 10);
2597 afs_stats_XferSumBytes
2598 [AFS_STATS_FS_XFERIDX_FETCHDATA] &= 0x3FF;
2599 if (bytesXferred < xferP->minBytes)
2600 xferP->minBytes = bytesXferred;
2601 if (bytesXferred > xferP->maxBytes)
2602 xferP->maxBytes = bytesXferred;
2605 * Tally the size of the object. Note: we tally the actual size,
2606 * NOT the number of bytes that made it out over the wire.
2608 if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
2609 (xferP->count[0])++;
2610 else if (bytesToXfer <=
2611 AFS_STATS_MAXBYTES_BUCKET1)
2612 (xferP->count[1])++;
2613 else if (bytesToXfer <=
2614 AFS_STATS_MAXBYTES_BUCKET2)
2615 (xferP->count[2])++;
2616 else if (bytesToXfer <=
2617 AFS_STATS_MAXBYTES_BUCKET3)
2618 (xferP->count[3])++;
2619 else if (bytesToXfer <=
2620 AFS_STATS_MAXBYTES_BUCKET4)
2621 (xferP->count[4])++;
2622 else if (bytesToXfer <=
2623 AFS_STATS_MAXBYTES_BUCKET5)
2624 (xferP->count[5])++;
2625 else if (bytesToXfer <=
2626 AFS_STATS_MAXBYTES_BUCKET6)
2627 (xferP->count[6])++;
2628 else if (bytesToXfer <=
2629 AFS_STATS_MAXBYTES_BUCKET7)
2630 (xferP->count[7])++;
2632 (xferP->count[8])++;
2634 afs_stats_GetDiff(elapsedTime, xferStartTime,
2636 afs_stats_AddTo((xferP->sumTime), elapsedTime);
2637 afs_stats_SquareAddTo((xferP->sqrTime),
2639 if (afs_stats_TimeLessThan
2640 (elapsedTime, (xferP->minTime))) {
2641 afs_stats_TimeAssign((xferP->minTime),
2644 if (afs_stats_TimeGreaterThan
2645 (elapsedTime, (xferP->maxTime))) {
2646 afs_stats_TimeAssign((xferP->maxTime),
2652 afs_CacheFetchProc(tcall, file, Position, tdc,
2654 #endif /* AFS_NOSTATS */
2659 EndRXAFS_FetchData(tcall, &tsmall->OutStatus,
2667 code1 = rx_EndCall(tcall, code);
2676 /* callback could have been broken (or expired) in a race here,
2677 * but we return the data anyway. It's as good as we knew about
2678 * when we started. */
2680 * validPos is updated by CacheFetchProc, and can only be
2681 * modifed under a dcache write lock, which we've blocked out
2683 size = tdc->validPos - Position; /* actual segment size */
2686 afs_CFileTruncate(file, size); /* prune it */
2688 if (!setLocks || slowPass) {
2689 ObtainWriteLock(&afs_xcbhash, 453);
2690 afs_DequeueCallback(avc);
2691 avc->states &= ~(CStatd | CUnique);
2692 avc->callback = NULL;
2693 ReleaseWriteLock(&afs_xcbhash);
2694 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2695 osi_dnlc_purgedp(avc);
2697 /* Something lost. Forget about performance, and go
2698 * back with a vcache write lock.
2700 afs_CFileTruncate(file, 0);
2701 afs_AdjustSize(tdc, 0);
2702 afs_CFileClose(file);
2703 osi_FreeLargeSpace(tsmall);
2705 ReleaseWriteLock(&tdc->lock);
2708 ReleaseReadLock(&avc->lock);
2710 goto RetryGetDCache;
2714 } while (afs_Analyze
2715 (tc, code, &avc->fid, areq,
2716 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2720 * avc->lock(R) if setLocks && !slowPass
2721 * avc->lock(W) if !setLocks || slowPass
2727 * In the case of replicated access, jot down info on the number of
2728 * attempts it took before we got through or gave up.
2731 if (numFetchLoops <= 1)
2732 (accP->refFirstReplicaOK)++;
2733 if (numFetchLoops > accP->maxReplicasPerRef)
2734 accP->maxReplicasPerRef = numFetchLoops;
2736 #endif /* AFS_NOSTATS */
2738 tdc->dflags &= ~DFFetching;
2739 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2740 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2741 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2742 tdc, ICL_TYPE_INT32, tdc->dflags);
2743 if (avc->execsOrWriters == 0)
2744 tdc->f.states &= ~DWriting;
2746 /* now, if code != 0, we have an error and should punt.
2747 * note that we have the vcache write lock, either because
2748 * !setLocks or slowPass.
2751 afs_CFileTruncate(file, 0);
2752 afs_AdjustSize(tdc, 0);
2753 afs_CFileClose(file);
2754 ZapDCE(tdc); /* sets DFEntryMod */
2755 if (vType(avc) == VDIR) {
2758 tdc->f.states &= ~(DRO|DBackup|DRW);
2759 afs_DCMoveBucket(tdc, 0, 0);
2760 ReleaseWriteLock(&tdc->lock);
2762 if (!afs_IsDynroot(avc)) {
2763 ObtainWriteLock(&afs_xcbhash, 454);
2764 afs_DequeueCallback(avc);
2765 avc->states &= ~(CStatd | CUnique);
2766 ReleaseWriteLock(&afs_xcbhash);
2767 if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2768 osi_dnlc_purgedp(avc);
2771 * avc->lock(W); assert(!setLocks || slowPass)
2773 osi_Assert(!setLocks || slowPass);
2775 tdc->f.states &= ~(DRO|DBackup|DRW);
2776 afs_DCMoveBucket(tdc, 0, 0);
2781 /* otherwise we copy in the just-fetched info */
2782 afs_CFileClose(file);
2783 afs_AdjustSize(tdc, size); /* new size */
2785 * Copy appropriate fields into vcache. Status is
2786 * copied later where we selectively acquire the
2787 * vcache write lock.
2790 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2792 setVcacheStatus = 1;
2793 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2794 tsmall->OutStatus.DataVersion);
2795 tdc->dflags |= DFEntryMod;
2796 afs_indexFlags[tdc->index] |= IFEverUsed;
2797 ConvertWToSLock(&tdc->lock);
2798 } /*Data version numbers don't match */
2801 * Data version numbers match.
2803 afs_stats_cmperf.dcacheHits++;
2804 } /*Data version numbers match */
2806 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2810 * avc->lock(R) if setLocks && !slowPass
2811 * avc->lock(W) if !setLocks || slowPass
2812 * tdc->lock(S) if tdc
2816 * See if this was a reference to a file in the local cell.
2818 if (afs_IsPrimaryCellNum(avc->fid.Cell))
2819 afs_stats_cmperf.dlocalAccesses++;
2821 afs_stats_cmperf.dremoteAccesses++;
2823 /* Fix up LRU info */
2826 MObtainWriteLock(&afs_xdcache, 602);
2827 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2828 hadd32(afs_indexCounter, 1);
2829 MReleaseWriteLock(&afs_xdcache);
2831 /* return the data */
2832 if (vType(avc) == VDIR)
2835 *aoffset = AFS_CHUNKOFFSET(abyte);
2836 *alen = (tdc->f.chunkBytes - *aoffset);
2837 ReleaseSharedLock(&tdc->lock);
2842 * avc->lock(R) if setLocks && !slowPass
2843 * avc->lock(W) if !setLocks || slowPass
2846 /* Fix up the callback and status values in the vcache */
2848 if (setLocks && !slowPass) {
2851 * This is our dirty little secret to parallel fetches.
2852 * We don't write-lock the vcache while doing the fetch,
2853 * but potentially we'll need to update the vcache after
2854 * the fetch is done.
2856 * Drop the read lock and try to re-obtain the write
2857 * lock. If the vcache still has the same DV, it's
2858 * ok to go ahead and install the new data.
2860 afs_hyper_t currentDV, statusDV;
2862 hset(currentDV, avc->m.DataVersion);
2864 if (setNewCallback && avc->callback != newCallback)
2868 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2869 tsmall->OutStatus.DataVersion);
2871 if (setVcacheStatus && avc->m.Length != tsmall->OutStatus.Length)
2873 if (setVcacheStatus && !hsame(currentDV, statusDV))
2877 ReleaseReadLock(&avc->lock);
2879 if (doVcacheUpdate) {
2880 ObtainWriteLock(&avc->lock, 615);
2881 if (!hsame(avc->m.DataVersion, currentDV)) {
2882 /* We lose. Someone will beat us to it. */
2884 ReleaseWriteLock(&avc->lock);
2889 /* With slow pass, we've already done all the updates */
2891 ReleaseWriteLock(&avc->lock);
2894 /* Check if we need to perform any last-minute fixes with a write-lock */
2895 if (!setLocks || doVcacheUpdate) {
2897 avc->callback = newCallback;
2898 if (tsmall && setVcacheStatus)
2899 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2901 ReleaseWriteLock(&avc->lock);
2905 osi_FreeLargeSpace(tsmall);
2908 } /*afs_GetDCache */
2912 * afs_WriteThroughDSlots
2915 * Sweep through the dcache slots and write out any modified
2916 * in-memory data back on to our caching store.
2922 * The afs_xdcache is write-locked through this whole affair.
2925 afs_WriteThroughDSlots(void)
2927 register struct dcache *tdc;
2928 register afs_int32 i, touchedit = 0;
2930 struct afs_q DirtyQ, *tq;
2932 AFS_STATCNT(afs_WriteThroughDSlots);
2935 * Because of lock ordering, we can't grab dcache locks while
2936 * holding afs_xdcache. So we enter xdcache, get a reference
2937 * for every dcache entry, and exit xdcache.
2939 MObtainWriteLock(&afs_xdcache, 283);
2941 for (i = 0; i < afs_cacheFiles; i++) {
2942 tdc = afs_indexTable[i];
2944 /* Grab tlock in case the existing refcount isn't zero */
2945 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2946 ObtainWriteLock(&tdc->tlock, 623);
2948 ReleaseWriteLock(&tdc->tlock);
2950 QAdd(&DirtyQ, &tdc->dirty);
2953 MReleaseWriteLock(&afs_xdcache);
2956 * Now, for each dcache entry we found, check if it's dirty.
2957 * If so, get write-lock, get afs_xdcache, which protects
2958 * afs_cacheInodep, and flush it. Don't forget to put back
2962 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2964 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2966 if (tdc->dflags & DFEntryMod) {
2969 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2971 /* Now that we have the write lock, double-check */
2972 if (wrLock && (tdc->dflags & DFEntryMod)) {
2973 tdc->dflags &= ~DFEntryMod;
2974 MObtainWriteLock(&afs_xdcache, 620);
2975 afs_WriteDCache(tdc, 1);
2976 MReleaseWriteLock(&afs_xdcache);
2980 ReleaseWriteLock(&tdc->lock);
2986 MObtainWriteLock(&afs_xdcache, 617);
2987 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2988 /* Touch the file to make sure that the mtime on the file is kept
2989 * up-to-date to avoid losing cached files on cold starts because
2990 * their mtime seems old...
2992 struct afs_fheader theader;
2994 theader.magic = AFS_FHMAGIC;
2995 theader.firstCSize = AFS_FIRSTCSIZE;
2996 theader.otherCSize = AFS_OTHERCSIZE;
2997 theader.version = AFS_CI_VERSION;
2998 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
3000 MReleaseWriteLock(&afs_xdcache);
3007 * Return a pointer to an freshly initialized dcache entry using
3008 * a memory-based cache. The tlock will be read-locked.
3011 * aslot : Dcache slot to look at.
3012 * tmpdc : Ptr to dcache entry.
3015 * Must be called with afs_xdcache write-locked.
3019 afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
3021 register struct dcache *tdc;
3024 AFS_STATCNT(afs_MemGetDSlot);
3025 if (CheckLock(&afs_xdcache) != -1)
3026 osi_Panic("getdslot nolock");
3027 if (aslot < 0 || aslot >= afs_cacheFiles)
3028 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
3029 tdc = afs_indexTable[aslot];
3031 QRemove(&tdc->lruq); /* move to queue head */
3032 QAdd(&afs_DLRU, &tdc->lruq);
3033 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
3034 ObtainWriteLock(&tdc->tlock, 624);
3036 ConvertWToRLock(&tdc->tlock);
3039 if (tmpdc == NULL) {
3040 if (!afs_freeDSList)
3041 afs_GetDownDSlot(4);
3042 if (!afs_freeDSList) {
3043 /* none free, making one is better than a panic */
3044 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
3045 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
3046 #ifdef KERNEL_HAVE_PIN
3047 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
3050 tdc = afs_freeDSList;
3051 afs_freeDSList = (struct dcache *)tdc->lruq.next;
3054 tdc->dflags = 0; /* up-to-date, not in free q */
3056 QAdd(&afs_DLRU, &tdc->lruq);
3057 if (tdc->lruq.prev == &tdc->lruq)
3058 osi_Panic("lruq 3");
3064 /* initialize entry */
3065 tdc->f.fid.Cell = 0;
3066 tdc->f.fid.Fid.Volume = 0;
3068 hones(tdc->f.versionNo);
3069 tdc->f.inode = aslot;
3070 tdc->dflags |= DFEntryMod;
3073 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
3076 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
3077 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
3078 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
3081 RWLOCK_INIT(&tdc->lock, "dcache lock");
3082 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
3083 RWLOCK_INIT(&tdc->mflock, "dcache flock");
3084 ObtainReadLock(&tdc->tlock);
3087 afs_indexTable[aslot] = tdc;
3090 } /*afs_MemGetDSlot */
3092 unsigned int last_error = 0, lasterrtime = 0;
3098 * Return a pointer to an freshly initialized dcache entry using
3099 * a UFS-based disk cache. The dcache tlock will be read-locked.
3102 * aslot : Dcache slot to look at.
3103 * tmpdc : Ptr to dcache entry.
3106 * afs_xdcache lock write-locked.
3109 afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
3111 register afs_int32 code;
3112 register struct dcache *tdc;
3116 AFS_STATCNT(afs_UFSGetDSlot);
3117 if (CheckLock(&afs_xdcache) != -1)
3118 osi_Panic("getdslot nolock");
3119 if (aslot < 0 || aslot >= afs_cacheFiles)
3120 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
3121 tdc = afs_indexTable[aslot];
3123 QRemove(&tdc->lruq); /* move to queue head */
3124 QAdd(&afs_DLRU, &tdc->lruq);
3125 /* Grab tlock in case refCount != 0 */
3126 ObtainWriteLock(&tdc->tlock, 625);
3128 ConvertWToRLock(&tdc->tlock);
3131 /* otherwise we should read it in from the cache file */
3133 * If we weren't passed an in-memory region to place the file info,
3134 * we have to allocate one.
3136 if (tmpdc == NULL) {
3137 if (!afs_freeDSList)
3138 afs_GetDownDSlot(4);
3139 if (!afs_freeDSList) {
3140 /* none free, making one is better than a panic */
3141 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
3142 tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
3143 #ifdef KERNEL_HAVE_PIN
3144 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
3147 tdc = afs_freeDSList;
3148 afs_freeDSList = (struct dcache *)tdc->lruq.next;
3151 tdc->dflags = 0; /* up-to-date, not in free q */
3153 QAdd(&afs_DLRU, &tdc->lruq);
3154 if (tdc->lruq.prev == &tdc->lruq)
3155 osi_Panic("lruq 3");
3162 * Seek to the aslot'th entry and read it in.
3165 afs_osi_Read(afs_cacheInodep,
3166 sizeof(struct fcache) * aslot +
3167 sizeof(struct afs_fheader), (char *)(&tdc->f),
3168 sizeof(struct fcache));
3170 if (code != sizeof(struct fcache))
3172 if (!afs_CellNumValid(tdc->f.fid.Cell))
3176 tdc->f.fid.Cell = 0;
3177 tdc->f.fid.Fid.Volume = 0;
3179 hones(tdc->f.versionNo);
3180 tdc->dflags |= DFEntryMod;
3181 #if defined(KERNEL_HAVE_UERROR)
3182 last_error = getuerror();
3184 lasterrtime = osi_Time();
3185 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
3186 tdc->f.states &= ~(DRO|DBackup|DRW);
3187 afs_DCMoveBucket(tdc, 0, 0);
3190 if (tdc->f.states & DRO) {
3191 afs_DCMoveBucket(tdc, 0, 2);
3192 } else if (tdc->f.states & DBackup) {
3193 afs_DCMoveBucket(tdc, 0, 1);
3195 afs_DCMoveBucket(tdc, 0, 1);
3201 if (tdc->f.chunk >= 0)
3202 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
3207 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
3208 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
3209 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
3212 RWLOCK_INIT(&tdc->lock, "dcache lock");
3213 RWLOCK_INIT(&tdc->tlock, "dcache tlock");
3214 RWLOCK_INIT(&tdc->mflock, "dcache flock");
3215 ObtainReadLock(&tdc->tlock);
3218 * If we didn't read into a temporary dcache region, update the
3219 * slot pointer table.
3222 afs_indexTable[aslot] = tdc;
3225 } /*afs_UFSGetDSlot */
3230 * Write a particular dcache entry back to its home in the
3233 * \param adc Pointer to the dcache entry to write.
3234 * \param atime If true, set the modtime on the file to the current time.
3236 * \note Environment:
3237 * Must be called with the afs_xdcache lock at least read-locked,
3238 * and dcache entry at least read-locked.
3239 * The reference count is not changed.
3243 afs_WriteDCache(register struct dcache *adc, int atime)
3245 register afs_int32 code;
3247 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
3249 AFS_STATCNT(afs_WriteDCache);
3250 osi_Assert(WriteLocked(&afs_xdcache));
3252 adc->f.modTime = osi_Time();
3254 * Seek to the right dcache slot and write the in-memory image out to disk.
3256 afs_cellname_write();
3258 afs_osi_Write(afs_cacheInodep,
3259 sizeof(struct fcache) * adc->index +
3260 sizeof(struct afs_fheader), (char *)(&adc->f),
3261 sizeof(struct fcache));
3262 if (code != sizeof(struct fcache))
3270 * Wake up users of a particular file waiting for stores to take
3273 * \param avc Ptr to related vcache entry.
3275 * \note Environment:
3276 * Nothing interesting.
3279 afs_wakeup(register struct vcache *avc)
3282 register struct brequest *tb;
3284 AFS_STATCNT(afs_wakeup);
3285 for (i = 0; i < NBRS; i++, tb++) {
3286 /* if request is valid and for this file, we've found it */
3287 if (tb->refCount > 0 && avc == tb->vc) {
3290 * If CSafeStore is on, then we don't awaken the guy
3291 * waiting for the store until the whole store has finished.
3292 * Otherwise, we do it now. Note that if CSafeStore is on,
3293 * the BStore routine actually wakes up the user, instead
3295 * I think this is redundant now because this sort of thing
3296 * is already being handled by the higher-level code.
3298 if ((avc->states & CSafeStore) == 0) {
3300 tb->flags |= BUVALID;
3301 if (tb->flags & BUWAIT) {
3302 tb->flags &= ~BUWAIT;
3314 * Given a file name and inode, set up that file to be an
3315 * active member in the AFS cache. This also involves checking
3316 * the usability of its data.
3318 * \param afile Name of the cache file to initialize.
3319 * \param ainode Inode of the file.
3321 * \note Environment:
3322 * This function is called only during initialization.
3325 afs_InitCacheFile(char *afile, ino_t ainode)
3327 register afs_int32 code;
3328 #if defined(AFS_LINUX22_ENV)
3329 struct dentry *filevp;
3331 struct vnode *filevp;
3335 struct osi_file *tfile;
3336 struct osi_stat tstat;
3337 register struct dcache *tdc;
3338 #if defined(LINUX_USE_FH)
3339 int max_len = sizeof(struct fid);
3342 AFS_STATCNT(afs_InitCacheFile);
3343 index = afs_stats_cmperf.cacheNumEntries;
3344 if (index >= afs_cacheFiles)
3347 MObtainWriteLock(&afs_xdcache, 282);
3348 tdc = afs_GetDSlot(index, NULL);
3349 ReleaseReadLock(&tdc->tlock);
3350 MReleaseWriteLock(&afs_xdcache);
3352 ObtainWriteLock(&tdc->lock, 621);
3353 MObtainWriteLock(&afs_xdcache, 622);
3355 code = gop_lookupname(afile, AFS_UIOSYS, 0, &filevp);
3357 ReleaseWriteLock(&afs_xdcache);
3358 ReleaseWriteLock(&tdc->lock);
3363 * We have a VN_HOLD on filevp. Get the useful info out and
3364 * return. We make use of the fact that the cache is in the
3365 * UFS file system, and just record the inode number.
3367 #ifdef AFS_LINUX22_ENV
3368 #if defined(LINUX_USE_FH)
3369 tdc->f.fh_type = osi_get_fh(filevp, &tdc->f.fh, &max_len);
3371 tdc->f.inode = VTOI(filevp->d_inode)->i_number;
3374 tdc->f.inode = afs_vnodeToInumber(filevp);
3376 #endif /* AFS_LINUX22_ENV */
3378 tdc->f.inode = ainode;
3381 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3383 #if defined(LINUX_USE_FH)
3384 tfile = osi_UFSOpen_fh(&tdc->f.fh, tdc->f.fh_type);
3386 tfile = osi_UFSOpen(ainode);
3388 code = afs_osi_Stat(tfile, &tstat);
3390 osi_Panic("initcachefile stat");
3393 * If file size doesn't match the cache info file, it's probably bad.
3395 if (tdc->f.chunkBytes != tstat.size)
3397 tdc->f.chunkBytes = 0;
3400 * If file changed within T (120?) seconds of cache info file, it's
3401 * probably bad. In addition, if slot changed within last T seconds,
3402 * the cache info file may be incorrectly identified, and so slot
3405 if (cacheInfoModTime < tstat.mtime + 120)
3407 if (cacheInfoModTime < tdc->f.modTime + 120)
3409 /* In case write through is behind, make sure cache items entry is
3410 * at least as new as the chunk.
3412 if (tdc->f.modTime < tstat.mtime)
3415 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3416 if (tstat.size != 0)
3417 osi_UFSTruncate(tfile, 0);
3418 tdc->f.states &= ~(DRO|DBackup|DRW);
3419 afs_DCMoveBucket(tdc, 0, 0);
3420 /* put entry in free cache slot list */
3421 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3422 afs_freeDCList = index;
3424 afs_indexFlags[index] |= IFFree;
3425 afs_indexUnique[index] = 0;
3428 * We must put this entry in the appropriate hash tables.
3429 * Note that i is still set from the above DCHash call
3431 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3432 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3433 afs_dchashTbl[code] = tdc->index;
3434 code = DVHash(&tdc->f.fid);
3435 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3436 afs_dvhashTbl[code] = tdc->index;
3437 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3439 /* has nontrivial amt of data */
3440 afs_indexFlags[index] |= IFEverUsed;
3441 afs_stats_cmperf.cacheFilesReused++;
3443 * Initialize index times to file's mod times; init indexCounter
3446 hset32(afs_indexTimes[index], tstat.atime);
3447 if (hgetlo(afs_indexCounter) < tstat.atime) {
3448 hset32(afs_indexCounter, tstat.atime);
3450 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3451 } /*File is not bad */
3453 osi_UFSClose(tfile);
3454 tdc->f.states &= ~DWriting;
3455 tdc->dflags &= ~DFEntryMod;
3456 /* don't set f.modTime; we're just cleaning up */
3457 afs_WriteDCache(tdc, 0);
3458 ReleaseWriteLock(&afs_xdcache);
3459 ReleaseWriteLock(&tdc->lock);
3461 afs_stats_cmperf.cacheNumEntries++;
3466 /*Max # of struct dcache's resident at any time*/
3468 * If 'dchint' is enabled then in-memory dcache min is increased because of
3474 * Initialize dcache related variables.
3484 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3486 register struct dcache *tdp;
3490 afs_freeDCList = NULLIDX;
3491 afs_discardDCList = NULLIDX;
3492 afs_freeDCCount = 0;
3493 afs_freeDSList = NULL;
3494 hzero(afs_indexCounter);
3496 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3502 if (achunk < 0 || achunk > 30)
3503 achunk = 13; /* Use default */
3504 AFS_SETCHUNKSIZE(achunk);
3510 if (aflags & AFSCALL_INIT_MEMCACHE) {
3512 * Use a memory cache instead of a disk cache
3514 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3515 afs_cacheType = &afs_MemCacheOps;
3516 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3517 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3518 /* ablocks is reported in 1K blocks */
3519 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3521 printf("afsd: memory cache too large for available memory.\n");
3522 printf("afsd: AFS files cannot be accessed.\n\n");
3524 afiles = ablocks = 0;
3526 printf("Memory cache: Allocating %d dcache entries...",
3529 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3530 afs_cacheType = &afs_UfsCacheOps;
3533 if (aDentries > 512)
3534 afs_dhashsize = 2048;
3535 /* initialize hash tables */
3537 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3539 (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3540 for (i = 0; i < afs_dhashsize; i++) {
3541 afs_dvhashTbl[i] = NULLIDX;
3542 afs_dchashTbl[i] = NULLIDX;
3544 afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3545 afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
3546 for (i = 0; i < afiles; i++) {
3547 afs_dvnextTbl[i] = NULLIDX;
3548 afs_dcnextTbl[i] = NULLIDX;
3551 /* Allocate and zero the pointer array to the dcache entries */
3552 afs_indexTable = (struct dcache **)
3553 afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3554 memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3556 (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3557 memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3559 (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
3560 memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3561 afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
3562 memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
3564 /* Allocate and thread the struct dcache entries themselves */
3565 tdp = afs_Initial_freeDSList =
3566 (struct dcache *)afs_osi_Alloc(aDentries * sizeof(struct dcache));
3567 memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
3568 #ifdef KERNEL_HAVE_PIN
3569 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3570 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3571 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3572 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3573 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3574 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3575 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3576 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3577 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3580 afs_freeDSList = &tdp[0];
3581 for (i = 0; i < aDentries - 1; i++) {
3582 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3583 RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3584 RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3585 RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3587 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3588 RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3589 RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3590 RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3592 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3593 afs_cacheBlocks = ablocks;
3594 afs_ComputeCacheParms(); /* compute parms based on cache size */
3596 afs_dcentries = aDentries;
3598 afs_stats_cmperf.cacheBucket0_Discarded =
3599 afs_stats_cmperf.cacheBucket1_Discarded =
3600 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3606 * Shuts down the cache.
3610 shutdown_dcache(void)
3614 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3615 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3616 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3617 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3618 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3619 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3620 afs_osi_Free(afs_Initial_freeDSList,
3621 afs_dcentries * sizeof(struct dcache));
3622 #ifdef KERNEL_HAVE_PIN
3623 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3624 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3625 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3626 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3627 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3628 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3629 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3633 for (i = 0; i < afs_dhashsize; i++) {
3634 afs_dvhashTbl[i] = NULLIDX;
3635 afs_dchashTbl[i] = NULLIDX;
3638 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3639 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3641 afs_blocksUsed = afs_dcentries = 0;
3642 afs_stats_cmperf.cacheBucket0_Discarded =
3643 afs_stats_cmperf.cacheBucket1_Discarded =
3644 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3645 hzero(afs_indexCounter);
3647 afs_freeDCCount = 0;
3648 afs_freeDCList = NULLIDX;
3649 afs_discardDCList = NULLIDX;
3650 afs_freeDSList = afs_Initial_freeDSList = 0;
3652 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3657 #if defined(AFS_DISCON_ENV)
3660 * Make a shadow copy of a dir's dcaches. It's used for disconnected
3661 * operations like remove/create/rename to keep the original directory data.
3662 * On reconnection, we can diff the original data with the server and get the
3663 * server changes and with the local data to get the local changes.
3665 * \param avc The dir vnode.
3667 * \return 0 for success.
3669 * \note The only lock allowed to be set is the dir's vcache entry, and it
3670 * must be set in write mode.
3671 * \note The vcache entry must be write locked.
3673 int afs_MakeShadowDir(struct vcache *avc)
3675 int j, i, index, code, ret_code = 0, offset, trans_size, block;
3676 struct dcache *tdc, *new_dc = NULL;
3677 struct osi_file *tfile_src, *tfile_dst;
3678 struct VenusFid shadow_fid;
3682 /* Is this a dir? */
3683 if (vType(avc) != VDIR)
3686 /* Generate a fid for the shadow dir. */
3687 shadow_fid.Cell = avc->fid.Cell;
3688 shadow_fid.Fid.Volume = avc->fid.Fid.Volume;
3689 afs_GenShadowFid(&shadow_fid);
3691 /* For each dcache, do copy it into a new fresh one. */
3692 i = DVHash(&avc->fid);
3693 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
3694 /* Making sure that this isn't going to get locked twice. */
3696 /* XXX: Moved it from outside of the loop.
3697 * Maybe it's not quite okay because of the use of
3698 * dvhashTbl (once) in the for statement.
3700 ObtainWriteLock(&afs_xdcache, 716);
3704 i = afs_dvnextTbl[index];
3705 if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
3706 tdc = afs_GetDSlot(index, NULL);
3708 ReleaseReadLock(&tdc->tlock);
3710 if (!FidCmp(&tdc->f.fid, &avc->fid)) {
3712 /* Got a dir's dcache. */
3715 /* Get a fresh dcache. */
3716 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3718 /* Unlock hash for now. Don't need it during operations on the
3719 * dcache. Oh, and we can't use it because of the locking
3722 /* XXX: So much for lock ierarchy, the afs_AllocDCache doesn't
3725 //ReleaseWriteLock(&afs_xdcache);
3727 ObtainReadLock(&tdc->lock);
3729 /* Set up the new fid. */
3730 /* Copy interesting data from original dir dcache. */
3731 new_dc->mflags = tdc->mflags;
3732 new_dc->dflags = tdc->dflags;
3733 new_dc->f.modTime = tdc->f.modTime;
3734 new_dc->f.versionNo = tdc->f.versionNo;
3735 new_dc->f.states = tdc->f.states;
3736 new_dc->f.chunk= tdc->f.chunk;
3737 new_dc->f.chunkBytes = tdc->f.chunkBytes;
3740 * Now add to the two hash chains - note that i is still set
3741 * from the above DCHash call.
3743 //ObtainWriteLock(&afs_xdcache, 713);
3745 j = DCHash(&shadow_fid, 0);
3746 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[j];
3747 afs_dchashTbl[j] = new_dc->index;
3749 j = DVHash(&shadow_fid);
3750 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[j];
3751 afs_dvhashTbl[j] = new_dc->index;
3752 afs_MaybeWakeupTruncateDaemon();
3754 ReleaseWriteLock(&afs_xdcache);
3756 /* Alloc a 4k block. */
3757 data = (char *) afs_osi_Alloc(4096);
3759 printf("afs_MakeShadowDir: could not alloc data\n");
3764 /* Open the files. */
3765 tfile_src = afs_CFileOpen(tdc->f.inode);
3766 tfile_dst = afs_CFileOpen(new_dc->f.inode);
3768 /* Init no of blocks to be read and offset. */
3769 block = (tdc->f.chunkBytes / 4096);
3772 /* And now copy dir dcache data into this dcache,
3775 while (block >= 0) {
3777 /* Last chunk might have less bytes to transfer. */
3780 trans_size = (tdc->f.chunkBytes % 4096);
3782 /* An exact no of 4k blocks. */
3787 /* Read a chunk from the dcache. */
3788 code = afs_CFileRead(tfile_src, offset, data, trans_size);
3789 if (code < trans_size) {
3790 /* Can't access file, stop doing stuff and return error. */
3795 /* Write it to the new dcache. */
3796 code = afs_CFileWrite(tfile_dst, offset, data, trans_size);
3797 if (code < trans_size) {
3804 } /* while (block) */
3806 afs_CFileClose(tfile_dst);
3807 afs_CFileClose(tfile_src);
3809 afs_osi_Free(data, 4096);
3811 ReleaseWriteLock(&new_dc->lock);
3812 ReleaseReadLock(&tdc->lock);
3814 afs_PutDCache(new_dc);
3815 } /* if dcache fid match */
3817 } /* if unuiquifier match */
3821 ReleaseWriteLock(&afs_xdcache);
3824 if (!avc->ddirty_flags) {
3825 ObtainWriteLock(&afs_DDirtyVCListLock, 763);
3826 AFS_DISCON_ADD_DIRTY(avc);
3827 ReleaseWriteLock(&afs_DDirtyVCListLock);
3829 avc->shVnode = shadow_fid.Fid.Vnode;
3830 avc->shUnique = shadow_fid.Fid.Unique;
3831 avc->ddirty_flags |= VDisconShadowed;
3838 * Delete the dcaches of a shadow dir.
3840 * \param avc The vcache containing the shadow fid.
3842 * \note avc must be write locked.
3844 void afs_DeleteShadowDir(struct vcache *avc)
3847 struct VenusFid shadow_fid;
3849 shadow_fid.Cell = avc->fid.Cell;
3850 shadow_fid.Fid.Volume = avc->fid.Fid.Volume;
3851 shadow_fid.Fid.Vnode = avc->shVnode;
3852 shadow_fid.Fid.Unique = avc->shUnique;
3854 tdc = afs_FindDCacheByFid(&shadow_fid);
3856 afs_HashOutDCache(tdc, 1);
3857 afs_DiscardDCache(tdc);
3860 /* Remove shadowed dir flag. */
3861 avc->ddirty_flags &= ~VDisconShadowed;