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)) {
+ afs_CacheTooFull = 0;
break;
}
if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
* 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) {
/*
* Get an entry from the list of discarded cache elements
*/
- tdc = afs_GetDSlot(afs_discardDCList, 0);
+ tdc = afs_GetNewDSlot(afs_discardDCList);
osi_Assert(tdc->refCount == 1);
ReleaseReadLock(&tdc->tlock);
}
#else
tdc->dflags &= ~DFEntryMod;
- afs_WriteDCache(tdc, 1);
+ osi_Assert(afs_WriteDCache(tdc, 1) == 0);
#endif
}
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) osi_Panic("afs_TryToSmush tdc");
if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
if (sync) {
if ((afs_indexFlags[index] & IFDataMod) == 0
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);
- if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
- totalChunks--;
- }
- ReleaseReadLock(&tdc->tlock);
- afs_PutDCache(tdc);
+ tdc = afs_GetValidDSlot(index);
+ if (tdc) {
+ if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
+ totalChunks--;
+ }
+ ReleaseReadLock(&tdc->tlock);
+ afs_PutDCache(tdc);
+ }
}
}
ReleaseWriteLock(&afs_xdcache);
ObtainWriteLock(&afs_xdcache, 278);
for (index = afs_dchashTbl[i]; index != NULLIDX;) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) osi_Panic("afs_FindDCache tdc");
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
break; /* leaving refCount high for caller */
|| ((lock & 2) && afs_freeDCList != NULLIDX)) {
afs_indexFlags[afs_freeDCList] &= ~IFFree;
- tdc = afs_GetDSlot(afs_freeDCList, 0);
+ tdc = afs_GetNewDSlot(afs_freeDCList);
osi_Assert(tdc->refCount == 1);
ReleaseReadLock(&tdc->tlock);
ObtainWriteLock(&tdc->lock, 604);
afs_freeDCCount--;
} else {
afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
- tdc = afs_GetDSlot(afs_discardDCList, 0);
+ tdc = afs_GetNewDSlot(afs_discardDCList);
osi_Assert(tdc->refCount == 1);
ReleaseReadLock(&tdc->tlock);
ObtainWriteLock(&tdc->lock, 605);
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;
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 */
us = NULLIDX;
for (index = afs_dchashTbl[i]; index != NULLIDX;) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ ReleaseWriteLock(&afs_xdcache);
+ goto done;
+ }
ReleaseReadLock(&tdc->tlock);
/*
* Locks held:
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)
size = maxGoodLength - Position;
if (size < 0)
size = 0; /* Handle random races */
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 + size > avc->f.truncPos) {
+ size = avc->f.truncPos - Position;
+ }
+ if (size < 0) {
+ size = 0;
+ }
}
}
if (afs_mariner && !tdc->f.chunk)
* 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++;
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;
}
} while (afs_Analyze
- (tc, code, &avc->f.fid, areq,
+ (tc, rxconn, code, &avc->f.fid, areq,
AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
/*
if (wrLock && (tdc->dflags & DFEntryMod)) {
tdc->dflags &= ~DFEntryMod;
ObtainWriteLock(&afs_xdcache, 620);
- afs_WriteDCache(tdc, 1);
+ osi_Assert(afs_WriteDCache(tdc, 1) == 0);
ReleaseWriteLock(&afs_xdcache);
touchedit = 1;
}
*
* Parameters:
* aslot : Dcache slot to look at.
- * tmpdc : Ptr to dcache entry.
*
* Environment:
* Must be called with afs_xdcache write-locked.
*/
struct dcache *
-afs_MemGetDSlot(afs_int32 aslot, struct dcache *tmpdc)
+afs_MemGetDSlot(afs_int32 aslot, int needvalid)
{
struct dcache *tdc;
int existing = 0;
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 = afs_osi_Alloc(sizeof(struct dcache));
- osi_Assert(tdc != NULL);
+
+ osi_Assert(needvalid);
+
+ 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.
*
* Environment:
* afs_xdcache lock write-locked.
*/
struct dcache *
-afs_UFSGetDSlot(afs_int32 aslot, struct dcache *tmpdc)
+afs_UFSGetDSlot(afs_int32 aslot, int needvalid)
{
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 = afs_osi_Alloc(sizeof(struct dcache));
- osi_Assert(tdc != 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 = 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();
+#endif
+ lasterrtime = osi_Time();
+ if (needvalid) {
+ struct osi_stat tstat;
+ if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
+ tstat.size = -1;
+ }
+ afs_warn("afs: disk cache read error in CacheItems off %d/%d "
+ "code %d/%d\n",
+ 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 (needvalid) {
+ 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 (needvalid && tdc->f.fid.Fid.Volume == 0) {
+ osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
+ (int)aslot, off);
+ }
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);
* 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 */
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;
}
return EINVAL;
ObtainWriteLock(&afs_xdcache, 282);
- tdc = afs_GetDSlot(index, NULL);
+ tdc = afs_GetNewDSlot(index);
ReleaseReadLock(&tdc->tlock);
ReleaseWriteLock(&afs_xdcache);
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);