/*
* Get an entry from the list of discarded cache elements
*/
- tdc = afs_GetNewDSlot(afs_discardDCList);
+ tdc = afs_GetUnusedDSlot(afs_discardDCList);
+ osi_Assert(tdc);
osi_Assert(tdc->refCount == 1);
ReleaseReadLock(&tdc->tlock);
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
int releaseTlock = 1;
tdc = afs_GetValidDSlot(index);
- if (!tdc) osi_Panic("afs_TryToSmush tdc");
+ if (!tdc) {
+ /* afs_TryToSmush is best-effort; we may not actually discard
+ * everything, so failure to discard a dcache due to an i/o
+ * error is okay. */
+ continue;
+ }
if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
if (sync) {
if ((afs_indexFlags[index] & IFDataMod) == 0
*/
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_GetValidDSlot(index);
- if (!tdc) osi_Panic("afs_FindDCache tdc");
+ 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. */
+ continue;
+ }
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);
|| ((lock & 2) && afs_freeDCList != NULLIDX)) {
afs_indexFlags[afs_freeDCList] &= ~IFFree;
- tdc = afs_GetNewDSlot(afs_freeDCList);
+ tdc = afs_GetUnusedDSlot(afs_freeDCList);
+ osi_Assert(tdc);
osi_Assert(tdc->refCount == 1);
ReleaseReadLock(&tdc->tlock);
ObtainWriteLock(&tdc->lock, 604);
afs_freeDCCount--;
} else {
afs_indexFlags[afs_discardDCList] &= ~IFDiscarded;
- tdc = afs_GetNewDSlot(afs_discardDCList);
+ tdc = afs_GetUnusedDSlot(afs_discardDCList);
+ osi_Assert(tdc);
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;
*/
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_GetValidDSlot(index);
if (!tdc) {
- ReleaseWriteLock(&afs_xdcache);
- goto done;
+ /* we got an i/o error when trying to get the given dslot,
+ * but do not bail out just yet; it is possible the dcache
+ * we're looking for is elsewhere, so it doesn't matter if
+ * we can't load this one. */
+ dslot_error = 1;
+ continue;
}
ReleaseReadLock(&tdc->tlock);
/*
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);
+ 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;
+ }
+
/* Make sure there is a free dcache entry for us to use */
if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
while (1) {
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)
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
/*
*
* Parameters:
* aslot : Dcache slot to look at.
+ * needvalid : Whether the specified slot should already exist
*
* Environment:
* Must be called with afs_xdcache write-locked.
*/
struct dcache *
-afs_MemGetDSlot(afs_int32 aslot, int needvalid)
+afs_MemGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
{
struct dcache *tdc;
int existing = 0;
return tdc;
}
- osi_Assert(!needvalid);
+ /* if 'indexvalid' is true, the slot must already exist and be populated
+ * somewhere. for memcache, the only place that dcache entries exist is
+ * in memory, so if we did not find it above, something is very wrong. */
+ osi_Assert(!indexvalid);
if (!afs_freeDSList)
afs_GetDownDSlot(4);
*
* Parameters:
* aslot : Dcache slot to look at.
+ * indexvalid : 1 if we know the slot we're giving is valid, and thus
+ * reading the dcache from the disk index should succeed. 0
+ * if we are initializing a new dcache, and so reading from
+ * the disk index may fail.
+ * datavalid : 0 if we are loading a dcache entry from the free or
+ * discard list, so we know the data in the given dcache is
+ * not valid. 1 if we are loading a known used dcache, so the
+ * data in the dcache must be valid.
*
* Environment:
* afs_xdcache lock write-locked.
*/
struct dcache *
-afs_UFSGetDSlot(afs_int32 aslot, int needvalid)
+afs_UFSGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
{
afs_int32 code;
struct dcache *tdc;
entryok = 0;
#if defined(KERNEL_HAVE_UERROR)
last_error = getuerror();
+#else
+ last_error = code;
#endif
lasterrtime = osi_Time();
- if (needvalid) {
+ if (indexvalid) {
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",
+ 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 */
}
if (!afs_CellNumValid(tdc->f.fid.Cell)) {
entryok = 0;
- if (needvalid) {
+ if (datavalid) {
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) {
+ if (datavalid && tdc->f.fid.Fid.Volume == 0) {
osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
(int)aslot, off);
}
- if (!entryok) {
+ if (!entryok || !datavalid) {
tdc->f.fid.Cell = 0;
tdc->f.fid.Fid.Volume = 0;
tdc->f.chunk = -1;
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 0;
}
-
/*!
* Given a file name and inode, set up that file to be an
* active member in the AFS cache. This also involves checking
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);
+ 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 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 */
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;
- }
-
if (aDentries > 512)
afs_dhashsize = 2048;
/* initialize hash tables */
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;
+ afiles = ablocks = 0;
+ } else
+ afs_warn("Memory cache: Allocating %d dcache entries...",
+ aDentries);
+ } else {
+ cacheDiskType = AFS_FCACHE_TYPE_UFS;
+ afs_cacheType = &afs_UfsCacheOps;
+ }
}
/*!