/*
* Implements:
*/
-#include "../afs/param.h" /*Should be always first*/
+#include <afsconfig.h>
+#include "../afs/param.h"
+
+RCSID("$Header$");
+
#include "../afs/sysincludes.h" /*Standard vendor system headers*/
#include "../afs/afsincludes.h" /*AFS-based standard headers*/
#include "../afs/afs_stats.h" /* statistics */
static void afs_GetDownD(int anumber, int *aneedSpace);
static void afs_FreeDiscardedDCache(void);
static void afs_DiscardDCache(struct dcache *);
+static void afs_FreeDCache(struct dcache *);
/* Imported variables */
extern afs_rwlock_t afs_xvcache;
* --------------------- Exported definitions ---------------------
*/
afs_lock_t afs_xdcache; /*Lock: alloc new disk cache entries*/
-afs_int32 afs_freeDCList; /*Free list for disk cache entries*/
-afs_int32 afs_freeDCCount; /*Count of elts in freeDCList*/
+afs_int32 afs_freeDCList; /*Free list for disk cache entries*/
+afs_int32 afs_freeDCCount; /*Count of elts in freeDCList*/
afs_int32 afs_discardDCList; /*Discarded disk cache entries*/
afs_int32 afs_discardDCCount; /*Count of elts in discardDCList*/
struct dcache *afs_freeDSList; /*Free list for disk slots */
struct osi_file *afs_cacheInodep = 0; /* file for CacheItems inode */
struct afs_q afs_DLRU; /*dcache LRU*/
afs_int32 afs_dhashsize = 1024;
-afs_int32 *afs_dvhashTbl; /*Data cache hash table*/
-afs_int32 *afs_dchashTbl; /*Data cache hash table*/
-afs_int32 *afs_dvnextTbl; /*Dcache hash table links */
-afs_int32 *afs_dcnextTbl; /*Dcache hash table links */
+afs_int32 *afs_dvhashTbl; /*Data cache hash table*/
+afs_int32 *afs_dchashTbl; /*Data cache hash table*/
+afs_int32 *afs_dvnextTbl; /*Dcache hash table links */
+afs_int32 *afs_dcnextTbl; /*Dcache hash table links */
struct dcache **afs_indexTable; /*Pointers to dcache entries*/
-afs_hyper_t *afs_indexTimes; /*Dcache entry Access times*/
-afs_int32 *afs_indexUnique; /*dcache entry Fid.Unique */
+afs_hyper_t *afs_indexTimes; /*Dcache entry Access times*/
+afs_int32 *afs_indexUnique; /*dcache entry Fid.Unique */
unsigned char *afs_indexFlags; /*(only one) Is there data there?*/
-afs_hyper_t afs_indexCounter; /*Fake time for marking index
+afs_hyper_t afs_indexCounter; /*Fake time for marking index
entries*/
afs_int32 afs_cacheFiles =0; /*Size of afs_indexTable*/
-afs_int32 afs_cacheBlocks; /*1K blocks in cache*/
-afs_int32 afs_cacheStats; /*Stat entries in cache*/
-afs_int32 afs_blocksUsed; /*Number of blocks in use*/
+afs_int32 afs_cacheBlocks; /*1K blocks in cache*/
+afs_int32 afs_cacheStats; /*Stat entries in cache*/
+afs_int32 afs_blocksUsed; /*Number of blocks in use*/
afs_int32 afs_blocksDiscarded; /*Blocks freed but not truncated */
-afs_int32 afs_fsfragsize = 1023; /*Underlying Filesystem minimum unit
+afs_int32 afs_fsfragsize = 1023; /*Underlying Filesystem minimum unit
*of disk allocation usually 1K
*this value is (truefrag -1 ) to
*save a bunch of subtracts... */
+#ifdef AFS_64BIT_CLIENT
+#ifdef AFS_VM_RDWR_ENV
+afs_size_t afs_vmMappingEnd; /* for large files (>= 2GB) the VM
+ * mapping an 32bit addressing machines
+ * can only be used below the 2 GB
+ * line. From this point upwards we
+ * must do direct I/O into the cache
+ * files. The value should be on a
+ * chunk boundary. */
+#endif /* AFS_VM_RDWR_ENV */
+#endif /* AFS_64BIT_CLIENT */
/* The following is used to ensure that new dcache's aren't obtained when
* the cache is nearly full.
int afs_TruncateDaemonRunning = 0;
int afs_CacheTooFull = 0;
-afs_int32 afs_dcentries; /* In-memory dcache entries */
+afs_int32 afs_dcentries; /* In-memory dcache entries */
int dcacheDisabled = 0;
}
} /*afs_StoreWarn*/
+void afs_MaybeWakeupTruncateDaemon() {
+ if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
+ afs_CacheTooFull = 1;
+ if (!afs_TruncateDaemonRunning)
+ afs_osi_Wakeup((char *)afs_CacheTruncateDaemon);
+ } else if (!afs_TruncateDaemonRunning &&
+ afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
+ afs_osi_Wakeup((char *)afs_CacheTruncateDaemon);
+ }
+}
+
/* Keep statistics on run time for afs_CacheTruncateDaemon. This is a
* struct so we need only export one symbol for AIX.
*/
afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
}
if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
+#ifdef AFS_AFSDB_ENV
+ afs_termState = AFSOP_STOP_AFSDB;
+#else
afs_termState = AFSOP_STOP_RXEVENT;
+#endif
afs_osi_Wakeup(&afs_termState);
break;
}
*/
void
-afs_AdjustSize(adc, anewSize)
+afs_AdjustSize(adc, newSize)
register struct dcache *adc;
- register afs_int32 anewSize;
+ register afs_int32 newSize;
{ /*afs_AdjustSize*/
register afs_int32 oldSize;
AFS_STATCNT(afs_AdjustSize);
- adc->flags |= DFEntryMod;
+
+ adc->dflags |= DFEntryMod;
oldSize = ((adc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
- adc->f.chunkBytes = anewSize;
- anewSize = ((anewSize + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
- if (anewSize > oldSize) {
+ adc->f.chunkBytes = newSize;
+ newSize = ((newSize + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
+ if (newSize > oldSize) {
/* We're growing the file, wakeup the daemon */
afs_MaybeWakeupTruncateDaemon();
}
- afs_blocksUsed += (anewSize - oldSize);
+ afs_blocksUsed += (newSize - oldSize);
afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
} /*afs_AdjustSize*/
*
* Environment:
* The anumber parameter is just a hint; at least one entry MUST be
- * moved, of we'll panic. We must be called with afs_xdcache
+ * moved, or we'll panic. We must be called with afs_xdcache
* write-locked. We should try to satisfy both anumber and aneedspace,
* whichever is more demanding - need to do several things:
* 1. only grab up to anumber victims if aneedSpace <= 0, not
/* first, hold all the victims, since we're going to release the lock
* during the truncate operation.
*/
- for(i=0; i < victimPtr; i++)
- victimDCs[i] = afs_GetDSlot(victims[i], 0);
+ for(i=0; i < victimPtr; i++) {
+ tdc = afs_GetDSlot(victims[i], 0);
+ /* We got tdc->tlock(R) here */
+ if (tdc->refCount == 1)
+ victimDCs[i] = tdc;
+ else
+ victimDCs[i] = 0;
+ ReleaseReadLock(&tdc->tlock);
+ if (!victimDCs[i]) afs_PutDCache(tdc);
+ }
for(i = 0; i < victimPtr; i++) {
/* q is first elt in dcache entry */
tdc = victimDCs[i];
* compare with 1, since we bumped it above when we called
* afs_GetDSlot to preserve the entry's identity.
*/
- if (tdc->refCount == 1) {
+ if (tdc && tdc->refCount == 1) {
unsigned char chunkFlags;
+ afs_size_t tchunkoffset;
afid = &tdc->f.fid;
/* xdcache is lower than the xvcache lock */
MReleaseWriteLock(&afs_xdcache);
skip = 0;
if (tdc->refCount > 1) skip = 1;
if (tvc) {
+ tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
chunkFlags = afs_indexFlags[tdc->index];
if (phase == 0 && osi_Active(tvc)) skip = 1;
if (phase > 0 && osi_Active(tvc) && (tvc->states & CDCLock)
if (chunkFlags & IFDataMod) skip = 1;
afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
- ICL_TYPE_INT32,
- (afs_int32)(chunkFlags & IFDirtyPages),
- ICL_TYPE_INT32, AFS_CHUNKTOBASE(tdc->f.chunk));
+ ICL_TYPE_INT32, tdc->index,
+ ICL_TYPE_OFFSET,
+ ICL_HANDLE_OFFSET(tchunkoffset));
#if defined(AFS_SUN5_ENV)
/*
if (skip) {
/* skip this guy and mark him as recently used */
afs_indexFlags[tdc->index] |= IFFlag;
+ afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
+ ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
+ ICL_TYPE_INT32, tdc->index,
+ ICL_TYPE_OFFSET,
+ ICL_HANDLE_OFFSET(tchunkoffset));
}
else {
/* flush this dude from the data cache and reclaim;
* melt it down for parts. Note that any concurrent
* (new possibility!) calls to GetDownD won't touch
* this guy because his reference count is > 0. */
+ afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
+ ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
+ ICL_TYPE_INT32, tdc->index,
+ ICL_TYPE_OFFSET,
+ ICL_HANDLE_OFFSET(tchunkoffset));
#ifndef AFS_DEC_ENV
AFS_STATCNT(afs_gget);
#endif
j = 1; /* we reclaimed at least one victim */
}
}
-#ifdef AFS_SUN5_ENVX
afs_PutDCache(tdc);
-#else
- tdc->refCount--; /* put it back */
-#endif
}
if (phase == 0) {
adc->f.fid.Fid.Volume = 0; /* invalid */
/* mark entry as modified */
- adc->flags |= DFEntryMod;
+ adc->dflags |= DFEntryMod;
/* all done */
return 0;
*
* Environment: called with afs_xdcache lock write-locked.
*/
-afs_FreeDCache(adc)
-register struct dcache *adc; {
+static void afs_FreeDCache(adc)
+ register struct dcache *adc;
+{
/* Thread on free list, update free list count and mark entry as
* freed in its indexFlags element. Also, ensure DCache entry gets
* written out (set DFEntryMod).
afs_freeDCList = adc->index;
afs_freeDCCount++;
afs_indexFlags[adc->index] |= IFFree;
- adc->flags |= DFEntryMod;
+ adc->dflags |= DFEntryMod;
if (afs_WaitForCacheDrain) {
if ((afs_blocksUsed - afs_blocksDiscarded) <=
afs_osi_Wakeup(&afs_WaitForCacheDrain);
}
}
- return 0;
}
/*
*
* Parameters:
* adc : Ptr to dcache entry.
+ *
+ * Environment:
+ * Must be called with afs_xdcache write-locked.
*/
static void
register afs_int32 size;
AFS_STATCNT(afs_DiscardDCache);
+
+ osi_Assert(adc->refCount == 1);
+
size = ((adc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
afs_blocksDiscarded += size;
afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
afs_discardDCCount++;
adc->f.fid.Fid.Volume = 0;
- adc->flags |= DFEntryMod;
+ adc->dflags |= DFEntryMod;
afs_indexFlags[adc->index] |= IFDiscarded;
if (afs_WaitForCacheDrain) {
* Get an entry from the list of discarded cache elements
*/
tdc = afs_GetDSlot(afs_discardDCList, 0);
+ osi_Assert(tdc->refCount == 1);
+ ReleaseReadLock(&tdc->tlock);
+
afs_discardDCList = afs_dvnextTbl[tdc->index];
afs_dvnextTbl[tdc->index] = NULLIDX;
afs_discardDCCount--;
size = ((tdc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;/* round up */
afs_blocksDiscarded -= size;
afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
+ /* We can lock because we just took it off the free list */
+ ObtainWriteLock(&tdc->lock, 626);
MReleaseWriteLock(&afs_xdcache);
/*
MObtainWriteLock(&afs_xdcache,511);
afs_indexFlags[tdc->index] &= ~IFDiscarded;
afs_FreeDCache(tdc);
- tdc->refCount--;
+ ReleaseWriteLock(&tdc->lock);
+ afs_PutDCache(tdc);
MReleaseWriteLock(&afs_xdcache);
}
*
* Parameters:
* anumber : Targeted number of disk slots to free up.
+ *
+ * Environment:
+ * Must be called with afs_xdcache write-locked.
*/
#if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
extern SV_TYPE afs_sgibksync;
QRemove(&tdc->lruq);
/* write-through if modified */
- if (tdc->flags & DFEntryMod) {
+ if (tdc->dflags & DFEntryMod) {
#if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
/*
* ask proxy to do this for us - we don't have the stack space
*/
- while (tdc->flags & DFEntryMod) {
+ while (tdc->dflags & DFEntryMod) {
int s;
AFS_GUNLOCK();
s = SPLOCK(afs_sgibklock);
AFS_GLOCK();
}
#else
- tdc->flags &= ~DFEntryMod;
+ tdc->dflags &= ~DFEntryMod;
afs_WriteDCache(tdc, 1);
#endif
}
} /*afs_GetDownDSlot*/
+/*
+ * afs_RefDCache
+ *
+ * Description:
+ * Increment the reference count on a disk cache entry,
+ * which already has a non-zero refcount. In order to
+ * increment the refcount of a zero-reference entry, you
+ * have to hold afs_xdcache.
+ *
+ * Parameters:
+ * adc : Pointer to the dcache entry to increment.
+ *
+ * Environment:
+ * Nothing interesting.
+ */
+afs_RefDCache(adc)
+ struct dcache *adc;
+{
+ ObtainWriteLock(&adc->tlock, 627);
+ if (adc->refCount < 0)
+ osi_Panic("RefDCache: negative refcount");
+ adc->refCount++;
+ ReleaseWriteLock(&adc->tlock);
+ return 0;
+}
+
/*
* afs_PutDCache
* Environment:
* Nothing interesting.
*/
-afs_PutDCache(ad)
- register struct dcache *ad;
+afs_PutDCache(adc)
+ register struct dcache *adc;
{ /*afs_PutDCache*/
AFS_STATCNT(afs_PutDCache);
-#ifndef AFS_SUN5_ENVX
- MObtainWriteLock(&afs_xdcache,276);
-#endif
- if (ad->refCount <= 0)
+ ObtainWriteLock(&adc->tlock, 276);
+ if (adc->refCount <= 0)
osi_Panic("putdcache");
- --ad->refCount;
-#ifdef AFS_SUN5_ENVX
- MReleaseWriteLock(&ad->lock);
-#else
- MReleaseWriteLock(&afs_xdcache);
-#endif
+ --adc->refCount;
+ ReleaseWriteLock(&adc->tlock);
return 0;
} /*afs_PutDCache*/
register int i;
AFS_STATCNT(afs_TryToSmush);
afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
- ICL_TYPE_INT32, avc->m.Length);
+ ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length));
sync = 1; /* XX Temp testing XX*/
#if defined(AFS_SUN5_ENV)
for(index = afs_dvhashTbl[i]; index != NULLIDX; index=i) {
i = afs_dvnextTbl[index]; /* next pointer this hash table */
if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
+ int releaseTlock = 1;
tdc = afs_GetDSlot(index, (struct dcache *)0);
if (!FidCmp(&tdc->f.fid, &avc->fid)) {
if (sync) {
if ((afs_indexFlags[index] & IFDataMod) == 0 &&
tdc->refCount == 1) {
+ ReleaseReadLock(&tdc->tlock);
+ releaseTlock = 0;
afs_FlushDCache(tdc);
}
} else
afs_indexTable[index] = 0;
}
- lockedPutDCache(tdc);
+ if (releaseTlock) ReleaseReadLock(&tdc->tlock);
+ afs_PutDCache(tdc);
}
}
#if defined(AFS_SUN5_ENV)
struct dcache *afs_FindDCache(avc, abyte)
register struct vcache *avc;
- afs_int32 abyte;
+ afs_size_t abyte;
{ /*afs_FindDCache*/
for(index = afs_dchashTbl[i]; index != NULLIDX;) {
if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
tdc = afs_GetDSlot(index, (struct dcache *)0);
+ ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
break; /* leaving refCount high for caller */
}
- lockedPutDCache(tdc);
+ afs_PutDCache(tdc);
}
index = afs_dcnextTbl[index];
}
register struct rx_call *acall;
struct osi_file *afile;
register afs_int32 alen;
+ afs_size_t *abytesToXferP;
+ afs_size_t *abytesXferredP;
struct vcache *avc;
int *shouldWake;
- afs_int32 *abytesToXferP;
- afs_int32 *abytesXferredP;
{ /* afs_UFSCacheStoreProc*/
afs_int32 code, got;
(*abytesXferredP) = 0;
#endif /* AFS_NOSTATS */
- afs_Trace3(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
- ICL_TYPE_INT32, avc->m.Length, ICL_TYPE_INT32, alen);
+ afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
+ ICL_TYPE_FID, &(avc->fid),
+ ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length),
+ ICL_TYPE_INT32, alen);
tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
while (alen > 0) {
tlen = (alen > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : alen);
got = afs_osi_Read(afile, -1, tbuffer, tlen);
if ((got < 0)
-#if !defined(AFS_SUN5_ENV) && !defined(AFS_OSF_ENV) && !defined(AFS_SGI64_ENV) && !defined(AFS_LINUX20_ENV)
+#if !defined(AFS_SUN5_ENV) && !defined(AFS_OSF_ENV) && !defined(AFS_SGI64_ENV) && !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN_ENV) && !defined(AFS_FBSD_ENV)
|| (got != tlen && getuerror())
#endif
) {
osi_FreeLargeSpace(tbuffer);
return EIO;
}
-#ifdef RX_ENABLE_LOCKS
- AFS_GUNLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ afs_Trace1(afs_iclSetp, CM_TRACE_STOREPROC2, ICL_TYPE_INT32, got);
+ RX_AFS_GUNLOCK();
code = rx_Write(acall, tbuffer, got); /* writing 0 bytes will
* push a short packet. Is that really what we want, just because the
* data didn't come back from the disk yet? Let's try it and see. */
-#ifdef RX_ENABLE_LOCKS
- AFS_GLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ RX_AFS_GLOCK();
#ifndef AFS_NOSTATS
(*abytesXferredP) += code;
#endif /* AFS_NOSTATS */
afs_wakeup(avc);
}
}
+ afs_Trace4(afs_iclSetp, CM_TRACE_STOREPROC, ICL_TYPE_POINTER, avc,
+ ICL_TYPE_FID, &(avc->fid),
+ ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->m.Length),
+ ICL_TYPE_INT32, alen);
osi_FreeLargeSpace(tbuffer);
return 0;
* acall : Ptr to the Rx call structure.
* afile : File descriptor for the cache file.
* abase : Base offset to fetch.
- * adc : Ptr to the dcache entry for the file.
+ * adc : Ptr to the dcache entry for the file, write-locked.
* avc : Ptr to the vcache entry for the file.
* abytesToXferP : Set to the number of bytes to xfer.
* NOTE: This parameter is only used if AFS_NOSTATS
*/
static int afs_UFSCacheFetchProc(acall, afile, abase, adc, avc,
- abytesToXferP, abytesXferredP)
+ abytesToXferP, abytesXferredP, lengthFound)
register struct rx_call *acall;
- afs_int32 abase;
+ afs_size_t abase;
+ afs_size_t *abytesToXferP;
+ afs_size_t *abytesXferredP;
struct dcache *adc;
struct vcache *avc;
struct osi_file *afile;
- afs_int32 *abytesToXferP;
- afs_int32 *abytesXferredP;
-
+ afs_int32 lengthFound;
{ /*UFS_CacheFetchProc*/
-
afs_int32 length;
register afs_int32 code;
register char *tbuffer;
register int tlen;
- int moredata;
+ int moredata = 0;
AFS_STATCNT(UFS_CacheFetchProc);
+ osi_Assert(WriteLocked(&adc->lock));
afile->offset = 0; /* Each time start from the beginning */
+ length = lengthFound;
#ifndef AFS_NOSTATS
(*abytesToXferP) = 0;
(*abytesXferredP) = 0;
#endif /* AFS_NOSTATS */
tbuffer = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
do {
-#ifdef RX_ENABLE_LOCKS
- AFS_GUNLOCK();
-#endif /* RX_ENABLE_LOCKS */
- code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
-#ifdef RX_ENABLE_LOCKS
- AFS_GLOCK();
-#endif /* RX_ENABLE_LOCKS */
- if (code != sizeof(afs_int32)) {
- osi_FreeLargeSpace(tbuffer);
- code = rx_Error(acall);
- return (code?code:-1); /* try to return code, not -1 */
+ if (moredata) {
+ RX_AFS_GUNLOCK();
+ code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
+ RX_AFS_GLOCK();
+ length = ntohl(length);
+ if (code != sizeof(afs_int32)) {
+ osi_FreeLargeSpace(tbuffer);
+ code = rx_Error(acall);
+ return (code?code:-1); /* try to return code, not -1 */
+ }
}
- length = ntohl(length);
/*
* The fetch protocol is extended for the AFS/DFS translator
* to allow multiple blocks of data, each with its own length,
#endif /* AFS_NOSTATS */
while (length > 0) {
tlen = (length > AFS_LRALLOCSIZ ? AFS_LRALLOCSIZ : length);
-#ifdef RX_ENABLE_LOCKS
- AFS_GUNLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ RX_AFS_GUNLOCK();
code = rx_Read(acall, tbuffer, tlen);
-#ifdef RX_ENABLE_LOCKS
- AFS_GLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ RX_AFS_GLOCK();
#ifndef AFS_NOSTATS
(*abytesXferredP) += code;
#endif /* AFS_NOSTATS */
if (code != tlen) {
osi_FreeLargeSpace(tbuffer);
+ afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64READ,
+ ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
+ ICL_TYPE_INT32, length);
return -34;
}
code = afs_osi_Write(afile, -1, tbuffer, tlen);
abase += tlen;
length -= tlen;
adc->validPos = abase;
- if (adc->flags & DFWaiting) {
- adc->flags &= ~DFWaiting;
- afs_osi_Wakeup(&adc->validPos);
- }
+ afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT,
+ ICL_TYPE_STRING, __FILE__,
+ ICL_TYPE_INT32, __LINE__,
+ ICL_TYPE_POINTER, adc,
+ ICL_TYPE_INT32, adc->dflags);
+ afs_osi_Wakeup(&adc->validPos);
}
} while (moredata);
osi_FreeLargeSpace(tbuffer);
* aflags : Settings as follows:
* 1 : Set locks
* 2 : Return after creating entry.
+ * 4 : called from afs_vnop_write.c
+ * *alen contains length of data to be written.
* OUT:
* aoffset : Set to the offset within the chunk where the resident
* byte is located.
struct AFSCallBack CallBack;
};
-/* these fields are protected by the lock on the vcache and luck
- * on the dcache */
-#define updateV2DC(l,v,d,src) { if (l) ObtainWriteLock(&((v)->lock),src);\
- if (hsame((v)->m.DataVersion, (d)->f.versionNo) && (v)->callback) { \
- (v)->quick.dc = (d); \
- (v)->quick.stamp = (d)->stamp = MakeStamp(); \
- (v)->quick.minLoc = AFS_CHUNKTOBASE((d)->f.chunk); \
- /* Don't think I need these next two lines forever */ \
- (v)->quick.len = (d)->f.chunkBytes; \
- (v)->h1.dchint = (d); } if(l) ReleaseWriteLock(&((v)->lock)); }
+/*
+ * Update the vnode-to-dcache hint if we can get the vnode lock
+ * right away. Assumes dcache entry is at least read-locked.
+ */
+void updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src) {
+ if (!lockVc || 0 == NBObtainWriteLock(&v->lock,src)) {
+ if (hsame(v->m.DataVersion, d->f.versionNo) && v->callback) {
+ v->quick.dc = d;
+ v->quick.stamp = d->stamp = MakeStamp();
+ v->quick.minLoc = AFS_CHUNKTOBASE(d->f.chunk);
+ /* Don't think I need these next two lines forever */
+ v->quick.len = d->f.chunkBytes;
+ v->h1.dchint = d;
+ }
+ if (lockVc) ReleaseWriteLock(&v->lock);
+ }
+}
struct dcache *afs_GetDCache(avc, abyte, areq, aoffset, alen, aflags)
- register struct vcache *avc; /*Held*/
- afs_int32 abyte;
+ register struct vcache *avc; /* Write-locked unless aflags & 1 */
+ afs_size_t abyte;
+ afs_size_t *aoffset, *alen;
int aflags;
- afs_int32 *aoffset, *alen;
register struct vrequest *areq;
{ /*afs_GetDCache*/
afs_int32 index;
afs_int32 us;
afs_int32 chunk;
- afs_int32 maxGoodLength; /* amount of good data at server */
+ afs_size_t maxGoodLength; /* amount of good data at server */
struct rx_call *tcall;
- afs_int32 Position = 0;
- afs_int32 size; /* size of segment to transfer */
- struct tlocal1 *tsmall;
+ afs_size_t Position = 0;
+#ifdef AFS_64BIT_CLIENT
+ afs_size_t tsize;
+#endif /* AFS_64BIT_CLIENT */
+ afs_int32 size, tlen; /* size of segment to transfer */
+ afs_size_t lengthFound; /* as returned from server */
+ struct tlocal1 *tsmall = 0;
register struct dcache *tdc;
register struct osi_file *file;
register struct conn *tc;
int downDCount = 0;
+ struct server *newCallback;
+ char setNewCallback;
+ char setVcacheStatus;
+ char doVcacheUpdate;
+ char slowPass = 0;
+ int doAdjustSize = 0;
+ int doReallyAdjustSize = 0;
+ int overWriteWholeChunk = 0;
+
XSTATS_DECLS
#ifndef AFS_NOSTATS
struct afs_stats_xferData *xferP; /* Ptr to this op's xfer struct */
osi_timeval_t xferStartTime, /*FS xfer start time*/
xferStopTime; /*FS xfer stop time*/
- afs_int32 bytesToXfer; /* # bytes to xfer*/
- afs_int32 bytesXferred; /* # bytes actually xferred*/
+ afs_size_t bytesToXfer; /* # bytes to xfer*/
+ afs_size_t bytesXferred; /* # bytes actually xferred*/
struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats*/
int fromReplica; /*Are we reading from a replica?*/
int numFetchLoops; /*# times around the fetch/analyze loop*/
if (dcacheDisabled)
return NULL;
+ setLocks = aflags & 1;
+
/*
* Determine the chunk number and offset within the chunk corresponding
* to the desired byte.
*/
- if (vType(avc) == VDIR) {
+ if (avc->fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
chunk = 0;
}
else {
chunk = AFS_CHUNK(abyte);
}
- setLocks = aflags & 1;
-
/* come back to here if we waited for the cache to drain. */
- RetryGetDCache:
+RetryGetDCache:
+
+ setNewCallback = setVcacheStatus = 0;
+
+ if (setLocks) {
+ if (slowPass)
+ ObtainWriteLock(&avc->lock, 616);
+ else
+ ObtainReadLock(&avc->lock);
+ }
+
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ */
+
shortcut = 0;
/* check hints first! (might could use bcmp or some such...) */
-
if (tdc = avc->h1.dchint) {
+ int dcLocked;
+
+ /*
+ * The locking order between afs_xdcache and dcache lock matters.
+ * The hint dcache entry could be anywhere, even on the free list.
+ * Locking afs_xdcache ensures that noone is trying to pull dcache
+ * entries from the free list, and thereby assuming them to be not
+ * referenced and not locked.
+ */
MObtainReadLock(&afs_xdcache);
- if ( (tdc->index != NULLIDX) && !FidCmp(&tdc->f.fid, &avc->fid) &&
+ dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
+
+ if (dcLocked &&
+ (tdc->index != NULLIDX) && !FidCmp(&tdc->f.fid, &avc->fid) &&
chunk == tdc->f.chunk &&
!(afs_indexFlags[tdc->index] & (IFFree|IFDiscarded))) {
/* got the right one. It might not be the right version, and it
/* All this code should be integrated better with what follows:
* I can save a good bit more time under a write lock if I do..
*/
- /* does avc need to be locked? */
- /* Note that the race labeled LOCKXXX is inconsequential: the xdcache
- * lock protects both the dcache slots AND the DLRU list. While
- * the slots and hash table and DLRU list all may change in the race,
- * THIS particular dcache structure cannot be recycled and its LRU
- * pointers must still be valid once we get the lock again. Still
- * we should either create another lock or invent a new method of
- * managing dcache structs -- CLOCK or something. */
- shortcut = 1;
-#ifdef AFS_SUN5_ENVX
- MObtainWriteLock(&tdc->lock,279);
-#endif
+ ObtainWriteLock(&tdc->tlock, 603);
tdc->refCount++;
- if (hsame(tdc->f.versionNo, avc->m.DataVersion)
- && !(tdc->flags & DFFetching)) {
- afs_stats_cmperf.dcacheHits++;
- MReleaseReadLock(&afs_xdcache);
+ ReleaseWriteLock(&tdc->tlock);
+
+ MReleaseReadLock(&afs_xdcache);
+ shortcut = 1;
- MObtainWriteLock(&afs_xdcache, 559); /* LOCKXXX */
+ if (hsame(tdc->f.versionNo, avc->m.DataVersion) &&
+ !(tdc->dflags & DFFetching)) {
+
+ afs_stats_cmperf.dcacheHits++;
+ MObtainWriteLock(&afs_xdcache, 559);
QRemove(&tdc->lruq);
QAdd(&afs_DLRU, &tdc->lruq);
MReleaseWriteLock(&afs_xdcache);
+
+ /* Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(S)
+ */
goto done;
}
-#ifdef AFS_SUN5_ENVX
- MReleaseWriteLock(&tdc->lock);
-#endif
+ } else {
+ if (dcLocked) ReleaseSharedLock(&tdc->lock);
+ MReleaseReadLock(&afs_xdcache);
}
- MReleaseReadLock(&afs_xdcache);
+
+ if (!shortcut)
+ tdc = 0;
}
- if (!shortcut)
- {
- /*
- * Hash on the [fid, chunk] and get the corresponding dcache index
- * after write-locking the dcache.
+ /* Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(S) if tdc
*/
- RetryLookup:
- i = DCHash(&avc->fid, chunk);
- afs_MaybeWakeupTruncateDaemon(); /* check to make sure our space is fine */
- MObtainWriteLock(&afs_xdcache,280);
- us = NULLIDX;
- for(index = afs_dchashTbl[i]; index != NULLIDX;) {
- if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, (struct dcache *)0);
- if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
- /* Move it up in the beginning of the list */
- if (afs_dchashTbl[i] != index) {
- afs_dcnextTbl[us] = afs_dcnextTbl[index];
- afs_dcnextTbl[index] = afs_dchashTbl[i];
- afs_dchashTbl[i] = index;
+
+ if (!tdc) { /* If the hint wasn't the right dcache entry */
+ /*
+ * Hash on the [fid, chunk] and get the corresponding dcache index
+ * after write-locking the dcache.
+ */
+RetryLookup:
+
+ /* Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ */
+
+ i = DCHash(&avc->fid, chunk);
+ /* check to make sure our space is fine */
+ afs_MaybeWakeupTruncateDaemon();
+
+ MObtainWriteLock(&afs_xdcache,280);
+ us = NULLIDX;
+ for (index = afs_dchashTbl[i]; index != NULLIDX; ) {
+ if (afs_indexUnique[index] == avc->fid.Fid.Unique) {
+ tdc = afs_GetDSlot(index, (struct dcache *)0);
+ ReleaseReadLock(&tdc->tlock);
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * afs_xdcache(W)
+ */
+ if (!FidCmp(&tdc->f.fid, &avc->fid) && chunk == tdc->f.chunk) {
+ /* Move it up in the beginning of the list */
+ if (afs_dchashTbl[i] != index) {
+ afs_dcnextTbl[us] = afs_dcnextTbl[index];
+ afs_dcnextTbl[index] = afs_dchashTbl[i];
+ afs_dchashTbl[i] = index;
+ }
+ MReleaseWriteLock(&afs_xdcache);
+ ObtainSharedLock(&tdc->lock, 606);
+ break; /* leaving refCount high for caller */
+ }
+ afs_PutDCache(tdc);
+ tdc = 0;
}
- MReleaseWriteLock(&afs_xdcache);
- break; /* leaving refCount high for caller */
+ us = index;
+ index = afs_dcnextTbl[index];
}
- lockedPutDCache(tdc);
- }
- us = index;
- index = afs_dcnextTbl[index];
- }
- /*
- * If we didn't find the entry, we'll create one.
- */
- if (index == NULLIDX) {
- afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER, avc,
- ICL_TYPE_INT32, chunk);
-
- if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
- while (1) {
- if (!setLocks) avc->states |= CDCLock;
- afs_GetDownD(5, (int*)0); /* just need slots */
- if (!setLocks) avc->states &= (~CDCLock);
- if (afs_discardDCList != NULLIDX || afs_freeDCList != NULLIDX)
- break;
- /* If we can't get space for 5 mins we give up and panic */
- if (++downDCount > 300)
- osi_Panic("getdcache");
- MReleaseWriteLock(&afs_xdcache);
- afs_osi_Wait(1000, 0, 0);
- goto RetryLookup;
+
+ /*
+ * If we didn't find the entry, we'll create one.
+ */
+ if (index == NULLIDX) {
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks
+ * avc->lock(W) if !setLocks
+ * afs_xdcache(W)
+ */
+ afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
+ avc, ICL_TYPE_INT32, chunk);
+
+ /* Make sure there is a free dcache entry for us to use */
+ if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
+ while (1) {
+ if (!setLocks) avc->states |= CDCLock;
+ afs_GetDownD(5, (int*)0); /* just need slots */
+ if (!setLocks) avc->states &= ~CDCLock;
+ if (afs_discardDCList != NULLIDX || afs_freeDCList != NULLIDX)
+ break;
+ /* If we can't get space for 5 mins we give up and panic */
+ if (++downDCount > 300)
+ osi_Panic("getdcache");
+ MReleaseWriteLock(&afs_xdcache);
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks
+ * avc->lock(W) if !setLocks
+ */
+ afs_osi_Wait(1000, 0, 0);
+ goto RetryLookup;
+ }
}
- }
- if (afs_discardDCList == NULLIDX ||
- ((aflags & 2) && afs_freeDCList != NULLIDX)) {
- afs_indexFlags[afs_freeDCList] &= ~IFFree;
- tdc = afs_GetDSlot(afs_freeDCList, 0);
- afs_freeDCList = afs_dvnextTbl[tdc->index];
- afs_freeDCCount--;
- } else {
- afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
- tdc = afs_GetDSlot(afs_discardDCList, 0);
- afs_discardDCList = afs_dvnextTbl[tdc->index];
- afs_discardDCCount--;
- size = ((tdc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;
- afs_blocksDiscarded -= size;
- afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
- if (aflags & 2) {
- /* Truncate the chunk so zeroes get filled properly */
- file = afs_CFileOpen(tdc->f.inode);
- afs_CFileTruncate(file, 0);
- afs_CFileClose(file);
- afs_AdjustSize(tdc, 0);
+
+ if (afs_discardDCList == NULLIDX ||
+ ((aflags & 2) && afs_freeDCList != NULLIDX)) {
+
+ afs_indexFlags[afs_freeDCList] &= ~IFFree;
+ tdc = afs_GetDSlot(afs_freeDCList, 0);
+ osi_Assert(tdc->refCount == 1);
+ ReleaseReadLock(&tdc->tlock);
+ ObtainWriteLock(&tdc->lock, 604);
+ afs_freeDCList = afs_dvnextTbl[tdc->index];
+ afs_freeDCCount--;
+ } else {
+ afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
+ tdc = afs_GetDSlot(afs_discardDCList, 0);
+ osi_Assert(tdc->refCount == 1);
+ ReleaseReadLock(&tdc->tlock);
+ ObtainWriteLock(&tdc->lock, 605);
+ afs_discardDCList = afs_dvnextTbl[tdc->index];
+ afs_discardDCCount--;
+ size = ((tdc->f.chunkBytes + afs_fsfragsize)^afs_fsfragsize)>>10;
+ afs_blocksDiscarded -= size;
+ afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
+ if (aflags & 2) {
+ /* Truncate the chunk so zeroes get filled properly */
+ file = afs_CFileOpen(tdc->f.inode);
+ afs_CFileTruncate(file, 0);
+ afs_CFileClose(file);
+ afs_AdjustSize(tdc, 0);
+ }
}
+
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks
+ * avc->lock(W) if !setLocks
+ * tdc->lock(W)
+ * afs_xdcache(W)
+ */
+
+ /*
+ * Fill in the newly-allocated dcache record.
+ */
+ afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
+ tdc->f.fid = avc->fid;
+ afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
+ hones(tdc->f.versionNo); /* invalid value */
+ tdc->f.chunk = chunk;
+ tdc->validPos = AFS_CHUNKTOBASE(chunk);
+ /* XXX */
+ if (tdc->lruq.prev == &tdc->lruq) osi_Panic("lruq 1");
+
+ /*
+ * Now add to the two hash chains - note that i is still set
+ * from the above DCHash call.
+ */
+ afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
+ afs_dchashTbl[i] = tdc->index;
+ i = DVHash(&avc->fid);
+ afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
+ afs_dvhashTbl[i] = tdc->index;
+ tdc->dflags = DFEntryMod;
+ tdc->mflags = 0;
+ tdc->f.states = 0;
+ afs_MaybeWakeupTruncateDaemon();
+ MReleaseWriteLock(&afs_xdcache);
+ ConvertWToSLock(&tdc->lock);
}
+ } /* vcache->dcache hint failed */
- /*
- * Fill in the newly-allocated dcache record.
- */
- afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
- tdc->f.fid = avc->fid;
- afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
- hones(tdc->f.versionNo); /* invalid value */
- tdc->f.chunk = chunk;
- /* XXX */
- if (tdc->lruq.prev == &tdc->lruq) osi_Panic("lruq 1");
- /*
- * Now add to the two hash chains - note that i is still set
- * from the above DCHash call.
- */
- afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
- afs_dchashTbl[i] = tdc->index;
- i = DVHash(&avc->fid);
- afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
- afs_dvhashTbl[i] = tdc->index;
- tdc->flags = DFEntryMod;
- tdc->f.states = 0;
- afs_MaybeWakeupTruncateDaemon();
- MReleaseWriteLock(&afs_xdcache);
- }
- } /* else hint failed... */
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(S)
+ */
afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
ICL_TYPE_POINTER, tdc,
ICL_TYPE_INT32, hgetlo(tdc->f.versionNo),
ICL_TYPE_INT32, hgetlo(avc->m.DataVersion));
/*
- * Here we have the unlocked entry in tdc, with its refCount
- * incremented. Note: we don't use the S-lock; it costs concurrency
- * when storing a file back to the server.
+ * Here we have the entry in tdc, with its refCount incremented.
+ * Note: we don't use the S-lock on avc; it costs concurrency when
+ * storing a file back to the server.
*/
- if (setLocks) ObtainReadLock(&avc->lock);
/*
* Not a newly created file so we need to check the file's length and
* read rpcs on newly created files (dv of 0) since only then we guarantee
* that this chunk's data hasn't been filled by another client.
*/
- if (!hiszero(avc->m.DataVersion))
- aflags &= ~4;
+ size = AFS_CHUNKSIZE(abyte);
+ if (aflags & 4) /* called from write */
+ tlen = *alen;
+ else /* called from read */
+ tlen = tdc->validPos - abyte;
+ Position = AFS_CHUNKTOBASE(chunk);
+ afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3,
+ ICL_TYPE_INT32, tlen,
+ ICL_TYPE_INT32, aflags,
+ ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(abyte),
+ ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(Position));
+ if ((aflags & 4) && (hiszero(avc->m.DataVersion)))
+ doAdjustSize = 1;
+ if ((aflags & 4) && (abyte == Position) && (tlen >= size))
+ overWriteWholeChunk = 1;
+ if (doAdjustSize || overWriteWholeChunk) {
#if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
-#ifdef AFS_SGI_ENV
-#ifdef AFS_SGI64_ENV
- if (aflags & 4) adjustsize = NBPP;
-#else
- if (aflags & 4) adjustsize = 8192;
-#endif
-#else
- if (aflags & 4) adjustsize = 4096;
-#endif
- if (AFS_CHUNKTOBASE(chunk)+adjustsize >= avc->m.Length &&
-#else
-#if defined(AFS_SUN_ENV) || defined(AFS_OSF_ENV)
- if (((aflags & 4) || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
-#else
- if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
-#endif
-#endif
- !hsame(avc->m.DataVersion, tdc->f.versionNo)) {
- /* no data in file to read at this position */
- if (setLocks) {
- ReleaseReadLock(&avc->lock);
- ObtainWriteLock(&avc->lock,64);
- }
- /* check again, now that we have a write lock */
-#if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
- if (AFS_CHUNKTOBASE(chunk)+adjustsize >= avc->m.Length &&
-#else
-#if defined(AFS_SUN_ENV) || defined(AFS_OSF_ENV)
- if (((aflags & 4) || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
-#else
- if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
-#endif
-#endif
- !hsame(avc->m.DataVersion, tdc->f.versionNo)) {
+ #ifdef AFS_SGI_ENV
+ #ifdef AFS_SGI64_ENV
+ if (doAdjustSize) adjustsize = NBPP;
+ #else /* AFS_SGI64_ENV */
+ if (doAdjustSize) adjustsize = 8192;
+ #endif /* AFS_SGI64_ENV */
+ #else /* AFS_SGI_ENV */
+ if (doAdjustSize) adjustsize = 4096;
+ #endif /* AFS_SGI_ENV */
+ if (AFS_CHUNKTOBASE(chunk)+adjustsize >= avc->m.Length &&
+#else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
+ #if defined(AFS_SUN_ENV) || defined(AFS_OSF_ENV)
+ if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->m.Length)) &&
+ #else
+ if (AFS_CHUNKTOBASE(chunk) >= avc->m.Length &&
+ #endif
+#endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
+ !hsame(avc->m.DataVersion, tdc->f.versionNo))
+ doReallyAdjustSize = 1;
+
+ if (doReallyAdjustSize || overWriteWholeChunk) {
+ /* no data in file to read at this position */
+ UpgradeSToWLock(&tdc->lock, 607);
+
file = afs_CFileOpen(tdc->f.inode);
afs_CFileTruncate(file, 0);
afs_CFileClose(file);
afs_AdjustSize(tdc, 0);
hset(tdc->f.versionNo, avc->m.DataVersion);
- tdc->flags |= DFEntryMod;
- }
- if (setLocks) {
- ReleaseWriteLock(&avc->lock);
- ObtainReadLock(&avc->lock);
+ tdc->dflags |= DFEntryMod;
+
+ ConvertWToSLock(&tdc->lock);
}
}
- if (setLocks) ReleaseReadLock(&avc->lock);
/*
- * We must read in the whole chunk iff the version number doesn't
+ * We must read in the whole chunk if the version number doesn't
* match.
*/
if (aflags & 2) {
/* don't need data, just a unique dcache entry */
+ ObtainWriteLock(&afs_xdcache, 608);
hset(afs_indexTimes[tdc->index], afs_indexCounter);
hadd32(afs_indexCounter, 1);
- updateV2DC(setLocks,avc,tdc,567);
+ ReleaseWriteLock(&afs_xdcache);
+
+ updateV2DC(setLocks, avc, tdc, 553);
+ if (vType(avc) == VDIR)
+ *aoffset = abyte;
+ else
+ *aoffset = AFS_CHUNKOFFSET(abyte);
+ if (tdc->validPos < abyte)
+ *alen = (afs_size_t) 0;
+ else
+ *alen = tdc->validPos - abyte;
+ ReleaseSharedLock(&tdc->lock);
+ if (setLocks) {
+ if (slowPass)
+ ReleaseWriteLock(&avc->lock);
+ else
+ ReleaseReadLock(&avc->lock);
+ }
return tdc; /* check if we're done */
}
- if (setLocks) ObtainReadLock(&avc->lock);
- if (!hsame(avc->m.DataVersion, tdc->f.versionNo)) {
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(S)
+ */
+ osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
+
+ setNewCallback = setVcacheStatus = 0;
+
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(S)
+ */
+ if (!hsame(avc->m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
/*
* Version number mismatch.
*/
- if (setLocks) {
- ReleaseReadLock(&avc->lock);
- ObtainWriteLock(&avc->lock,65);
- }
+ UpgradeSToWLock(&tdc->lock, 609);
/*
* If data ever existed for this vnode, and this is a text object,
/*
* By here, the cache entry is always write-locked. We can
* deadlock if we call osi_Flush with the cache entry locked...
+ * Unlock the dcache too.
*/
- ReleaseWriteLock(&avc->lock);
+ ReleaseWriteLock(&tdc->lock);
+ if (setLocks && !slowPass)
+ ReleaseReadLock(&avc->lock);
+ else
+ ReleaseWriteLock(&avc->lock);
+
osi_FlushText(avc);
/*
* Call osi_FlushPages in open, read/write, and map, since it
* is too hard here to figure out if we should lock the
* pvnLock.
*/
- ObtainWriteLock(&avc->lock,66);
+ if (setLocks && !slowPass)
+ ObtainReadLock(&avc->lock);
+ else
+ ObtainWriteLock(&avc->lock, 66);
+ ObtainWriteLock(&tdc->lock, 610);
}
- /* Watch for standard race condition */
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(W)
+ */
+
+ /* Watch for standard race condition around osi_FlushText */
if (hsame(avc->m.DataVersion, tdc->f.versionNo)) {
- updateV2DC(0,avc,tdc,569); /* set hint */
- if (setLocks) ReleaseWriteLock(&avc->lock);
- afs_stats_cmperf.dcacheHits++;
+ updateV2DC(setLocks, avc, tdc, 569); /* set hint */
+ afs_stats_cmperf.dcacheHits++;
+ ConvertWToSLock(&tdc->lock);
goto done;
}
/* Sleep here when cache needs to be drained. */
- if (setLocks &&
+ if (setLocks && !slowPass &&
(afs_blocksUsed > (CM_WAITFORDRAINPCT*afs_cacheBlocks)/100)) {
/* Make sure truncate daemon is running */
afs_MaybeWakeupTruncateDaemon();
+ ObtainWriteLock(&tdc->tlock, 614);
tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
- ReleaseWriteLock(&avc->lock);
+ ReleaseWriteLock(&tdc->tlock);
+ ReleaseWriteLock(&tdc->lock);
+ ReleaseReadLock(&avc->lock);
while ((afs_blocksUsed-afs_blocksDiscarded) >
(CM_WAITFORDRAINPCT*afs_cacheBlocks)/100) {
afs_WaitForCacheDrain = 1;
afs_MaybeFreeDiscardedDCache();
/* need to check if someone else got the chunk first. */
goto RetryGetDCache;
- }
+ }
/* Do not fetch data beyond truncPos. */
maxGoodLength = avc->m.Length;
file = afs_CFileOpen(tdc->f.inode);
afs_RemoveVCB(&avc->fid);
tdc->f.states |= DWriting;
- tdc->flags |= DFFetching;
- tdc->validPos = Position; /*Last valid position in this chunk*/
- if (tdc->flags & DFFetchReq) {
- tdc->flags &= ~DFFetchReq;
+ tdc->dflags |= DFFetching;
+ tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
+ if (tdc->mflags & DFFetchReq) {
+ tdc->mflags &= ~DFFetchReq;
afs_osi_Wakeup(&tdc->validPos);
}
tsmall = (struct tlocal1 *) osi_AllocLargeSpace(sizeof(struct tlocal1));
+ setVcacheStatus = 0;
#ifndef AFS_NOSTATS
/*
* Remember if we are doing the reading from a replicated volume,
(accP->unreplicatedRefs)++;
#endif /* AFS_NOSTATS */
/* this is a cache miss */
- afs_stats_cmperf.dcacheMisses++;
- afs_Trace3(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
- ICL_TYPE_INT32, Position, ICL_TYPE_INT32, size);
+ afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
+ ICL_TYPE_FID, &(avc->fid),
+ ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(Position),
+ ICL_TYPE_INT32, size);
+
+ if (size) afs_stats_cmperf.dcacheMisses++;
+ code = 0;
+ /*
+ * Dynamic root support: fetch data from local memory.
+ */
+ if (afs_IsDynroot(avc)) {
+ char *dynrootDir;
+ int dynrootLen;
+
+ afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
+
+ dynrootDir += Position;
+ dynrootLen -= Position;
+ if (size > dynrootLen)
+ size = dynrootLen;
+ if (size < 0) size = 0;
+ code = afs_CFileWrite(file, 0, dynrootDir, size);
+ afs_PutDynroot();
+
+ if (code == size)
+ code = 0;
+ else
+ code = -1;
+
+ tdc->validPos = Position + size;
+ afs_CFileTruncate(file, size); /* prune it */
+ } else
+ /*
+ * Not a dynamic vnode: do the real fetch.
+ */
do {
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(W)
+ */
+
tc = afs_Conn(&avc->fid, areq, SHARED_LOCK);
if (tc) {
+ afs_int32 length_hi, length, bytes;
#ifndef AFS_NOSTATS
numFetchLoops++;
if (fromReplica)
(accP->numReplicasAccessed)++;
#endif /* AFS_NOSTATS */
- avc->callback = tc->srvr->server;
- ConvertWToSLock(&avc->lock);
+ if (!setLocks || slowPass) {
+ avc->callback = tc->srvr->server;
+ } else {
+ newCallback = tc->srvr->server;
+ setNewCallback = 1;
+ }
i = osi_Time();
-#ifdef RX_ENABLE_LOCKS
- AFS_GUNLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ RX_AFS_GUNLOCK();
tcall = rx_NewCall(tc->id);
-#ifdef RX_ENABLE_LOCKS
- AFS_GLOCK();
-#endif /* RX_ENABLE_LOCKS */
-
+ RX_AFS_GLOCK();
XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_FETCHDATA);
-#ifdef RX_ENABLE_LOCKS
- AFS_GUNLOCK();
-#endif /* RX_ENABLE_LOCKS */
+#ifdef AFS_64BIT_CLIENT
+ length_hi = code = 0;
+ if (!afs_serverHasNo64Bit(tc)) {
+ tsize = size;
+ RX_AFS_GUNLOCK();
+ code = StartRXAFS_FetchData64(tcall,
+ (struct AFSFid *) &avc->fid.Fid,
+ Position, tsize);
+ if (code != 0) {
+ RX_AFS_GLOCK();
+ afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
+ ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code);
+ } else {
+ bytes = rx_Read(tcall, (char *)&length_hi, sizeof(afs_int32));
+ RX_AFS_GLOCK();
+ afs_Trace2(afs_iclSetp, CM_TRACE_FETCH64CODE,
+ ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code);
+ if (bytes == sizeof(afs_int32)) {
+ length_hi = ntohl(length_hi);
+ } else {
+ length_hi = 0;
+ code = rx_Error(tcall);
+ RX_AFS_GUNLOCK();
+ code1 = rx_EndCall(tcall, code);
+ RX_AFS_GLOCK();
+ tcall = (struct rx_call *) 0;
+ }
+ }
+ }
+ if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
+ if (Position > 0x7FFFFFFF) {
+ code = EFBIG;
+ } else {
+ afs_int32 pos;
+ pos = Position;
+ RX_AFS_GUNLOCK();
+ if (!tcall)
+ tcall = rx_NewCall(tc->id);
+ code = StartRXAFS_FetchData(tcall,
+ (struct AFSFid *) &avc->fid.Fid, pos, size);
+ RX_AFS_GLOCK();
+ }
+ afs_serverSetNo64Bit(tc);
+ }
+ if (code == 0) {
+ RX_AFS_GUNLOCK();
+ bytes = rx_Read(tcall, (char *)&length, sizeof(afs_int32));
+ RX_AFS_GLOCK();
+ if (bytes == sizeof(afs_int32)) {
+ length = ntohl(length);
+ } else {
+ code = rx_Error(tcall);
+ }
+ }
+ FillInt64(lengthFound, length_hi, length);
+ afs_Trace3(afs_iclSetp, CM_TRACE_FETCH64LENG,
+ ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code,
+ ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(lengthFound));
+#else /* AFS_64BIT_CLIENT */
+ RX_AFS_GUNLOCK();
code = StartRXAFS_FetchData(tcall,
(struct AFSFid *) &avc->fid.Fid,
Position, size);
-#ifdef RX_ENABLE_LOCKS
- AFS_GLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ RX_AFS_GLOCK();
+ if (code == 0) {
+ RX_AFS_GUNLOCK();
+ bytes = rx_Read(tcall, (char *)&length, sizeof(afs_int32));
+ RX_AFS_GLOCK();
+ if (bytes == sizeof(afs_int32)) {
+ length = ntohl(length);
+ } else {
+ code = rx_Error(tcall);
+ }
+ }
+#endif /* AFS_64BIT_CLIENT */
if (code == 0) {
#ifndef AFS_NOSTATS
xferP = &(afs_stats_cmfullperf.rpc.fsXferTimes[AFS_STATS_FS_XFERIDX_FETCHDATA]);
osi_GetuTime(&xferStartTime);
- code = afs_CacheFetchProc(tcall, file, Position, tdc, avc,
- &bytesToXfer, &bytesXferred);
+ code = afs_CacheFetchProc(tcall, file,
+ (afs_size_t) Position, tdc, avc,
+ &bytesToXfer, &bytesXferred, length);
osi_GetuTime(&xferStopTime);
(xferP->numXfers)++;
if (bytesXferred > xferP->maxBytes)
xferP->maxBytes = bytesXferred;
- /*
- * Tally the size of the object. Note: we tally the actual size,
- * NOT the number of bytes that made it out over the wire.
- */
+ /*
+ * Tally the size of the object. Note: we tally the actual size,
+ * NOT the number of bytes that made it out over the wire.
+ */
if (bytesToXfer <= AFS_STATS_MAXBYTES_BUCKET0)
(xferP->count[0])++;
else
if (afs_stats_TimeGreaterThan(elapsedTime, (xferP->maxTime))) {
afs_stats_TimeAssign((xferP->maxTime), elapsedTime);
}
- }
+ }
#else
- code = afs_CacheFetchProc(tcall, file, Position, tdc, avc, 0, 0);
+ code = afs_CacheFetchProc(tcall, file, Position, tdc, avc, 0, 0, length);
#endif /* AFS_NOSTATS */
}
if (code == 0) {
-#ifdef RX_ENABLE_LOCKS
- AFS_GUNLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ RX_AFS_GUNLOCK();
code = EndRXAFS_FetchData(tcall,
&tsmall->OutStatus,
&tsmall->CallBack,
&tsmall->tsync);
-#ifdef RX_ENABLE_LOCKS
- AFS_GLOCK();
-#endif /* RX_ENABLE_LOCKS */
+ RX_AFS_GLOCK();
}
XSTATS_END_TIME;
- code1 = rx_EndCall(tcall, code);
- UpgradeSToWLock(&avc->lock,27);
+ RX_AFS_GUNLOCK();
+ if (tcall)
+ code1 = rx_EndCall(tcall, code);
+ RX_AFS_GLOCK();
}
else {
code = -1;
* when we started. */
/*
* validPos is updated by CacheFetchProc, and can only be
- * modifed under an S or W lock, which we've blocked out
+ * modifed under a dcache write lock, which we've blocked out
*/
size = tdc->validPos - Position; /* actual segment size */
if (size < 0) size = 0;
afs_CFileTruncate(file, size); /* prune it */
}
- else {
- ObtainWriteLock(&afs_xcbhash, 453);
- afs_DequeueCallback(avc);
- avc->states &= ~(CStatd | CUnique);
- avc->callback = (struct server *)0;
- ReleaseWriteLock(&afs_xcbhash);
- if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
- osi_dnlc_purgedp(avc);
- }
+ else {
+ if (!setLocks || slowPass) {
+ ObtainWriteLock(&afs_xcbhash, 453);
+ afs_DequeueCallback(avc);
+ avc->states &= ~(CStatd | CUnique);
+ avc->callback = (struct server *)0;
+ ReleaseWriteLock(&afs_xcbhash);
+ if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
+ osi_dnlc_purgedp(avc);
+ } else {
+ /* Something lost. Forget about performance, and go
+ * back with a vcache write lock.
+ */
+ afs_CFileTruncate(file, 0);
+ afs_AdjustSize(tdc, 0);
+ afs_CFileClose(file);
+ osi_FreeLargeSpace(tsmall);
+ tsmall = 0;
+ ReleaseWriteLock(&tdc->lock);
+ afs_PutDCache(tdc);
+ tdc = 0;
+ ReleaseReadLock(&avc->lock);
+ slowPass = 1;
+ goto RetryGetDCache;
+ }
+ }
} while
(afs_Analyze(tc, code, &avc->fid, areq,
AFS_STATS_FS_RPCIDX_FETCHDATA,
SHARED_LOCK, (struct cell *)0));
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(W)
+ */
+
#ifndef AFS_NOSTATS
/*
* In the case of replicated access, jot down info on the number of
}
#endif /* AFS_NOSTATS */
- tdc->flags &= ~DFFetching;
- if (tdc->flags & DFWaiting) {
- tdc->flags &= ~DFWaiting;
- afs_osi_Wakeup(&tdc->validPos);
- }
+ tdc->dflags &= ~DFFetching;
+ afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT,
+ ICL_TYPE_STRING, __FILE__,
+ ICL_TYPE_INT32, __LINE__,
+ ICL_TYPE_POINTER, tdc,
+ ICL_TYPE_INT32, tdc->dflags);
+ afs_osi_Wakeup(&tdc->validPos);
if (avc->execsOrWriters == 0) tdc->f.states &= ~DWriting;
- /* now, if code != 0, we have an error and should punt */
+ /* now, if code != 0, we have an error and should punt.
+ * note that we have the vcache write lock, either because
+ * !setLocks or slowPass.
+ */
if (code) {
afs_CFileTruncate(file, 0);
afs_AdjustSize(tdc, 0);
if (vType(avc) == VDIR) {
DZap(&tdc->f.inode);
}
-#ifdef AFS_SUN5_ENVX
+ ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
-#else
- tdc->refCount--;
-#endif
ObtainWriteLock(&afs_xcbhash, 454);
afs_DequeueCallback(avc);
avc->states &= ~( CStatd | CUnique );
ReleaseWriteLock(&afs_xcbhash);
if (avc->fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
osi_dnlc_purgedp(avc);
- if (setLocks) ReleaseWriteLock(&avc->lock);
- osi_FreeLargeSpace(tsmall);
+ /*
+ * Locks held:
+ * avc->lock(W); assert(!setLocks || slowPass)
+ */
+ osi_Assert(!setLocks || slowPass);
tdc = (struct dcache *) 0;
goto done;
}
afs_CFileClose(file);
afs_AdjustSize(tdc, size); /* new size */
/*
- * Copy appropriate fields into vcache
+ * Copy appropriate fields into vcache. Status is
+ * copied later where we selectively acquire the
+ * vcache write lock.
*/
- afs_ProcessFS(avc, &tsmall->OutStatus, areq);
+ if (slowPass)
+ afs_ProcessFS(avc, &tsmall->OutStatus, areq);
+ else
+ setVcacheStatus = 1;
hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh, tsmall->OutStatus.DataVersion);
- tdc->flags |= DFEntryMod;
+ tdc->dflags |= DFEntryMod;
afs_indexFlags[tdc->index] |= IFEverUsed;
- if (setLocks) ReleaseWriteLock(&avc->lock);
- osi_FreeLargeSpace(tsmall);
+ ConvertWToSLock(&tdc->lock);
} /*Data version numbers don't match*/
else {
/*
- * Data version numbers match. Release locks if we locked
- * them, and remember we've had a cache hit.
+ * Data version numbers match.
*/
- if (setLocks)
- ReleaseReadLock(&avc->lock);
afs_stats_cmperf.dcacheHits++;
} /*Data version numbers match*/
- updateV2DC(setLocks,avc,tdc,332); /* set hint */
+ updateV2DC(setLocks, avc, tdc, 335); /* set hint */
done:
/*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ * tdc->lock(S) if tdc
+ */
+
+ /*
* See if this was a reference to a file in the local cell.
*/
if (avc->fid.Cell == LOCALCELL)
/* Fix up LRU info */
if (tdc) {
- hset(afs_indexTimes[tdc->index], afs_indexCounter);
- hadd32(afs_indexCounter, 1);
+ MObtainWriteLock(&afs_xdcache, 602);
+ hset(afs_indexTimes[tdc->index], afs_indexCounter);
+ hadd32(afs_indexCounter, 1);
+ MReleaseWriteLock(&afs_xdcache);
- /* return the data */
- if (vType(avc) == VDIR)
+ /* return the data */
+ if (vType(avc) == VDIR)
*aoffset = abyte;
- else
+ else
*aoffset = AFS_CHUNKOFFSET(abyte);
- *alen = (tdc->f.chunkBytes - *aoffset);
+ *alen = *aoffset + tdc->f.chunkBytes - abyte;
+ ReleaseSharedLock(&tdc->lock);
}
- return tdc;
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks && !slowPass
+ * avc->lock(W) if !setLocks || slowPass
+ */
+
+ /* Fix up the callback and status values in the vcache */
+ doVcacheUpdate = 0;
+ if (setLocks && !slowPass) {
+ /* DCLOCKXXX
+ *
+ * This is our dirty little secret to parallel fetches.
+ * We don't write-lock the vcache while doing the fetch,
+ * but potentially we'll need to update the vcache after
+ * the fetch is done.
+ *
+ * Drop the read lock and try to re-obtain the write
+ * lock. If the vcache still has the same DV, it's
+ * ok to go ahead and install the new data.
+ */
+ afs_hyper_t currentDV, statusDV;
+
+ hset(currentDV, avc->m.DataVersion);
+
+ if (setNewCallback && avc->callback != newCallback)
+ doVcacheUpdate = 1;
+
+ if (tsmall) {
+ hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
+ tsmall->OutStatus.DataVersion);
+
+ if (setVcacheStatus && avc->m.Length != tsmall->OutStatus.Length)
+ doVcacheUpdate = 1;
+ if (setVcacheStatus && !hsame(currentDV, statusDV))
+ doVcacheUpdate = 1;
+ }
+
+ ReleaseReadLock(&avc->lock);
+
+ if (doVcacheUpdate) {
+ ObtainWriteLock(&avc->lock, 615);
+ if (!hsame(avc->m.DataVersion, currentDV)) {
+ /* We lose. Someone will beat us to it. */
+ doVcacheUpdate = 0;
+ ReleaseWriteLock(&avc->lock);
+ }
+ }
+ }
+
+ /* With slow pass, we've already done all the updates */
+ if (slowPass) {
+ ReleaseWriteLock(&avc->lock);
+ }
+
+ /* Check if we need to perform any last-minute fixes with a write-lock */
+ if (!setLocks || doVcacheUpdate) {
+ if (setNewCallback) avc->callback = newCallback;
+ if (tsmall && setVcacheStatus) afs_ProcessFS(avc, &tsmall->OutStatus, areq);
+ if (setLocks) ReleaseWriteLock(&avc->lock);
+ }
+ if (tsmall) osi_FreeLargeSpace(tsmall);
+
+ return tdc;
} /*afs_GetDCache*/
register struct dcache *tdc;
register afs_int32 i, touchedit=0;
+ struct dcache **ents;
+ int entmax, entcount;
AFS_STATCNT(afs_WriteThroughDSlots);
+
+ /*
+ * Because of lock ordering, we can't grab dcache locks while
+ * holding afs_xdcache. So we enter xdcache, get a reference
+ * for every dcache entry, and exit xdcache.
+ */
MObtainWriteLock(&afs_xdcache,283);
+ entmax = afs_cacheFiles;
+ ents = afs_osi_Alloc(entmax * sizeof(struct dcache *));
+ entcount = 0;
for(i = 0; i < afs_cacheFiles; i++) {
tdc = afs_indexTable[i];
- if (tdc && (tdc->flags & DFEntryMod)) {
- tdc->flags &= ~DFEntryMod;
- afs_WriteDCache(tdc, 1);
- touchedit = 1;
+
+ /* Grab tlock in case the existing refcount isn't zero */
+ if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
+ ObtainWriteLock(&tdc->tlock, 623);
+ tdc->refCount++;
+ ReleaseWriteLock(&tdc->tlock);
+
+ ents[entcount++] = tdc;
}
}
+ MReleaseWriteLock(&afs_xdcache);
+
+ /*
+ * Now, for each dcache entry we found, check if it's dirty.
+ * If so, get write-lock, get afs_xdcache, which protects
+ * afs_cacheInodep, and flush it. Don't forget to put back
+ * the refcounts.
+ */
+ for (i = 0; i < entcount; i++) {
+ tdc = ents[i];
+
+ if (tdc->dflags & DFEntryMod) {
+ int wrLock;
+
+ wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
+
+ /* Now that we have the write lock, double-check */
+ if (wrLock && (tdc->dflags & DFEntryMod)) {
+ tdc->dflags &= ~DFEntryMod;
+ MObtainWriteLock(&afs_xdcache, 620);
+ afs_WriteDCache(tdc, 1);
+ MReleaseWriteLock(&afs_xdcache);
+ touchedit = 1;
+ }
+ if (wrLock) ReleaseWriteLock(&tdc->lock);
+ }
+
+ afs_PutDCache(tdc);
+ }
+ afs_osi_Free(ents, entmax * sizeof(struct dcache *));
+
+ MObtainWriteLock(&afs_xdcache, 617);
if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
- /* Touch the file to make sure that the mtime on the file is kept up-to-date
- * to avoid losing cached files on cold starts because their mtime seems old...
+ /* Touch the file to make sure that the mtime on the file is kept
+ * up-to-date to avoid losing cached files on cold starts because
+ * their mtime seems old...
*/
struct afs_fheader theader;
*
* Description:
* Return a pointer to an freshly initialized dcache entry using
- * a memory-based cache.
+ * a memory-based cache. The tlock will be read-locked.
*
* Parameters:
* aslot : Dcache slot to look at.
* tmpdc : Ptr to dcache entry.
*
* Environment:
- * Nothing interesting.
+ * Must be called with afs_xdcache write-locked.
*/
struct dcache *afs_MemGetDSlot(aslot, tmpdc)
register afs_int32 code;
register struct dcache *tdc;
register char *tfile;
+ int existing = 0;
AFS_STATCNT(afs_MemGetDSlot);
if (CheckLock(&afs_xdcache) != -1) osi_Panic("getdslot nolock");
if (tdc) {
QRemove(&tdc->lruq); /* move to queue head */
QAdd(&afs_DLRU, &tdc->lruq);
+ /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
+ ObtainWriteLock(&tdc->tlock, 624);
tdc->refCount++;
+ ConvertWToRLock(&tdc->tlock);
return tdc;
}
if (tmpdc == (struct dcache *)0) {
} else {
tdc = afs_freeDSList;
afs_freeDSList = (struct dcache *) tdc->lruq.next;
+ existing = 1;
}
- tdc->flags = 0; /* up-to-date, not in free q */
+ tdc->dflags = 0; /* up-to-date, not in free q */
+ tdc->mflags = 0;
QAdd(&afs_DLRU, &tdc->lruq);
if (tdc->lruq.prev == &tdc->lruq) osi_Panic("lruq 3");
}
tdc->f.chunk = -1;
hones(tdc->f.versionNo);
tdc->f.inode = aslot;
- tdc->flags |= DFEntryMod;
+ tdc->dflags |= DFEntryMod;
tdc->refCount = 1;
tdc->index = aslot;
afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
+
+ if (existing) {
+ osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
+ osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
+ osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
+ }
+
+ RWLOCK_INIT(&tdc->lock, "dcache lock");
+ RWLOCK_INIT(&tdc->tlock, "dcache tlock");
+ RWLOCK_INIT(&tdc->mflock, "dcache flock");
+ ObtainReadLock(&tdc->tlock);
if (tmpdc == (struct dcache *)0)
afs_indexTable[aslot] = tdc;
*
* Description:
* Return a pointer to an freshly initialized dcache entry using
- * a UFS-based disk cache.
+ * a UFS-based disk cache. The dcache tlock will be read-locked.
*
* Parameters:
* aslot : Dcache slot to look at.
register afs_int32 code;
register struct dcache *tdc;
+ int existing = 0;
AFS_STATCNT(afs_UFSGetDSlot);
if (CheckLock(&afs_xdcache) != -1) osi_Panic("getdslot nolock");
if (aslot < 0 || aslot >= afs_cacheFiles) osi_Panic("getdslot slot");
tdc = afs_indexTable[aslot];
if (tdc) {
-#ifdef AFS_SUN5_ENVX
- mutex_enter(&tdc->lock);
-#endif
QRemove(&tdc->lruq); /* move to queue head */
QAdd(&afs_DLRU, &tdc->lruq);
+ /* Grab tlock in case refCount != 0 */
+ ObtainWriteLock(&tdc->tlock, 625);
tdc->refCount++;
+ ConvertWToRLock(&tdc->tlock);
return tdc;
}
/* otherwise we should read it in from the cache file */
} else {
tdc = afs_freeDSList;
afs_freeDSList = (struct dcache *) tdc->lruq.next;
+ existing = 1;
}
- tdc->flags = 0; /* up-to-date, not in free q */
+ tdc->dflags = 0; /* up-to-date, not in free q */
+ tdc->mflags = 0;
QAdd(&afs_DLRU, &tdc->lruq);
if (tdc->lruq.prev == &tdc->lruq) osi_Panic("lruq 3");
}
tdc->ihint = 0;
}
-#ifdef AFS_SUN5_ENVX
- mutex_enter(&tdc->lock);
-#endif
/*
* Seek to the aslot'th entry and read it in.
*/
tdc->f.fid.Fid.Volume = 0;
tdc->f.chunk = -1;
hones(tdc->f.versionNo);
- tdc->flags |= DFEntryMod;
-#if !defined(AFS_SUN5_ENV) && !defined(AFS_OSF_ENV) && !defined(AFS_SGI64_ENV) && !defined(AFS_LINUX20_ENV)
+ tdc->dflags |= DFEntryMod;
+#if !defined(AFS_SUN5_ENV) && !defined(AFS_OSF_ENV) && !defined(AFS_SGI64_ENV) && !defined(AFS_LINUX20_ENV) && !defined(AFS_DARWIN_ENV) && !defined(AFS_FBSD_ENV)
last_error = getuerror();
#endif
lasterrtime = osi_Time();
tdc->refCount = 1;
tdc->index = aslot;
+ if (existing) {
+ osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
+ osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
+ osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
+ }
+
+ RWLOCK_INIT(&tdc->lock, "dcache lock");
+ RWLOCK_INIT(&tdc->tlock, "dcache tlock");
+ RWLOCK_INIT(&tdc->mflock, "dcache flock");
+ ObtainReadLock(&tdc->tlock);
+
/*
* If we didn't read into a temporary dcache region, update the
* slot pointer table.
* atime : If true, set the modtime on the file to the current time.
*
* Environment:
- * Must be called with the afs_xdcache lock at least read-locked.
+ * Must be called with the afs_xdcache lock at least read-locked,
+ * and dcache entry at least read-locked.
* The reference count is not changed.
*/
{ /*afs_InitCacheFile*/
register afs_int32 code;
-#ifdef AFS_LINUX22_ENV
+#if defined(AFS_LINUX22_ENV)
struct dentry *filevp;
#else
struct vnode *filevp;
MObtainWriteLock(&afs_xdcache,282);
tdc = afs_GetDSlot(index, (struct dcache *)0);
+ ReleaseReadLock(&tdc->tlock);
MReleaseWriteLock(&afs_xdcache);
+
+ ObtainWriteLock(&tdc->lock, 621);
+ MObtainWriteLock(&afs_xdcache, 622);
if (afile) {
code = gop_lookupname(afile,
AFS_UIOSYS,
(struct vnode **) 0,
&filevp);
if (code) {
+ ReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
return code;
}
tdc->f.inode = VTOI(filevp->d_inode)->i_number;
dput(filevp);
#else
-#if defined(AFS_SGI62_ENV) || defined(AFS_HAVE_VXFS)
- tdc->f.inode = VnodeToIno(filevp);
-#else
- tdc->f.inode = VTOI(filevp)->i_number;
-#endif
+ tdc->f.inode = afs_vnodeToInumber(filevp);
#ifdef AFS_DEC_ENV
grele(filevp);
#else
osi_UFSClose(tfile);
tdc->f.states &= ~DWriting;
- tdc->flags &= ~DFEntryMod;
+ tdc->dflags &= ~DFEntryMod;
/* don't set f.modTime; we're just cleaning up */
afs_WriteDCache(tdc, 0);
+ ReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
afs_stats_cmperf.cacheNumEntries++;
return 0;
/* Allocate and zero the pointer array to the dcache entries */
afs_indexTable = (struct dcache **)
afs_osi_Alloc(sizeof(struct dcache *) * afiles);
- bzero((char *)afs_indexTable, sizeof(struct dcache *) * afiles);
+ memset((char *)afs_indexTable, 0, sizeof(struct dcache *) * afiles);
afs_indexTimes = (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
- bzero((char *)afs_indexTimes, afiles * sizeof(afs_hyper_t));
+ memset((char *)afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
afs_indexUnique = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
- bzero((char *)afs_indexUnique, afiles * sizeof(afs_uint32));
+ memset((char *)afs_indexUnique, 0, afiles * sizeof(afs_uint32));
afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
- bzero((char *)afs_indexFlags, afiles * sizeof(char));
+ memset((char *)afs_indexFlags, 0, afiles * sizeof(char));
/* Allocate and thread the struct dcache entries themselves */
tdp = afs_Initial_freeDSList =
(struct dcache *) afs_osi_Alloc(aDentries * sizeof(struct dcache));
- bzero((char *)tdp, aDentries * sizeof(struct dcache));
+ memset((char *)tdp, 0, aDentries * sizeof(struct dcache));
#ifdef AFS_AIX32_ENV
pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles);/* XXX */
pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
afs_dcentries = aDentries;
afs_blocksUsed = 0;
QInit(&afs_DLRU);
-
-
}
/*