/*
* Copyright 2000, International Business Machines Corporation and others.
- *$All Rights Reserved.
- *
+ * All Rights Reserved.
+ *
* This software has been released under the terms of the IBM Public
* License. For details, see the LICENSE file in the top-level source
* directory or online at http://www.openafs.org/dl/license10.html
#include "afs/afs_cbqueue.h"
#include "afs/afs_osidnlc.h"
+#include <opr/ffs.h>
+
/* Forward declarations. */
static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
-static void afs_FreeDiscardedDCache(void);
+static int afs_FreeDiscardedDCache(void);
static void afs_DiscardDCache(struct dcache *);
static void afs_FreeDCache(struct dcache *);
/* For split 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 = AFS_MIN_FRAGSIZE; /*!< Underlying Filesystem minimum unit
+afs_int32 afs_fsfragsize = AFS_MIN_FRAGSIZE; /*!< Underlying Filesystem minimum unit
*of disk allocation usually 1K
*this value is (truefrag -1 ) to
*save a bunch of subtracts... */
int dcacheDisabled = 0;
struct afs_cacheOps afs_UfsCacheOps = {
-#if defined(AFS_SGI_ENV) && !defined(__c99)
+#ifndef HAVE_STRUCT_LABEL_SUPPORT
osi_UFSOpen,
osi_UFSTruncate,
afs_osi_Read,
afs_osi_Write,
osi_UFSClose,
- afs_UFSRead,
- afs_UFSWrite,
+ afs_UFSReadUIO,
+ afs_UFSWriteUIO,
afs_UFSGetDSlot,
afs_UFSGetVolSlot,
afs_UFSHandleLink,
.fread = afs_osi_Read,
.fwrite = afs_osi_Write,
.close = osi_UFSClose,
- .vread = afs_UFSRead,
- .vwrite = afs_UFSWrite,
+ .vreadUIO = afs_UFSReadUIO,
+ .vwriteUIO = afs_UFSWriteUIO,
.GetDSlot = afs_UFSGetDSlot,
.GetVolSlot = afs_UFSGetVolSlot,
.HandleLink = afs_UFSHandleLink,
};
struct afs_cacheOps afs_MemCacheOps = {
-#if (defined(AFS_SGI_ENV) && !defined(__c99))
+#ifndef HAVE_STRUCT_LABEL_SUPPORT
afs_MemCacheOpen,
afs_MemCacheTruncate,
afs_MemReadBlk,
afs_MemWriteBlk,
afs_MemCacheClose,
- afs_MemRead,
- afs_MemWrite,
+ afs_MemReadUIO,
+ afs_MemWriteUIO,
afs_MemGetDSlot,
afs_MemGetVolSlot,
afs_MemHandleLink,
.fread = afs_MemReadBlk,
.fwrite = afs_MemWriteBlk,
.close = afs_MemCacheClose,
- .vread = afs_MemRead,
- .vwrite = afs_MemWrite,
+ .vreadUIO = afs_MemReadUIO,
+ .vwriteUIO = afs_MemWriteUIO,
.GetDSlot = afs_MemGetDSlot,
.GetVolSlot = afs_MemGetVolSlot,
.HandleLink = afs_MemHandleLink,
int cacheDiskType; /*Type of backing disk for cache */
struct afs_cacheOps *afs_cacheType;
+
+/*
+ * The PFlush algorithm makes use of the fact that Fid.Unique is not used in
+ * below hash algorithms. Change it if need be so that flushing algorithm
+ * doesn't move things from one hash chain to another.
+ */
+/*Vnode, Chunk -> Hash table index */
+int DCHash(struct VenusFid *fid, afs_int32 chunk)
+{
+ afs_uint32 buf[3];
+
+ buf[0] = fid->Fid.Volume;
+ buf[1] = fid->Fid.Vnode;
+ buf[2] = chunk;
+ return opr_jhash(buf, 3, 0) & (afs_dhashsize - 1);
+}
+/*Vnode -> Other hash table index */
+int DVHash(struct VenusFid *fid)
+{
+ return opr_jhash_int2(fid->Fid.Volume, fid->Fid.Vnode, 0) &
+ (afs_dhashsize - 1);
+}
+
/*!
* Where is this vcache's entry associated dcache located/
* \param avc The vcache entry.
* 2 : RO
*/
static afs_int32
-afs_DCGetBucket(struct vcache *avc)
+afs_DCGetBucket(struct vcache *avc)
{
- if (!splitdcache)
+ if (!splitdcache)
return 1;
-
+
/* This should be replaced with some sort of user configurable function */
if (avc->f.states & CRO) {
return 2;
* \param newSize The new size to be adjusted to.
*
*/
-static void
+static void
afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
{
afs_int32 adjustSize = newSize - oldSize;
- if (!splitdcache)
+ if (!splitdcache)
return;
- switch (adc->bucket)
+ switch (adc->bucket)
{
case 0:
afs_blocksUsed_0 += adjustSize;
/*!
* Move a dcache from one bucket to another.
- *
+ *
* \param adc Operate on this dcache.
* \param size Size in bucket (?).
* \param newBucket Destination bucket.
*
*/
-static void
+static void
afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
{
- if (!splitdcache)
+ if (!splitdcache)
return;
- /* Substract size from old bucket. */
- switch (adc->bucket)
+ /* Substract size from old bucket. */
+ switch (adc->bucket)
{
case 0:
afs_blocksUsed_0 -= size;
/* Set new bucket and increase destination bucket size. */
adc->bucket = newBucket;
- switch (adc->bucket)
+ switch (adc->bucket)
{
case 0:
afs_blocksUsed_0 += size;
afs_blocksUsed_2 += size;
break;
}
-
+
return;
}
/*!
* Init split caches size.
*/
-static void
-afs_DCSizeInit(void)
+static void
+afs_DCSizeInit(void)
{
afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
}
* \param bucket
*/
static afs_int32
-afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
+afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
{
- if (!splitdcache)
+ if (!splitdcache)
return 0;
afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
/* Short cut: if we don't know about it, try to kill it */
- if (phase < 2 && afs_blocksUsed_0)
+ if (phase < 2 && afs_blocksUsed_0)
return 0;
-
- if (afs_pct1 > afs_tpct1)
+
+ if (afs_pct1 > afs_tpct1)
return 1;
if (afs_pct2 > afs_tpct2)
return 2;
*/
void
-afs_StoreWarn(register afs_int32 acode, afs_int32 avolume,
- register afs_int32 aflags)
+afs_StoreWarn(afs_int32 acode, afs_int32 avolume,
+ afs_int32 aflags)
{
static char problem_fmt[] =
"afs: failed to store file in volume %d (%s)\n";
* struct so we need only export one symbol for AIX.
*/
static struct CTD_stats {
- osi_timeval_t CTD_beforeSleep;
- osi_timeval_t CTD_afterSleep;
- osi_timeval_t CTD_sleepTime;
- osi_timeval_t CTD_runTime;
+ osi_timeval32_t CTD_beforeSleep;
+ osi_timeval32_t CTD_afterSleep;
+ osi_timeval32_t CTD_sleepTime;
+ osi_timeval32_t CTD_runTime;
int CTD_nSleeps;
} CTD_stats;
u_int afs_min_cache = 0;
/*!
+ * If there are waiters for the cache to drain, wake them if
+ * the number of free or discarded cache blocks reaches the
+ * CM_CACHESIZEDDRAINEDPCT limit.
+ *
+ * \note Environment:
+ * This routine must be called with the afs_xdcache lock held
+ * (in write mode).
+ */
+static void
+afs_WakeCacheWaitersIfDrained(void)
+{
+ if (afs_WaitForCacheDrain) {
+ if ((afs_blocksUsed - afs_blocksDiscarded) <=
+ PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
+ afs_WaitForCacheDrain = 0;
+ afs_osi_Wakeup(&afs_WaitForCacheDrain);
+ }
+ }
+}
+
+/*!
* Keeps the cache clean and free by truncating uneeded files, when used.
- * \param
- * \return
+ * \param
+ * \return
*/
void
afs_CacheTruncateDaemon(void)
{
- osi_timeval_t CTD_tmpTime;
+ osi_timeval32_t CTD_tmpTime;
u_int counter;
u_int cb_lowat;
u_int dc_hiwat =
afs_min_cache =
(((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
- osi_GetuTime(&CTD_stats.CTD_afterSleep);
+ osi_GetTime(&CTD_stats.CTD_afterSleep);
afs_TruncateDaemonRunning = 1;
while (1) {
cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
ObtainWriteLock(&afs_xdcache, 266);
- if (afs_CacheTooFull) {
+ if (afs_CacheTooFull || afs_WaitForCacheDrain) {
int space_needed, slots_needed;
/* if we get woken up, we should try to clean something out */
for (counter = 0; counter < 10; counter++) {
space_needed =
afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
+ if (space_needed < 0)
+ space_needed = 0;
slots_needed =
dc_hiwat - afs_freeDCCount - afs_discardDCCount;
- afs_GetDownD(slots_needed, &space_needed, 0);
+ if (slots_needed < 0)
+ slots_needed = 0;
+ if (slots_needed || space_needed)
+ afs_GetDownD(slots_needed, &space_needed, 0);
if ((space_needed <= 0) && (slots_needed <= 0)) {
break;
}
if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
break;
}
- if (!afs_CacheIsTooFull())
+ if (!afs_CacheIsTooFull()) {
afs_CacheTooFull = 0;
+ afs_WakeCacheWaitersIfDrained();
+ }
} /* end of cache cleanup */
ReleaseWriteLock(&afs_xdcache);
*/
while (afs_blocksDiscarded && !afs_WaitForCacheDrain
&& (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
- afs_FreeDiscardedDCache();
+ int code = afs_FreeDiscardedDCache();
+ if (code) {
+ /* If we can't free any discarded dcache entries, that's okay.
+ * We're just doing this in the background; if someone needs
+ * discarded entries freed, they will try it themselves and/or
+ * signal us that the cache is too full. In any case, we'll
+ * try doing this again the next time we run through the loop.
+ */
+ break;
+ }
}
/* See if we need to continue to run. Someone may have
&& (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
/* Collect statistics on truncate daemon. */
CTD_stats.CTD_nSleeps++;
- osi_GetuTime(&CTD_stats.CTD_beforeSleep);
+ osi_GetTime(&CTD_stats.CTD_beforeSleep);
afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
CTD_stats.CTD_beforeSleep);
afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
afs_TruncateDaemonRunning = 1;
- osi_GetuTime(&CTD_stats.CTD_afterSleep);
+ osi_GetTime(&CTD_stats.CTD_afterSleep);
afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
CTD_stats.CTD_afterSleep);
afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
*/
void
-afs_AdjustSize(register struct dcache *adc, register afs_int32 newSize)
+afs_AdjustSize(struct dcache *adc, afs_int32 newSize)
{
- register afs_int32 oldSize;
+ afs_int32 oldSize;
AFS_STATCNT(afs_AdjustSize);
+ if (newSize > afs_OtherCSize && !(adc->f.fid.Fid.Vnode & 1)) {
+ /* No non-dir cache files should be larger than the chunk size.
+ * (Directory blobs are fetched in a single chunk file, so directories
+ * can be larger.) If someone is requesting that a chunk is larger than
+ * the chunk size, something strange is happening. Log a message about
+ * it, to give a hint to subsequent strange behavior, if any occurs. */
+ static int warned;
+ if (!warned) {
+ warned = 1;
+ afs_warn("afs: Warning: dcache %d is very large (%d > %d). This "
+ "should not happen, but trying to continue regardless. If "
+ "AFS starts hanging or behaving strangely, this might be "
+ "why.\n",
+ adc->index, newSize, afs_OtherCSize);
+ }
+ }
+
adc->dflags |= DFEntryMod;
oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
adc->f.chunkBytes = newSize;
* 1. only grab up to anumber victims if aneedSpace <= 0, not
* the whole set of MAXATONCE.
* 2. dynamically choose MAXATONCE to reflect severity of
- * demand: something like (*aneedSpace >> (logChunk - 9))
+ * demand: something like (*aneedSpace >> (logChunk - 9))
*
* \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
* indicates that the cache is not properly configured/tuned or
afs_int32 i, j;
afs_hyper_t vtime;
int skip, phase;
- register struct vcache *tvc;
+ struct vcache *tvc;
afs_uint32 victims[MAXATONCE];
struct dcache *victimDCs[MAXATONCE];
afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
afs_uint32 maxVictimPtr; /* where it is */
int discard;
int curbucket;
-#if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
- int vfslocked;
-#endif
-
-#if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
- vfslocked = VFS_LOCK_GIANT(afs_globalVFS);
-#endif
AFS_STATCNT(afs_GetDownD);
if (CheckLock(&afs_xdcache) != -1)
osi_Panic("getdownd nolock");
/* decrement anumber first for all dudes in free list */
- /* SHOULD always decrement anumber first, even if aneedSpace >0,
+ /* SHOULD always decrement anumber first, even if aneedSpace >0,
* because we should try to free space even if anumber <=0 */
if (!aneedSpace || *aneedSpace <= 0) {
anumber -= afs_freeDCCount;
if (anumber <= 0) {
-#if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
- VFS_UNLOCK_GIANT(vfslocked);
-#endif
return; /* enough already free */
}
}
/* rewrite so phases include a better eligiblity for gc test*/
/*
* The phase variable manages reclaims. Set to 0, the first pass,
- * we don't reclaim active entries, or other than target bucket.
+ * we don't reclaim active entries, or other than target bucket.
* Set to 1, we reclaim even active ones in target bucket.
* Set to 2, we reclaim any inactive one.
- * Set to 3, we reclaim even active ones.
+ * Set to 3, we reclaim even active ones. On Solaris, we also reclaim
+ * entries whose corresponding vcache has a nonempty multiPage list, when
+ * possible.
*/
if (splitdcache) {
phase = 0;
* during the truncate operation.
*/
for (i = 0; i < victimPtr; i++) {
- tdc = afs_GetDSlot(victims[i], 0);
+ tdc = afs_GetValidDSlot(victims[i]);
/* We got tdc->tlock(R) here */
- if (tdc->refCount == 1)
+ if (tdc && tdc->refCount == 1)
victimDCs[i] = tdc;
else
victimDCs[i] = 0;
- ReleaseReadLock(&tdc->tlock);
- if (!victimDCs[i])
- afs_PutDCache(tdc);
+ if (tdc) {
+ ReleaseReadLock(&tdc->tlock);
+ if (!victimDCs[i])
+ afs_PutDCache(tdc);
+ }
}
for (i = 0; i < victimPtr; i++) {
/* q is first elt in dcache entry */
* have to verify, before proceeding, that there are no other
* references to this dcache entry, even now. Note that we
* compare with 1, since we bumped it above when we called
- * afs_GetDSlot to preserve the entry's identity.
+ * afs_GetValidDSlot to preserve the entry's identity.
*/
if (tdc && tdc->refCount == 1) {
unsigned char chunkFlags;
ReleaseWriteLock(&afs_xdcache);
ObtainWriteLock(&tvc->vlock, 543);
- if (tvc->multiPage) {
- skip = 1;
- goto endmultipage;
+ if (!QEmpty(&tvc->multiPage)) {
+ if (phase < 3 || osi_VM_MultiPageConflict(tvc, tdc)) {
+ skip = 1;
+ goto endmultipage;
+ }
}
/* block locking pages */
tvc->vstates |= VPageCleaning;
j = 1; /* we reclaimed at least one victim */
}
}
- afs_PutDCache(tdc);
+ if (tdc)
+ afs_PutDCache(tdc);
} /* end of for victims loop */
if (phase < 5) {
}
} /* big while loop */
-#if defined(AFS_FBSD80_ENV) && !defined(UKERNEL)
- VFS_UNLOCK_GIANT(vfslocked);
-#endif
-
return;
} /*afs_GetDownD */
* (in write mode).
*/
void
-afs_FlushDCache(register struct dcache *adc)
+afs_FlushDCache(struct dcache *adc)
{
AFS_STATCNT(afs_FlushDCache);
/*
} else {
afs_FreeDCache(adc);
}
-
- if (afs_WaitForCacheDrain) {
- if (afs_blocksUsed <=
- PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
- afs_WaitForCacheDrain = 0;
- afs_osi_Wakeup(&afs_WaitForCacheDrain);
- }
- }
} /*afs_FlushDCache */
* \note Environment: called with afs_xdcache lock write-locked.
*/
static void
-afs_FreeDCache(register struct dcache *adc)
+afs_FreeDCache(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
afs_indexFlags[adc->index] |= IFFree;
adc->dflags |= DFEntryMod;
- if (afs_WaitForCacheDrain) {
- if ((afs_blocksUsed - afs_blocksDiscarded) <=
- PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
- afs_WaitForCacheDrain = 0;
- afs_osi_Wakeup(&afs_WaitForCacheDrain);
- }
- }
+ afs_WakeCacheWaitersIfDrained();
} /* afs_FreeDCache */
/*!
*/
static void
-afs_DiscardDCache(register struct dcache *adc)
+afs_DiscardDCache(struct dcache *adc)
{
- register afs_int32 size;
+ afs_int32 size;
AFS_STATCNT(afs_DiscardDCache);
adc->dflags |= DFEntryMod;
afs_indexFlags[adc->index] |= IFDiscarded;
- if (afs_WaitForCacheDrain) {
- if ((afs_blocksUsed - afs_blocksDiscarded) <=
- PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
- afs_WaitForCacheDrain = 0;
- afs_osi_Wakeup(&afs_WaitForCacheDrain);
- }
+ afs_WakeCacheWaitersIfDrained();
+} /*afs_DiscardDCache */
+
+/**
+ * Get a dcache entry from the discard or free list
+ *
+ * @param[out] adc On success, a dcache from the given list. Otherwise, NULL.
+ * @param[in] indexp A pointer to the head of the dcache free list or discard
+ * list (afs_freeDCList, or afs_discardDCList)
+ *
+ * @return 0 on success. If there are no dcache slots available, return ENOSPC.
+ * If we encountered an error in disk i/o while trying to find a
+ * dcache, return EIO.
+ *
+ * @pre afs_xdcache is write-locked
+ */
+static int
+afs_GetDSlotFromList(struct dcache **adc, afs_int32 *indexp)
+{
+ struct dcache *tdc;
+
+ *adc = NULL;
+
+ if (*indexp == NULLIDX) {
+ return ENOSPC;
}
-} /*afs_DiscardDCache */
+ tdc = afs_GetUnusedDSlot(*indexp);
+ if (tdc == NULL) {
+ return EIO;
+ }
+
+ osi_Assert(tdc->refCount == 1);
+ ReleaseReadLock(&tdc->tlock);
+ *indexp = afs_dvnextTbl[tdc->index];
+ afs_dvnextTbl[tdc->index] = NULLIDX;
+
+ *adc = tdc;
+ return 0;
+}
/*!
* Free the next element on the list of discarded cache elements.
+ *
+ * Returns -1 if we encountered an error preventing us from freeing a
+ * discarded dcache, or 0 on success.
*/
-static void
+static int
afs_FreeDiscardedDCache(void)
{
- register struct dcache *tdc;
- register struct osi_file *tfile;
- register afs_int32 size;
+ struct dcache *tdc;
+ struct osi_file *tfile;
+ afs_int32 size;
AFS_STATCNT(afs_FreeDiscardedDCache);
ObtainWriteLock(&afs_xdcache, 510);
if (!afs_blocksDiscarded) {
ReleaseWriteLock(&afs_xdcache);
- return;
+ return 0;
}
/*
* Get an entry from the list of discarded cache elements
*/
- tdc = afs_GetDSlot(afs_discardDCList, 0);
- osi_Assert(tdc->refCount == 1);
- ReleaseReadLock(&tdc->tlock);
+ (void)afs_GetDSlotFromList(&tdc, &afs_discardDCList);
+ if (!tdc) {
+ ReleaseWriteLock(&afs_xdcache);
+ return -1;
+ }
- 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;
* Truncate the element to reclaim its space
*/
tfile = afs_CFileOpen(&tdc->f.inode);
+ osi_Assert(tfile);
afs_CFileTruncate(tfile, 0);
afs_CFileClose(tfile);
afs_AdjustSize(tdc, 0);
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
ReleaseWriteLock(&afs_xdcache);
+
+ return 0;
}
/*!
while (afs_blocksDiscarded
&& (afs_blocksUsed >
PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
- afs_FreeDiscardedDCache();
+ int code = afs_FreeDiscardedDCache();
+ if (code) {
+ /* Callers depend on us to get the afs_blocksDiscarded count down.
+ * If we cannot do that, the callers can spin by calling us over
+ * and over. Panic for now until we can figure out something
+ * better. */
+ osi_Panic("Error freeing discarded dcache");
+ }
}
return 0;
}
if (tdc->refCount == 0) {
if ((ix = tdc->index) == NULLIDX)
osi_Panic("getdowndslot");
- /* pull the entry out of the lruq and put it on the free list */
- QRemove(&tdc->lruq);
/* write-through if modified */
if (tdc->dflags & DFEntryMod) {
AFS_GLOCK();
}
#else
+ int code;
+
+ code = afs_WriteDCache(tdc, 1);
+ if (code) {
+ /*
+ * We couldn't flush it at this time; return early because
+ * if afs_WriteDCache() failed once it is likely to
+ * continue failing for subsequent dcaches.
+ */
+ return;
+ }
tdc->dflags &= ~DFEntryMod;
- afs_WriteDCache(tdc, 1);
#endif
}
- /* finally put the entry in the free list */
+ /* pull the entry out of the lruq and put it on the free list */
+ QRemove(&tdc->lruq);
afs_indexTable[ix] = NULL;
afs_indexFlags[ix] &= ~IFEverUsed;
tdc->index = NULLIDX;
* Nothing interesting.
*/
int
-afs_PutDCache(register struct dcache *adc)
+afs_PutDCache(struct dcache *adc)
{
AFS_STATCNT(afs_PutDCache);
ObtainWriteLock(&adc->tlock, 276);
* Both pvnLock and lock are write held.
*/
void
-afs_TryToSmush(register struct vcache *avc, afs_ucred_t *acred, int sync)
+afs_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
{
- register struct dcache *tdc;
- register int index;
- register int i;
+ struct dcache *tdc;
+ int index;
+ int i;
AFS_STATCNT(afs_TryToSmush);
afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
i = afs_dvnextTbl[index]; /* next pointer this hash table */
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
int releaseTlock = 1;
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ /* afs_TryToSmush is best-effort; we may not actually discard
+ * everything, so failure to discard dcaches due to an i/o
+ * error is okay. */
+ break;
+ }
if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
if (sync) {
if ((afs_indexFlags[index] & IFDataMod) == 0
#endif
ReleaseWriteLock(&afs_xdcache);
/*
- * It's treated like a callback so that when we do lookups we'll
+ * It's treated like a callback so that when we do lookups we'll
* invalidate the unique bit if any
* trytoSmush occured during the lookup call
*/
* Description
* Given the cached info for a file, return the number of chunks that
* are not available from the dcache.
- *
+ *
* Parameters:
* avc: Pointer to the (held) vcache entry to look in.
- *
+ *
* Returns:
* The number of chunks which are not currently cached.
- *
+ *
* Environment:
* The vcache entry is held upon entry.
*/
*/
if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
totalChunks = 1;
-
+
/*
printf("Should have %d chunks for %u bytes\n",
totalChunks, (totalLength + 1));
for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
i = afs_dvnextTbl[index];
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ break;
+ }
if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
- totalChunks--;
+ totalChunks--;
}
ReleaseReadLock(&tdc->tlock);
afs_PutDCache(tdc);
*/
struct dcache *
-afs_FindDCache(register struct vcache *avc, afs_size_t abyte)
+afs_FindDCache(struct vcache *avc, afs_size_t abyte)
{
afs_int32 chunk;
- register afs_int32 i, index;
- register struct dcache *tdc = NULL;
+ afs_int32 i, index;
+ struct dcache *tdc = NULL;
AFS_STATCNT(afs_FindDCache);
chunk = AFS_CHUNK(abyte);
*/
i = DCHash(&avc->f.fid, chunk);
ObtainWriteLock(&afs_xdcache, 278);
- for (index = afs_dchashTbl[i]; index != NULLIDX;) {
+ for (index = afs_dchashTbl[i]; index != NULLIDX; index = afs_dcnextTbl[index]) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ /* afs_FindDCache is best-effort; we may not find the given
+ * file/offset, so if we cannot find the given dcache due to
+ * i/o errors, that is okay. */
+ index = NULLIDX;
+ break;
+ }
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
break; /* leaving refCount high for caller */
}
afs_PutDCache(tdc);
}
- index = afs_dcnextTbl[index];
}
if (index != NULLIDX) {
hset(afs_indexTimes[tdc->index], afs_indexCounter);
hadd32(afs_indexCounter, 1);
ReleaseWriteLock(&afs_xdcache);
return tdc;
- }
+ }
ReleaseWriteLock(&afs_xdcache);
return NULL;
} /*afs_FindDCache */
+/* only call these from afs_AllocDCache() */
+static int
+afs_AllocFreeDSlot(struct dcache **adc)
+{
+ int code;
+ struct dcache *tdc;
+
+ code = afs_GetDSlotFromList(&tdc, &afs_freeDCList);
+ if (code) {
+ return code;
+ }
+ afs_indexFlags[tdc->index] &= ~IFFree;
+ ObtainWriteLock(&tdc->lock, 604);
+ afs_freeDCCount--;
+
+ *adc = tdc;
+ return 0;
+}
+static int
+afs_AllocDiscardDSlot(struct dcache **adc, afs_int32 lock)
+{
+ int code;
+ struct dcache *tdc;
+ afs_uint32 size = 0;
+ struct osi_file *file;
+
+ code = afs_GetDSlotFromList(&tdc, &afs_discardDCList);
+ if (code) {
+ return code;
+ }
+ afs_indexFlags[tdc->index] &= ~IFDiscarded;
+ ObtainWriteLock(&tdc->lock, 605);
+ afs_discardDCCount--;
+ size =
+ ((tdc->f.chunkBytes +
+ afs_fsfragsize) ^ afs_fsfragsize) >> 10;
+ tdc->f.states &= ~(DRO|DBackup|DRW);
+ afs_DCMoveBucket(tdc, size, 0);
+ afs_blocksDiscarded -= size;
+ afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
+ if ((lock & 2)) {
+ /* Truncate the chunk so zeroes get filled properly */
+ file = afs_CFileOpen(&tdc->f.inode);
+ osi_Assert(file);
+ afs_CFileTruncate(file, 0);
+ afs_CFileClose(file);
+ afs_AdjustSize(tdc, 0);
+ }
+
+ *adc = tdc;
+ return 0;
+}
/*!
* Get a fresh dcache from the free or discarded list.
*
+ * \param adc Set to the new dcache on success, and NULL on error.
* \param avc Who's dcache is this going to be?
* \param chunk The position where it will be placed in.
* \param lock How are locks held.
* - avc (R if (lock & 1) set and W otherwise)
* \note It write locks the new dcache. The caller must unlock it.
*
- * \return The new dcache.
+ * \return If we're out of dslots, ENOSPC. If we encountered disk errors, EIO.
+ * On success, return 0.
*/
-struct dcache *
-afs_AllocDCache(struct vcache *avc, afs_int32 chunk, afs_int32 lock,
- struct VenusFid *ashFid)
+static int
+afs_AllocDCache(struct dcache **adc, struct vcache *avc, afs_int32 chunk,
+ afs_int32 lock, struct VenusFid *ashFid)
{
+ int code;
struct dcache *tdc = NULL;
- afs_uint32 size = 0;
- struct osi_file *file;
- if (afs_discardDCList == NULLIDX
- || ((lock & 2) && afs_freeDCList != NULLIDX)) {
+ *adc = NULL;
- 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--;
+ /* if (lock & 2), prefer 'free' dcaches; otherwise, prefer 'discard'
+ * dcaches. In either case, try both if our first choice doesn't work due
+ * to ENOSPC. */
+ if ((lock & 2)) {
+ code = afs_AllocFreeDSlot(&tdc);
+ if (code == ENOSPC) {
+ code = afs_AllocDiscardDSlot(&tdc, lock);
+ }
} 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;
- tdc->f.states &= ~(DRO|DBackup|DRW);
- afs_DCMoveBucket(tdc, size, 0);
- afs_blocksDiscarded -= size;
- afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
- if (lock & 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);
+ code = afs_AllocDiscardDSlot(&tdc, lock);
+ if (code == ENOSPC) {
+ code = afs_AllocFreeDSlot(&tdc);
}
}
+ if (code) {
+ return code;
+ }
/*
* Locks held:
if (tdc->lruq.prev == &tdc->lruq)
osi_Panic("lruq 1");
- return tdc;
+ *adc = tdc;
+ return 0;
+}
+
+static int
+IsDCacheSizeOK(struct dcache *adc, struct vcache *avc, afs_int32 chunk_bytes,
+ afs_size_t file_length, afs_uint32 versionNo, int from_net)
+{
+ afs_size_t expected_bytes;
+ afs_size_t chunk_start = AFS_CHUNKTOBASE(adc->f.chunk);
+
+ if (vType(avc) == VDIR) {
+ /*
+ * Directory blobs may be constructed locally (see afs_LocalHero), and
+ * the size of the blob may differ slightly compared to what's on the
+ * fileserver. So, skip size checks for directories.
+ */
+ return 1;
+ }
+
+ if ((avc->f.states & CDirty)) {
+ /*
+ * Our vcache may have writes that are local to our cache, but not yet
+ * written to the fileserver. In such a situation, we may have dcaches
+ * for that file that are "short". For example:
+ *
+ * Say we have a file that is 0 bytes long. A process opens that file,
+ * and writes some data to offset 5M (keeping the file open). Another
+ * process comes along and reads data from offset 1M. We'll try to
+ * fetch data at offset 1M, and the fileserver will respond with 0
+ * bytes, since our locally-written data hasn't been written to the
+ * fileserver yet (on the fileserver, the file is still 0-bytes long).
+ * So our dcache at offset 1M will have 0 bytes.
+ *
+ * So if CDirty is set, don't do any size/length checks at all, since
+ * we have no idea if the avc length is valid.
+ */
+ return 1;
+ }
+
+ if (!from_net && (adc->f.states & DRW)) {
+ /*
+ * The dcache data we're looking at is from our local cache (not from a
+ * fileserver), and it's for data in an RW volume. For cached RW data,
+ * there are some edge cases that can cause the below length checks to
+ * trigger false positives.
+ *
+ * For example: if the local client writes 4 bytes to a new file at
+ * offset 0, and then 4 bytes at offset 0x400000, the file will be
+ * 0x400004 bytes long, but the first dcache chunk will only contain 4
+ * bytes. If such a file is fetched from a fileserver, the first chunk
+ * will have a full chunk of data (most of it zeroes), but on the
+ * client that did the write, the sparse data will not appear in the
+ * dcache.
+ *
+ * Such false positives should only be possible with RW data, since
+ * non-RW data is never generated locally. So to avoid the false
+ * positives, assume the dcache length is OK for RW data if the dcache
+ * came from our local cache (and not directly from a fileserver).
+ */
+ return 1;
+ }
+
+ if (file_length < chunk_start) {
+ expected_bytes = 0;
+
+ } else {
+ expected_bytes = file_length - chunk_start;
+
+ if (vType(avc) != VDIR && expected_bytes > AFS_CHUNKTOSIZE(adc->f.chunk)) {
+ /* A non-dir chunk cannot have more bytes than the chunksize. */
+ expected_bytes = AFS_CHUNKTOSIZE(adc->f.chunk);
+ }
+ }
+
+ if (chunk_bytes != expected_bytes) {
+ static const afs_uint32 one_hour = 60 * 60;
+ static afs_uint32 last_warn;
+ afs_uint32 now = osi_Time();
+
+ if (now < last_warn) {
+ /* clock went backwards */
+ last_warn = now;
+ }
+
+ if (now - last_warn > one_hour) {
+ unsigned int mtime = adc->f.modTime;
+
+ last_warn = now;
+
+ if (from_net) {
+ /*
+ * The dcache we're looking at didn't come from the cache, but is
+ * being populated from the net. Don't print out its mtime in that
+ * case; that would be misleading since that's the mtime from the
+ * last time this dcache slot was written to.
+ */
+ mtime = 0;
+ }
+
+ afs_warn("afs: Detected corrupt dcache for file %d.%u.%u.%u: chunk %d "
+ "(offset %lu) has %d bytes, but it should have %lu bytes\n",
+ adc->f.fid.Cell,
+ adc->f.fid.Fid.Volume,
+ adc->f.fid.Fid.Vnode,
+ adc->f.fid.Fid.Unique,
+ adc->f.chunk,
+ (unsigned long)chunk_start,
+ chunk_bytes,
+ (unsigned long)expected_bytes);
+ afs_warn("afs: (dcache %p, file length %lu, DV %u, dcache mtime %u, "
+ "index %d, dflags 0x%x, mflags 0x%x, states 0x%x, vcache "
+ "states 0x%x)\n",
+ adc,
+ (unsigned long)file_length,
+ versionNo,
+ mtime,
+ adc->index,
+ (unsigned)adc->dflags,
+ (unsigned)adc->mflags,
+ (unsigned)adc->f.states,
+ avc->f.states);
+ afs_warn("afs: Ignoring the dcache for now, but this may indicate "
+ "corruption in the AFS cache, or a bug.\n");
+ }
+ return 0;
+ }
+ return 1;
+}
+
+/*!
+ * Check if a dcache is "fresh". That is, if the dcache's DV matches the DV of
+ * the vcache for that file, and the dcache looks "sane" (its length makes
+ * sense, when considering the length of the given avc).
+ *
+ * \param adc The dcache to check
+ * \param avc The vcache for adc
+ *
+ * \return 1 if the dcache is "fresh". 0 otherwise.
+ */
+int
+afs_IsDCacheFresh(struct dcache *adc, struct vcache *avc)
+{
+ if (!hsame(adc->f.versionNo, avc->f.m.DataVersion)) {
+ return 0;
+ }
+
+ /*
+ * If we've reached here, the DV in adc matches the DV of our avc. Check if
+ * the number of bytes in adc agrees with the avc file length, as a sanity
+ * check. If they don't match, we'll pretend the DVs don't match, so the
+ * bad dcache data will not be used, and we'll probably re-fetch the chunk
+ * data, replacing the bad chunk.
+ */
+
+ if (!IsDCacheSizeOK(adc, avc, adc->f.chunkBytes, avc->f.m.Length,
+ hgetlo(adc->f.versionNo), 0)) {
+ return 0;
+ }
+
+ return 1;
}
/*
updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
{
if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
- if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
+ if (afs_IsDCacheFresh(d, v) && v->callback)
v->dchint = d;
if (lockVc)
ReleaseWriteLock(&v->lock);
/* avc - Write-locked unless aflags & 1 */
struct dcache *
-afs_GetDCache(register struct vcache *avc, afs_size_t abyte,
- register struct vrequest *areq, afs_size_t * aoffset,
+afs_GetDCache(struct vcache *avc, afs_size_t abyte,
+ struct vrequest *areq, afs_size_t * aoffset,
afs_size_t * alen, int aflags)
{
- register afs_int32 i, code, shortcut;
+ afs_int32 i, code, shortcut;
#if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
- register afs_int32 adjustsize = 0;
+ afs_int32 adjustsize = 0;
#endif
int setLocks;
afs_int32 index;
afs_int32 us;
afs_int32 chunk;
- afs_size_t maxGoodLength; /* amount of good data at server */
afs_size_t Position = 0;
afs_int32 size, tlen; /* size of segment to transfer */
struct afs_FetchOutput *tsmall = 0;
- register struct dcache *tdc;
- register struct osi_file *file;
- register struct afs_conn *tc;
+ struct dcache *tdc;
+ struct osi_file *file;
+ struct afs_conn *tc;
int downDCount = 0;
struct server *newCallback = NULL;
char setNewCallback;
int doAdjustSize = 0;
int doReallyAdjustSize = 0;
int overWriteWholeChunk = 0;
+ struct rx_connection *rxconn;
#ifndef AFS_NOSTATS
struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
if (dcLocked && (tdc->index != NULLIDX)
&& !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
&& !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
- /* got the right one. It might not be the right version, and it
+ /* got the right one. It might not be the right version, and it
* might be fetching, but it's the right dcache entry.
*/
/* All this code should be integrated better with what follows:
ReleaseReadLock(&afs_xdcache);
shortcut = 1;
- if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
+ if (afs_IsDCacheFresh(tdc, avc)
&& !(tdc->dflags & DFFetching)) {
afs_stats_cmperf.dcacheHits++;
*/
if (!tdc) { /* If the hint wasn't the right dcache entry */
+ int dslot_error = 0;
/*
* Hash on the [fid, chunk] and get the corresponding dcache index
* after write-locking the dcache.
ObtainWriteLock(&afs_xdcache, 280);
us = NULLIDX;
- for (index = afs_dchashTbl[i]; index != NULLIDX;) {
+ for (index = afs_dchashTbl[i]; index != NULLIDX; us = index, index = afs_dcnextTbl[index]) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ /* we got an i/o error when trying to get the given dslot.
+ * it's possible the dslot we're looking for is elsewhere,
+ * but most likely the disk cache is currently unusable, so
+ * all afs_GetValidDSlot calls will fail, so just bail out. */
+ dslot_error = 1;
+ index = NULLIDX;
+ break;
+ }
ReleaseReadLock(&tdc->tlock);
/*
* Locks held:
afs_PutDCache(tdc);
tdc = 0;
}
- us = index;
- index = afs_dcnextTbl[index];
}
/*
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 (dslot_error) {
+ /* We couldn't find the dcache we want, but we hit some i/o
+ * errors when trying to find it, so we're not sure if the
+ * dcache we want is in the cache or not. Error out, so we
+ * don't try to possibly create 2 separate dcaches for the
+ * same exact data. */
+ ReleaseWriteLock(&afs_xdcache);
+ goto done;
+ }
+
if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
- while (1) {
- if (!setLocks)
- avc->f.states |= CDCLock;
- /* just need slots */
- afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
- if (!setLocks)
- avc->f.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");
+ if (!setLocks)
+ avc->f.states |= CDCLock;
+ /* just need slots */
+ afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
+ if (!setLocks)
+ avc->f.states &= ~CDCLock;
+ }
+ code = afs_AllocDCache(&tdc, avc, chunk, aflags, NULL);
+ if (code) {
+ ReleaseWriteLock(&afs_xdcache);
+ if (code == ENOSPC) {
+ /* It looks like afs_AllocDCache failed because we don't
+ * have any free dslots to use. Maybe if we wait a little
+ * while, we'll be able to free up some slots, so try for 5
+ * minutes, then bail out. */
+ if (++downDCount > 300) {
+ afs_warn("afs: Unable to get free cache space for file "
+ "%u:%u.%u.%u for 5 minutes; failing with an i/o error\n",
+ avc->f.fid.Cell,
+ avc->f.fid.Fid.Volume,
+ avc->f.fid.Fid.Vnode,
+ avc->f.fid.Fid.Unique);
+ goto done;
}
- ReleaseWriteLock(&afs_xdcache);
- /*
- * Locks held:
- * avc->lock(R) if setLocks
- * avc->lock(W) if !setLocks
- */
- afs_osi_Wait(1000, 0, 0);
- goto RetryLookup;
- }
+ afs_osi_Wait(1000, 0, 0);
+ goto RetryLookup;
+ }
+
+ /* afs_AllocDCache failed, but not because we're out of free
+ * dslots. Something must be screwy with the cache, so bail out
+ * immediately without waiting. */
+ afs_warn("afs: Error while alloc'ing cache slot for file "
+ "%u:%u.%u.%u; failing with an i/o error\n",
+ avc->f.fid.Cell,
+ avc->f.fid.Fid.Volume,
+ avc->f.fid.Fid.Vnode,
+ avc->f.fid.Fid.Unique);
+ goto done;
}
- tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
+ /*
+ * Locks held:
+ * avc->lock(R) if setLocks
+ * avc->lock(W) if !setLocks
+ * tdc->lock(W)
+ * afs_xdcache(W)
+ */
/*
* Now add to the two hash chains - note that i is still set
if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
#endif
#endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
- !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
+ !afs_IsDCacheFresh(tdc, avc))
doReallyAdjustSize = 1;
if (doReallyAdjustSize || overWriteWholeChunk) {
/* no data in file to read at this position */
UpgradeSToWLock(&tdc->lock, 607);
file = afs_CFileOpen(&tdc->f.inode);
+ osi_Assert(file);
afs_CFileTruncate(file, 0);
afs_CFileClose(file);
afs_AdjustSize(tdc, 0);
* avc->lock(W) if !setLocks || slowPass
* tdc->lock(S)
*/
- if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
+ if (!afs_IsDCacheFresh(tdc, avc) && !overWriteWholeChunk) {
/*
* Version number mismatch.
*/
*/
/* Watch for standard race condition around osi_FlushText */
- if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
+ if (afs_IsDCacheFresh(tdc, avc)) {
updateV2DC(setLocks, avc, tdc, 569); /* set hint */
afs_stats_cmperf.dcacheHits++;
ConvertWToSLock(&tdc->lock);
goto RetryGetDCache;
}
- /* Do not fetch data beyond truncPos. */
- maxGoodLength = avc->f.m.Length;
- if (avc->f.truncPos < maxGoodLength)
- maxGoodLength = avc->f.truncPos;
Position = AFS_CHUNKBASE(abyte);
if (vType(avc) == VDIR) {
size = avc->f.m.Length;
}
size = 999999999; /* max size for transfer */
} else {
+ afs_size_t maxGoodLength;
+
+ /* estimate how much data we're expecting back from the server,
+ * and reserve space in the dcache entry for it */
+
+ maxGoodLength = avc->f.m.Length;
+ if (avc->f.truncPos < maxGoodLength)
+ maxGoodLength = avc->f.truncPos;
+
size = AFS_CHUNKSIZE(abyte); /* expected max size */
- /* don't read past end of good data on server */
- if (Position + size > maxGoodLength)
+ if (Position > maxGoodLength) { /* If we're beyond EOF */
+ size = 0;
+ } else if (Position + size > maxGoodLength) {
size = maxGoodLength - Position;
- if (size < 0)
- size = 0; /* Handle random races */
+ }
+ osi_Assert(size >= 0);
+
if (size > tdc->f.chunkBytes) {
- /* pre-reserve space for file */
+ /* pre-reserve estimated space for file */
afs_AdjustSize(tdc, size); /* changes chunkBytes */
- /* max size for transfer still in size */
+ }
+
+ if (size) {
+ /* For the actual fetch, do not limit the request to the
+ * length of the file. If this results in a read past EOF on
+ * the server, the server will just reply with less data than
+ * requested. If we limit ourselves to only requesting data up
+ * to the avc file length, we open ourselves up to races if the
+ * file is extended on the server at about the same time.
+ *
+ * However, we must restrict ourselves to the avc->f.truncPos
+ * length, since this represents an outstanding local
+ * truncation of the file that will be committed to the
+ * fileserver when we actually write the fileserver contents.
+ * If we do not restrict the fetch length based on
+ * avc->f.truncPos, a different truncate operation extending
+ * the file length could cause the old data after
+ * avc->f.truncPos to reappear, instead of extending the file
+ * with NUL bytes. */
+ size = AFS_CHUNKSIZE(abyte);
+ if (Position > avc->f.truncPos) {
+ size = 0;
+ } else if (Position + size > avc->f.truncPos) {
+ size = avc->f.truncPos - Position;
+ }
+ osi_Assert(size >= 0);
}
}
if (afs_mariner && !tdc->f.chunk)
*/
DZap(tdc); /* pages in cache may be old */
file = afs_CFileOpen(&tdc->f.inode);
+ if (!file) {
+ /* We can't access the file in the disk cache backing this dcache;
+ * bail out. */
+ ReleaseWriteLock(&tdc->lock);
+ afs_PutDCache(tdc);
+ tdc = NULL;
+ goto done;
+ }
afs_RemoveVCB(&avc->f.fid);
tdc->f.states |= DWriting;
tdc->dflags |= DFFetching;
ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
tdc->dflags);
}
- tsmall =
- (struct afs_FetchOutput *)osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
+ tsmall = osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
setVcacheStatus = 0;
#ifndef AFS_NOSTATS
/*
* tdc->lock(W)
*/
- tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK);
+ tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
if (tc) {
#ifndef AFS_NOSTATS
numFetchLoops++;
#endif /* AFS_NOSTATS */
if (!setLocks || slowPass) {
- avc->callback = tc->srvr->server;
+ avc->callback = tc->parent->srvr->server;
} else {
- newCallback = tc->srvr->server;
+ newCallback = tc->parent->srvr->server;
setNewCallback = 1;
}
i = osi_Time();
- code = afs_CacheFetchProc(tc, file, Position, tdc,
+ code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
avc, size, tsmall);
} else
code = -1;
if (code == 0) {
- /* callback could have been broken (or expired) in a race here,
+ /* callback could have been broken (or expired) in a race here,
* but we return the data anyway. It's as good as we knew about
* when we started. */
- /*
- * validPos is updated by CacheFetchProc, and can only be
- * modifed under a dcache write lock, which we've blocked out
+ /*
+ * validPos is updated by CacheFetchProc, and can only be
+ * modifed under a dcache write lock, which we've blocked out
*/
+ afs_size_t length;
+
size = tdc->validPos - Position; /* actual segment size */
if (size < 0)
size = 0;
afs_CFileTruncate(file, size); /* prune it */
- } else {
+
+ /* Check that the amount of data that we fetched for the
+ * dcache makes sense. */
+ FillInt64(length, tsmall->OutStatus.Length_hi, tsmall->OutStatus.Length);
+ if (!IsDCacheSizeOK(tdc, avc, size,
+ length,
+ tsmall->OutStatus.DataVersion, 1)) {
+ code = EIO;
+ }
+ }
+ if (code) {
if (!setLocks || slowPass) {
- ObtainWriteLock(&afs_xcbhash, 453);
- afs_DequeueCallback(avc);
- avc->f.states &= ~(CStatd | CUnique);
- avc->callback = NULL;
- ReleaseWriteLock(&afs_xcbhash);
- if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
- osi_dnlc_purgedp(avc);
+ afs_StaleVCacheFlags(avc, AFS_STALEVC_CLEARCB, CUnique);
} else {
/* Something lost. Forget about performance, and go
* back with a vcache write lock.
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
tdc = 0;
+
+ /*
+ * Call afs_Analyze to manage the connection references
+ * and handle the error code (possibly mark servers
+ * down, etc). We are going to retry getting the
+ * dcache regardless, so we just ignore the retry hint
+ * returned by afs_Analyze on this call.
+ */
+ (void)afs_Analyze(tc, rxconn, code, &avc->f.fid, areq,
+ AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL);
+
ReleaseReadLock(&avc->lock);
+
slowPass = 1;
goto RetryGetDCache;
}
}
} while (afs_Analyze
- (tc, code, &avc->f.fid, areq,
+ (tc, rxconn, code, &avc->f.fid, areq,
AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
/*
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
if (!afs_IsDynroot(avc)) {
- ObtainWriteLock(&afs_xcbhash, 454);
- afs_DequeueCallback(avc);
- avc->f.states &= ~(CStatd | CUnique);
- ReleaseWriteLock(&afs_xcbhash);
- if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
- osi_dnlc_purgedp(avc);
+ afs_StaleVCacheFlags(avc, 0, CUnique);
/*
* Locks held:
* avc->lock(W); assert(!setLocks || slowPass)
* Environment:
* The afs_xdcache is write-locked through this whole affair.
*/
-void
+int
afs_WriteThroughDSlots(void)
{
- register struct dcache *tdc;
- register afs_int32 i, touchedit = 0;
+ struct dcache *tdc;
+ afs_int32 i, touchedit = 0;
+ int code = 0;
struct afs_q DirtyQ, *tq;
#define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
- for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
+ for (tq = DirtyQ.prev; tq != &DirtyQ && code == 0; tq = QPrev(tq)) {
tdc = DQTODC(tq);
if (tdc->dflags & DFEntryMod) {
int wrLock;
if (wrLock && (tdc->dflags & DFEntryMod)) {
tdc->dflags &= ~DFEntryMod;
ObtainWriteLock(&afs_xdcache, 620);
- afs_WriteDCache(tdc, 1);
+ code = afs_WriteDCache(tdc, 1);
ReleaseWriteLock(&afs_xdcache);
- touchedit = 1;
+ if (code) {
+ /* We didn't successfully write out the dslot; make sure we
+ * try again later */
+ tdc->dflags |= DFEntryMod;
+ } else {
+ touchedit = 1;
+ }
}
if (wrLock)
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
}
+ if (code) {
+ return code;
+ }
+
ObtainWriteLock(&afs_xdcache, 617);
if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
/* Touch the file to make sure that the mtime on the file is kept
*/
struct afs_fheader theader;
- theader.magic = AFS_FHMAGIC;
- theader.firstCSize = AFS_FIRSTCSIZE;
- theader.otherCSize = AFS_OTHERCSIZE;
- theader.version = AFS_CI_VERSION;
- theader.dataSize = sizeof(struct fcache);
+ afs_InitFHeader(&theader);
afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
}
ReleaseWriteLock(&afs_xdcache);
+ return 0;
}
/*
*
* Parameters:
* aslot : Dcache slot to look at.
- * tmpdc : Ptr to dcache entry.
+ * type : What 'type' of dslot to get; see the dslot_state enum
*
* Environment:
* Must be called with afs_xdcache write-locked.
*/
struct dcache *
-afs_MemGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
+afs_MemGetDSlot(afs_int32 aslot, dslot_state type)
{
- register struct dcache *tdc;
+ struct dcache *tdc;
int existing = 0;
AFS_STATCNT(afs_MemGetDSlot);
ConvertWToRLock(&tdc->tlock);
return tdc;
}
- if (tmpdc == NULL) {
- if (!afs_freeDSList)
- afs_GetDownDSlot(4);
- if (!afs_freeDSList) {
- /* none free, making one is better than a panic */
- afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
- tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
+
+ /* if we got here, the given slot is not in memory in our list of known
+ * slots. for memcache, the only place a dslot can exist is in memory, so
+ * if the caller is expecting to get back a known dslot, and we've reached
+ * here, something is very wrong. DSLOT_NEW is the only type of dslot that
+ * may not exist; for all others, the caller assumes the given dslot
+ * already exists. so, 'type' had better be DSLOT_NEW here, or something is
+ * very wrong. */
+ osi_Assert(type == DSLOT_NEW);
+
+ if (!afs_freeDSList)
+ afs_GetDownDSlot(4);
+ if (!afs_freeDSList) {
+ /* none free, making one is better than a panic */
+ afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
+ tdc = afs_osi_Alloc(sizeof(struct dcache));
+ osi_Assert(tdc != NULL);
#ifdef KERNEL_HAVE_PIN
- pin((char *)tdc, sizeof(struct dcache)); /* XXX */
+ pin((char *)tdc, sizeof(struct dcache)); /* XXX */
#endif
- } else {
- tdc = afs_freeDSList;
- afs_freeDSList = (struct dcache *)tdc->lruq.next;
- existing = 1;
- }
- 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");
} else {
- tdc = tmpdc;
- tdc->f.states = 0;
+ tdc = afs_freeDSList;
+ afs_freeDSList = (struct dcache *)tdc->lruq.next;
+ existing = 1;
}
+ 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");
/* initialize entry */
tdc->f.fid.Cell = 0;
AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
ObtainReadLock(&tdc->tlock);
- if (tmpdc == NULL)
- afs_indexTable[aslot] = tdc;
+ afs_indexTable[aslot] = tdc;
return tdc;
} /*afs_MemGetDSlot */
*
* Parameters:
* aslot : Dcache slot to look at.
- * tmpdc : Ptr to dcache entry.
+ * type : What 'type' of dslot to get; see the dslot_state enum
*
* Environment:
* afs_xdcache lock write-locked.
*/
struct dcache *
-afs_UFSGetDSlot(register afs_int32 aslot, register struct dcache *tmpdc)
+afs_UFSGetDSlot(afs_int32 aslot, dslot_state type)
{
- register afs_int32 code;
- register struct dcache *tdc;
+ afs_int32 code;
+ struct dcache *tdc;
int existing = 0;
int entryok;
+ int off;
AFS_STATCNT(afs_UFSGetDSlot);
if (CheckLock(&afs_xdcache) != -1)
ConvertWToRLock(&tdc->tlock);
return tdc;
}
+
/* otherwise we should read it in from the cache file */
- /*
- * If we weren't passed an in-memory region to place the file info,
- * we have to allocate one.
- */
- if (tmpdc == NULL) {
- if (!afs_freeDSList)
- afs_GetDownDSlot(4);
- if (!afs_freeDSList) {
- /* none free, making one is better than a panic */
- afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
- tdc = (struct dcache *)afs_osi_Alloc(sizeof(struct dcache));
+ if (!afs_freeDSList)
+ afs_GetDownDSlot(4);
+ if (!afs_freeDSList) {
+ /* none free, making one is better than a panic */
+ afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
+ tdc = afs_osi_Alloc(sizeof(struct dcache));
+ osi_Assert(tdc != NULL);
#ifdef KERNEL_HAVE_PIN
- pin((char *)tdc, sizeof(struct dcache)); /* XXX */
+ pin((char *)tdc, sizeof(struct dcache)); /* XXX */
#endif
- } else {
- tdc = afs_freeDSList;
- afs_freeDSList = (struct dcache *)tdc->lruq.next;
- existing = 1;
- }
- 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");
} else {
- tdc = tmpdc;
- tdc->f.states = 0;
+ tdc = afs_freeDSList;
+ afs_freeDSList = (struct dcache *)tdc->lruq.next;
+ existing = 1;
}
+ 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");
/*
* Seek to the aslot'th entry and read it in.
*/
+ off = sizeof(struct fcache)*aslot + sizeof(struct afs_fheader);
code =
afs_osi_Read(afs_cacheInodep,
- sizeof(struct fcache) * aslot +
- sizeof(struct afs_fheader), (char *)(&tdc->f),
+ off, (char *)(&tdc->f),
sizeof(struct fcache));
entryok = 1;
- if (code != sizeof(struct fcache))
+ if (code != sizeof(struct fcache)) {
entryok = 0;
- if (!afs_CellNumValid(tdc->f.fid.Cell))
+#if defined(KERNEL_HAVE_UERROR)
+ last_error = getuerror();
+#else
+ last_error = code;
+#endif
+ lasterrtime = osi_Time();
+ if (type != DSLOT_NEW) {
+ /* If we are requesting a non-DSLOT_NEW slot, this is an error.
+ * non-DSLOT_NEW slots are supposed to already exist, so if we
+ * failed to read in the slot, something is wrong. */
+ struct osi_stat tstat;
+ if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
+ tstat.size = -1;
+ }
+ afs_warn("afs: disk cache read error in CacheItems slot %d "
+ "off %d/%d code %d/%d\n",
+ (int)aslot,
+ off, (int)tstat.size,
+ (int)code, (int)sizeof(struct fcache));
+ /* put tdc back on the free dslot list */
+ QRemove(&tdc->lruq);
+ tdc->index = NULLIDX;
+ tdc->lruq.next = (struct afs_q *)afs_freeDSList;
+ afs_freeDSList = tdc;
+ return NULL;
+ }
+ }
+ if (!afs_CellNumValid(tdc->f.fid.Cell)) {
entryok = 0;
+ if (type == DSLOT_VALID) {
+ osi_Panic("afs: needed valid dcache but index %d off %d has "
+ "invalid cell num %d\n",
+ (int)aslot, off, (int)tdc->f.fid.Cell);
+ }
+ }
+
+ if (type == DSLOT_VALID && tdc->f.fid.Fid.Volume == 0) {
+ osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
+ (int)aslot, off);
+ }
+
+ if (type == DSLOT_UNUSED) {
+ /* the requested dslot is known to exist, but contain invalid data
+ * (this happens when we're using a dslot from the free or discard
+ * list). be sure not to re-use the data in it, so force invalidation.
+ */
+ entryok = 0;
+ }
if (!entryok) {
tdc->f.fid.Cell = 0;
tdc->f.chunk = -1;
hones(tdc->f.versionNo);
tdc->dflags |= DFEntryMod;
-#if defined(KERNEL_HAVE_UERROR)
- last_error = getuerror();
-#endif
- lasterrtime = osi_Time();
afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, 0, 0);
} else {
- if (&tdc->f != 0) {
- if (tdc->f.states & DRO) {
- afs_DCMoveBucket(tdc, 0, 2);
- } else if (tdc->f.states & DBackup) {
- afs_DCMoveBucket(tdc, 0, 1);
- } else {
- afs_DCMoveBucket(tdc, 0, 1);
- }
- }
+ if (tdc->f.states & DRO) {
+ afs_DCMoveBucket(tdc, 0, 2);
+ } else if (tdc->f.states & DBackup) {
+ afs_DCMoveBucket(tdc, 0, 1);
+ } else {
+ afs_DCMoveBucket(tdc, 0, 1);
+ }
}
tdc->refCount = 1;
tdc->index = aslot;
* If we didn't read into a temporary dcache region, update the
* slot pointer table.
*/
- if (tmpdc == NULL)
- afs_indexTable[aslot] = tdc;
+ afs_indexTable[aslot] = tdc;
return tdc;
} /*afs_UFSGetDSlot */
*/
int
-afs_WriteDCache(register struct dcache *adc, int atime)
+afs_WriteDCache(struct dcache *adc, int atime)
{
- register afs_int32 code;
+ afs_int32 code;
if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
return 0;
osi_Assert(WriteLocked(&afs_xdcache));
if (atime)
adc->f.modTime = osi_Time();
+
+ if ((afs_indexFlags[adc->index] & (IFFree | IFDiscarded)) == 0 &&
+ adc->f.fid.Fid.Volume == 0) {
+ /* If a dcache slot is not on the free or discard list, it must be
+ * in the hash table. Thus, the volume must be non-zero, since that
+ * is how we determine whether or not to unhash the entry when kicking
+ * it out of the cache. Do this check now, since otherwise this can
+ * cause hash table corruption and a panic later on after we read the
+ * entry back in. */
+ osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
+ adc->index, (unsigned)afs_indexFlags[adc->index]);
+ }
+
/*
* Seek to the right dcache slot and write the in-memory image out to disk.
*/
sizeof(struct fcache) * adc->index +
sizeof(struct afs_fheader), (char *)(&adc->f),
sizeof(struct fcache));
- if (code != sizeof(struct fcache))
+ if (code != sizeof(struct fcache)) {
+ afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
+ (long)(sizeof(struct fcache) * adc->index + sizeof(struct afs_fheader)),
+ (int)code, (int)sizeof(struct fcache));
return EIO;
+ }
return 0;
}
* Nothing interesting.
*/
int
-afs_wakeup(register struct vcache *avc)
+afs_wakeup(struct vcache *avc)
{
- register int i;
- register struct brequest *tb;
+ int i;
+ struct brequest *tb;
tb = afs_brs;
AFS_STATCNT(afs_wakeup);
for (i = 0; i < NBRS; i++, tb++) {
* is already being handled by the higher-level code.
*/
if ((avc->f.states & CSafeStore) == 0) {
- tb->code = 0;
+ tb->code_raw = tb->code_checkcode = 0;
tb->flags |= BUVALID;
if (tb->flags & BUWAIT) {
tb->flags &= ~BUWAIT;
return 0;
}
-
/*!
* Given a file name and inode, set up that file to be an
* active member in the AFS cache. This also involves checking
int
afs_InitCacheFile(char *afile, ino_t ainode)
{
- register afs_int32 code;
+ afs_int32 code;
afs_int32 index;
int fileIsBad;
struct osi_file *tfile;
struct osi_stat tstat;
- register struct dcache *tdc;
+ struct dcache *tdc;
AFS_STATCNT(afs_InitCacheFile);
index = afs_stats_cmperf.cacheNumEntries;
return EINVAL;
ObtainWriteLock(&afs_xdcache, 282);
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetNewDSlot(index);
ReleaseReadLock(&tdc->tlock);
ReleaseWriteLock(&afs_xdcache);
ObtainWriteLock(&tdc->lock, 621);
ObtainWriteLock(&afs_xdcache, 622);
- if (afile) {
- code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
- if (code) {
- ReleaseWriteLock(&afs_xdcache);
- ReleaseWriteLock(&tdc->lock);
- afs_PutDCache(tdc);
- return code;
- }
+ if (!afile && !ainode) {
+ tfile = NULL;
+ fileIsBad = 1;
} else {
- /* Add any other 'complex' inode types here ... */
+ if (afile) {
+ code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
+ if (code) {
+ ReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&tdc->lock);
+ afs_PutDCache(tdc);
+ return code;
+ }
+ } else {
+ /* Add any other 'complex' inode types here ... */
#if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
- tdc->f.inode.ufs = ainode;
+ tdc->f.inode.ufs = ainode;
#else
- osi_Panic("Can't init cache with inode numbers when complex inodes are "
- "in use\n");
+ osi_Panic("Can't init cache with inode numbers when complex inodes are "
+ "in use\n");
#endif
- }
- fileIsBad = 0;
- if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
- fileIsBad = 1;
- tfile = osi_UFSOpen(&tdc->f.inode);
- code = afs_osi_Stat(tfile, &tstat);
- if (code)
- osi_Panic("initcachefile stat");
+ }
+ fileIsBad = 0;
+ if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
+ fileIsBad = 1;
+ tfile = osi_UFSOpen(&tdc->f.inode);
+ if (!tfile) {
+ ReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&tdc->lock);
+ afs_PutDCache(tdc);
+ return ENOENT;
+ }
- /*
- * If file size doesn't match the cache info file, it's probably bad.
- */
- if (tdc->f.chunkBytes != tstat.size)
- fileIsBad = 1;
+ code = afs_osi_Stat(tfile, &tstat);
+ if (code)
+ osi_Panic("initcachefile stat");
+
+ /*
+ * If file size doesn't match the cache info file, it's probably bad.
+ */
+ if (tdc->f.chunkBytes != tstat.size)
+ fileIsBad = 1;
+ /*
+ * If file changed within T (120?) seconds of cache info file, it's
+ * probably bad. In addition, if slot changed within last T seconds,
+ * the cache info file may be incorrectly identified, and so slot
+ * may be bad.
+ */
+ if (cacheInfoModTime < tstat.mtime + 120)
+ fileIsBad = 1;
+ if (cacheInfoModTime < tdc->f.modTime + 120)
+ fileIsBad = 1;
+ /* In case write through is behind, make sure cache items entry is
+ * at least as new as the chunk.
+ */
+ if (tdc->f.modTime < tstat.mtime)
+ fileIsBad = 1;
+ }
tdc->f.chunkBytes = 0;
- /*
- * If file changed within T (120?) seconds of cache info file, it's
- * probably bad. In addition, if slot changed within last T seconds,
- * the cache info file may be incorrectly identified, and so slot
- * may be bad.
- */
- if (cacheInfoModTime < tstat.mtime + 120)
- fileIsBad = 1;
- if (cacheInfoModTime < tdc->f.modTime + 120)
- fileIsBad = 1;
- /* In case write through is behind, make sure cache items entry is
- * at least as new as the chunk.
- */
- if (tdc->f.modTime < tstat.mtime)
- fileIsBad = 1;
if (fileIsBad) {
tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
- if (tstat.size != 0)
+ if (tfile && tstat.size != 0)
osi_UFSTruncate(tfile, 0);
tdc->f.states &= ~(DRO|DBackup|DRW);
afs_DCMoveBucket(tdc, 0, 0);
afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
} /*File is not bad */
- osi_UFSClose(tfile);
+ if (tfile)
+ osi_UFSClose(tfile);
tdc->f.states &= ~DWriting;
tdc->dflags &= ~DFEntryMod;
/* don't set f.modTime; we're just cleaning up */
- afs_WriteDCache(tdc, 0);
+ osi_Assert(afs_WriteDCache(tdc, 0) == 0);
ReleaseWriteLock(&afs_xdcache);
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
* Initialize dcache related variables.
*
* \param afiles
- * \param ablocks
+ * \param ablocks
* \param aDentries
* \param achunk
* \param aflags
*
*/
-void
+int
afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
{
- register struct dcache *tdp;
+ struct dcache *tdp;
int i;
int code;
+ int afs_dhashbits;
afs_freeDCList = NULLIDX;
afs_discardDCList = NULLIDX;
if (!aDentries)
aDentries = DDSIZE;
- if (aflags & AFSCALL_INIT_MEMCACHE) {
- /*
- * Use a memory cache instead of a disk cache
- */
- cacheDiskType = AFS_FCACHE_TYPE_MEM;
- afs_cacheType = &afs_MemCacheOps;
- afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
- ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
- /* ablocks is reported in 1K blocks */
- code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
- if (code != 0) {
- afs_warn("afsd: memory cache too large for available memory.\n");
- afs_warn("afsd: AFS files cannot be accessed.\n\n");
- dcacheDisabled = 1;
- afiles = ablocks = 0;
- } else
- afs_warn("Memory cache: Allocating %d dcache entries...",
- aDentries);
- } else {
- cacheDiskType = AFS_FCACHE_TYPE_UFS;
- afs_cacheType = &afs_UfsCacheOps;
- }
-
+ /* afs_dhashsize defaults to 1024 */
if (aDentries > 512)
afs_dhashsize = 2048;
+ /* Try to keep the average chain length around two unless the table
+ * would be ridiculously big. */
+ if (aDentries > 4096) {
+ afs_dhashbits = opr_fls(aDentries) - 3;
+ /* Cap the hash tables to 32k entries. */
+ if (afs_dhashbits > 15)
+ afs_dhashbits = 15;
+ afs_dhashsize = opr_jhash_size(afs_dhashbits);
+ }
/* initialize hash tables */
- afs_dvhashTbl =
- (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
- afs_dchashTbl =
- (afs_int32 *) afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
+ afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
+ osi_Assert(afs_dvhashTbl != NULL);
+ afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
+ osi_Assert(afs_dchashTbl != NULL);
for (i = 0; i < afs_dhashsize; i++) {
afs_dvhashTbl[i] = NULLIDX;
afs_dchashTbl[i] = NULLIDX;
}
- afs_dvnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
- afs_dcnextTbl = (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_int32));
+ afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
+ osi_Assert(afs_dvnextTbl != NULL);
+ afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
+ osi_Assert(afs_dcnextTbl != NULL);
for (i = 0; i < afiles; i++) {
afs_dvnextTbl[i] = NULLIDX;
afs_dcnextTbl[i] = NULLIDX;
}
/* Allocate and zero the pointer array to the dcache entries */
- afs_indexTable = (struct dcache **)
- afs_osi_Alloc(sizeof(struct dcache *) * afiles);
+ afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
+ osi_Assert(afs_indexTable != NULL);
memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
- afs_indexTimes =
- (afs_hyper_t *) afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
+ afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
+ osi_Assert(afs_indexTimes != NULL);
memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
- afs_indexUnique =
- (afs_int32 *) afs_osi_Alloc(afiles * sizeof(afs_uint32));
+ afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
+ osi_Assert(afs_indexUnique != NULL);
memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
- afs_indexFlags = (u_char *) afs_osi_Alloc(afiles * sizeof(u_char));
+ afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
+ osi_Assert(afs_indexFlags != NULL);
memset(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));
+ afs_osi_Alloc(aDentries * sizeof(struct dcache));
+ osi_Assert(tdp != NULL);
memset(tdp, 0, aDentries * sizeof(struct dcache));
#ifdef KERNEL_HAVE_PIN
pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
afs_dcentries = aDentries;
afs_blocksUsed = 0;
- afs_stats_cmperf.cacheBucket0_Discarded =
- afs_stats_cmperf.cacheBucket1_Discarded =
+ afs_stats_cmperf.cacheBucket0_Discarded =
+ afs_stats_cmperf.cacheBucket1_Discarded =
afs_stats_cmperf.cacheBucket2_Discarded = 0;
afs_DCSizeInit();
QInit(&afs_DLRU);
+
+ if (aflags & AFSCALL_INIT_MEMCACHE) {
+ /*
+ * Use a memory cache instead of a disk cache
+ */
+ cacheDiskType = AFS_FCACHE_TYPE_MEM;
+ afs_cacheType = &afs_MemCacheOps;
+ afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
+ ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
+ /* ablocks is reported in 1K blocks */
+ code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
+ if (code != 0) {
+ afs_warn("afsd: memory cache too large for available memory.\n");
+ afs_warn("afsd: AFS files cannot be accessed.\n\n");
+ dcacheDisabled = 1;
+ return code;
+ } else
+ afs_warn("Memory cache: Allocating %d dcache entries...",
+ aDentries);
+ } else {
+ cacheDiskType = AFS_FCACHE_TYPE_UFS;
+ afs_cacheType = &afs_UfsCacheOps;
+ }
+ return 0;
}
/*!
afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
afs_blocksUsed = afs_dcentries = 0;
- afs_stats_cmperf.cacheBucket0_Discarded =
- afs_stats_cmperf.cacheBucket1_Discarded =
+ afs_stats_cmperf.cacheBucket0_Discarded =
+ afs_stats_cmperf.cacheBucket1_Discarded =
afs_stats_cmperf.cacheBucket2_Discarded = 0;
hzero(afs_indexCounter);
/*!
* Get a dcache ready for writing, respecting the current cache size limits
*
- * len is required because afs_GetDCache with flag == 4 expects the length
- * field to be filled. It decides from this whether it's necessary to fetch
- * data into the chunk before writing or not (when the whole chunk is
+ * len is required because afs_GetDCache with flag == 4 expects the length
+ * field to be filled. It decides from this whether it's necessary to fetch
+ * data into the chunk before writing or not (when the whole chunk is
* overwritten!).
*
* \param avc The vcache to fetch a dcache for
* \param noLock
*
* \return If successful, a reference counted dcache with tdc->lock held. Lock
- * must be released and afs_PutDCache() called to free dcache.
+ * must be released and afs_PutDCache() called to free dcache.
* NULL on failure
*
- * \note avc->lock must be held on entry. Function may release and reobtain
+ * \note avc->lock must be held on entry. Function may release and reobtain
* avc->lock and GLOCK.
*/
struct dcache *
-afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
+afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
afs_size_t len, struct vrequest *areq,
int noLock)
{
tdc = afs_FindDCache(avc, filePos);
if (tdc) {
ObtainWriteLock(&tdc->lock, 658);
- if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
+ if (!afs_IsDCacheFresh(tdc, avc)
|| (tdc->dflags & DFFetching)) {
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
return tdc;
}
-#if defined(AFS_DISCON_ENV)
-
/*!
* Make a shadow copy of a dir's dcache. It's used for disconnected
* operations like remove/create/rename to keep the original directory data.
ObtainWriteLock(&afs_xdcache, 716);
/* Get a fresh dcache. */
- new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
+ (void)afs_AllocDCache(&new_dc, avc, 0, 0, &shadow_fid);
+ osi_Assert(new_dc);
ObtainReadLock(&adc->mflock);
new_dc->f.chunkBytes = adc->f.chunkBytes;
ReleaseReadLock(&adc->mflock);
-
+
/* Now add to the two hash chains */
i = DCHash(&shadow_fid, 0);
afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
ReleaseWriteLock(&afs_xdcache);
/* Alloc a 4k block. */
- data = (char *) afs_osi_Alloc(4096);
+ data = afs_osi_Alloc(4096);
if (!data) {
afs_warn("afs_MakeShadowDir: could not alloc data\n");
ret_code = ENOMEM;
/* Open the files. */
tfile_src = afs_CFileOpen(&adc->f.inode);
tfile_dst = afs_CFileOpen(&new_dc->f.inode);
+ osi_Assert(tfile_src);
+ osi_Assert(tfile_dst);
/* And now copy dir dcache data into this dcache,
* 4k at a time.
ObtainWriteLock(&afs_xvcache, 763);
ObtainWriteLock(&afs_disconDirtyLock, 765);
QAdd(&afs_disconShadow, &avc->shadowq);
- osi_vnhold(avc, 0);
+ osi_Assert((afs_RefVCache(avc) == 0));
ReleaseWriteLock(&afs_disconDirtyLock);
ReleaseWriteLock(&afs_xvcache);
* \param alen The new length of the file
*
*/
-void
+void
afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
{
struct dcache *tdc;
afs_int32 start, end;
/* We're doing this to deal with the situation where we extend
- * by writing after lseek()ing past the end of the file . If that
- * extension skips chunks, then those chunks won't be created, and
- * GetDCache will assume that they have to be fetched from the server.
- * So, for each chunk between the current file position, and the new
+ * by writing after lseek()ing past the end of the file . If that
+ * extension skips chunks, then those chunks won't be created, and
+ * GetDCache will assume that they have to be fetched from the server.
+ * So, for each chunk between the current file position, and the new
* length we GetDCache for that chunk.
*/
- if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
+ if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
return;
if (avc->f.m.Length == 0)
start = 0;
- else
+ else
start = AFS_CHUNK(avc->f.m.Length)+1;
end = AFS_CHUNK(apos);
start++;
}
}
-
-#endif
git://git.openafs.org / openafs.git / blobdiff