/*
* Copyright 2000, International Business Machines Corporation and others.
* 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
* We're write-locked upon entry.
*/
-int
-afs_StoreMini(register struct vcache *avc, struct vrequest *areq)
+static int
+afs_StoreMini(struct vcache *avc, struct vrequest *areq)
{
- register struct afs_conn *tc;
+ struct afs_conn *tc;
struct AFSStoreStatus InStatus;
struct AFSFetchStatus OutStatus;
struct AFSVolSync tsync;
- register afs_int32 code;
- register struct rx_call *tcall;
+ afs_int32 code;
+ struct rx_call *tcall;
+ struct rx_connection *rxconn;
afs_size_t tlen, xlen = 0;
XSTATS_DECLS;
AFS_STATCNT(afs_StoreMini);
avc->f.states &= ~CExtendedFile;
do {
- tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK);
+ tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
if (tc) {
#ifdef AFS_64BIT_CLIENT
retry:
#endif
RX_AFS_GUNLOCK();
- tcall = rx_NewCall(tc->id);
+ tcall = rx_NewCall(rxconn);
RX_AFS_GLOCK();
/* Set the client mod time since we always want the file
* to have the client's mod time and not the server's one
l2 = tlen;
if ((avc->f.m.Length > 0x7fffffff) ||
(tlen > 0x7fffffff) ||
- ((0x7fffffff - tlen) < avc->f.m.Length))
- return EFBIG;
+ ((0x7fffffff - tlen) < avc->f.m.Length)) {
+ code = EFBIG;
+ goto error;
+ }
code =
StartRXAFS_StoreData(tcall,
(struct AFSFid *)&avc->f.fid.Fid,
if (code == 0) {
code = EndRXAFS_StoreData(tcall, &OutStatus, &tsync);
}
+#ifdef AFS_64BIT_CLIENT
+ error:
+#endif
code = rx_EndCall(tcall, code);
RX_AFS_GLOCK();
XSTATS_END_TIME;
} else
code = -1;
} while (afs_Analyze
- (tc, code, &avc->f.fid, areq, AFS_STATS_FS_RPCIDX_STOREDATA,
+ (tc, rxconn, code, &avc->f.fid, areq, AFS_STATS_FS_RPCIDX_STOREDATA,
SHARED_LOCK, NULL));
- if (code == 0) {
+ if (code == 0)
afs_ProcessFS(avc, &OutStatus, areq);
- } else {
- /* blew it away */
- afs_InvalidateAllSegments(avc);
- }
- return code;
+ return code;
} /*afs_StoreMini */
-unsigned int storeallmissing = 0;
-#define lmin(a,b) (((a) < (b)) ? (a) : (b))
/*
* afs_StoreAllSegments
*
int
-afs_StoreAllSegments(register struct vcache *avc, struct vrequest *areq,
+afs_StoreAllSegments(struct vcache *avc, struct vrequest *areq,
int sync)
{
- register struct dcache *tdc;
- register afs_int32 code = 0;
- register afs_int32 index;
- register afs_int32 origCBs, foreign = 0;
+ struct dcache *tdc;
+ afs_int32 code = 0;
+ afs_int32 index;
+ afs_int32 origCBs, foreign = 0;
int hash;
afs_hyper_t newDV, oldDV; /* DV when we start, and finish, respectively */
- struct dcache **dcList, **dclist;
+ struct dcache **dcList;
unsigned int i, j, minj, moredata, high, off;
- afs_size_t tlen;
afs_size_t maxStoredLength; /* highest offset we've written to server. */
- int safety;
+ int safety, marineronce = 0;
AFS_STATCNT(afs_StoreAllSegments);
hset(newDV, avc->f.m.DataVersion);
hash = DVHash(&avc->f.fid);
foreign = (avc->f.states & CForeign);
- dcList = (struct dcache **)osi_AllocLargeSpace(AFS_LRALLOCSIZ);
+ dcList = osi_AllocLargeSpace(AFS_LRALLOCSIZ);
afs_Trace2(afs_iclSetp, CM_TRACE_STOREALL, ICL_TYPE_POINTER, avc,
ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
#if !defined(AFS_AIX32_ENV) && !defined(AFS_SGI65_ENV)
* on the memcache case since that's we adjust the file's size
* and finish flushing partial vm pages.
*/
- if ((cacheDiskType != AFS_FCACHE_TYPE_MEM) || (sync & AFS_LASTSTORE))
+ if ((cacheDiskType != AFS_FCACHE_TYPE_MEM) ||
+ (sync & AFS_VMSYNC_INVAL) || (sync & AFS_VMSYNC) ||
+ (sync & AFS_LASTSTORE))
#endif /* !AFS_AIX32_ENV && !AFS_SGI65_ENV */
{
/* If we're not diskless, reading a file may stress the VM
* - Have to get a write lock on xdcache because GetDSlot might need it (if
* the chunk doesn't have a dcache struct).
* This seems like overkill in most cases.
- * - I'm not sure that it's safe to do "index = .hvNextp", then unlock
+ * - I'm not sure that it's safe to do "index = .hvNextp", then unlock
* xdcache, then relock xdcache and try to use index. It is done
* a lot elsewhere in the CM, but I'm not buying that argument.
* - should be able to check IFDataMod without doing the GetDSlot (just
* hold afs_xdcache). That way, it's easy to do this without the
* writelock on afs_xdcache, and we save unneccessary disk
- * operations. I don't think that works, 'cuz the next pointers
+ * operations. I don't think that works, 'cuz the next pointers
* are still on disk.
*/
origCBs = afs_allCBs;
maxStoredLength = 0;
- tlen = avc->f.m.Length;
minj = 0;
do {
- memset((char *)dcList, 0, NCHUNKSATONCE * sizeof(struct dcache *));
+ memset(dcList, 0, NCHUNKSATONCE * sizeof(struct dcache *));
high = 0;
moredata = FALSE;
- /* lock and start over from beginning of hash chain
+ /* lock and start over from beginning of hash chain
* in order to avoid a race condition. */
- MObtainWriteLock(&afs_xdcache, 284);
+ ObtainWriteLock(&afs_xdcache, 284);
index = afs_dvhashTbl[hash];
for (j = 0; index != NULLIDX;) {
if ((afs_indexFlags[index] & IFDataMod)
&& (afs_indexUnique[index] == avc->f.fid.Fid.Unique)) {
- tdc = afs_GetDSlot(index, 0); /* refcount+1. */
+ tdc = afs_GetValidDSlot(index); /* refcount+1. */
+ if (!tdc) {
+ ReleaseWriteLock(&afs_xdcache);
+ code = EIO;
+ goto done;
+ }
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid) && tdc->f.chunk >= minj) {
off = tdc->f.chunk - minj;
if (off < NCHUNKSATONCE) {
if (dcList[off])
osi_Panic("dclist slot already in use!");
+ if (afs_mariner && !marineronce) {
+ /* first chunk only */
+ afs_MarinerLog("store$Storing", avc);
+ marineronce++;
+ }
dcList[off] = tdc;
if (off > high)
high = off;
}
index = afs_dvnextTbl[index];
}
- MReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&afs_xdcache);
/* this guy writes chunks, puts back dcache structs, and bumps newDV */
/* "moredata" just says "there are more dirty chunks yet to come".
*/
if (j) {
- struct AFSStoreStatus InStatus;
- struct AFSFetchStatus OutStatus;
- int doProcessFS = 0;
- afs_size_t base, bytes;
- afs_uint32 nchunks;
- int nomore;
- unsigned int first = 0;
- struct afs_conn *tc;
- for (bytes = 0, j = 0; !code && j <= high; j++) {
- if (dcList[j]) {
- ObtainSharedLock(&(dcList[j]->lock), 629);
- if (!bytes)
- first = j;
- bytes += dcList[j]->f.chunkBytes;
- if ((dcList[j]->f.chunkBytes < afs_OtherCSize)
- && (dcList[j]->f.chunk - minj < high)
- && dcList[j + 1]) {
- int sbytes = afs_OtherCSize - dcList[j]->f.chunkBytes;
- bytes += sbytes;
- }
- }
- if (bytes && (j == high || !dcList[j + 1])) {
- /* base = AFS_CHUNKTOBASE(dcList[first]->f.chunk); */
- base = AFS_CHUNKTOBASE(first + minj);
- /*
- *
- * take a list of dcache structs and send them all off to the server
- * the list must be in order, and the chunks contiguous.
- * Note - there is no locking done by this code currently. For
- * safety's sake, xdcache could be locked over the entire call.
- * However, that pretty well ties up all the threads. Meantime, all
- * the chunks _MUST_ have their refcounts bumped.
- * The writes done before a store back will clear setuid-ness
- * in cache file.
- * We can permit CacheStoreProc to wake up the user process IFF we
- * are doing the last RPC for this close, ie, storing back the last
- * set of contiguous chunks of a file.
- */
-
- dclist = &dcList[first];
- nchunks = 1 + j - first;
- nomore = !(moredata || (j != high));
- InStatus.ClientModTime = avc->f.m.Date;
- InStatus.Mask = AFS_SETMODTIME;
- if (sync & AFS_SYNC) {
- InStatus.Mask |= AFS_FSYNC;
- }
- tlen = lmin(avc->f.m.Length, avc->f.truncPos);
- afs_Trace4(afs_iclSetp, CM_TRACE_STOREDATA64,
- ICL_TYPE_FID, &avc->f.fid.Fid, ICL_TYPE_OFFSET,
- ICL_HANDLE_OFFSET(base), ICL_TYPE_OFFSET,
- ICL_HANDLE_OFFSET(bytes), ICL_TYPE_OFFSET,
- ICL_HANDLE_OFFSET(tlen));
-
- do {
- tc = afs_Conn(&avc->f.fid, areq, 0);
-#ifdef AFS_64BIT_CLIENT
- restart:
-#endif
- code = afs_CacheStoreProc(tc, dclist,
- avc, bytes, base, tlen,
- &newDV, &doProcessFS, &OutStatus,
- nchunks, &nomore);
-#ifdef AFS_64BIT_CLIENT
- if (code == RXGEN_OPCODE && !afs_serverHasNo64Bit(tc)) {
- afs_serverSetNo64Bit(tc);
- goto restart;
- }
-#endif /* AFS_64BIT_CLIENT */
- } while (afs_Analyze
- (tc, code, &avc->f.fid, areq,
- AFS_STATS_FS_RPCIDX_STOREDATA, SHARED_LOCK,
- NULL));
-
- /* put back all remaining locked dcache entries */
- for (i = 0; i < nchunks; i++) {
- tdc = dclist[i];
- if (!code) {
- if (afs_indexFlags[tdc->index] & IFDataMod) {
- /*
- * LOCKXXX -- should hold afs_xdcache(W) when
- * modifying afs_indexFlags.
- */
- afs_indexFlags[tdc->index] &= ~IFDataMod;
- afs_stats_cmperf.cacheCurrDirtyChunks--;
- afs_indexFlags[tdc->index] &= ~IFDirtyPages;
- if (sync & AFS_VMSYNC_INVAL) {
- /* since we have invalidated all the pages of this
- ** vnode by calling osi_VM_TryToSmush, we can
- ** safely mark this dcache entry as not having
- ** any pages. This vnode now becomes eligible for
- ** reclamation by getDownD.
- */
- afs_indexFlags[tdc->index] &= ~IFAnyPages;
- }
- }
- }
- UpgradeSToWLock(&tdc->lock, 628);
- tdc->f.states &= ~DWriting; /* correct? */
- tdc->dflags |= DFEntryMod;
- ReleaseWriteLock(&tdc->lock);
- afs_PutDCache(tdc);
- /* Mark the entry as released */
- dclist[i] = NULL;
- }
-
- if (doProcessFS) {
- /* Now copy out return params */
- UpgradeSToWLock(&avc->lock, 28); /* keep out others for a while */
- afs_ProcessFS(avc, &OutStatus, areq);
- /* Keep last (max) size of file on server to see if
- * we need to call afs_StoreMini to extend the file.
- */
- if (!moredata)
- maxStoredLength = OutStatus.Length;
- ConvertWToSLock(&avc->lock);
- doProcessFS = 0;
- }
-
- if (code) {
- for (j++; j <= high; j++) {
- if (dcList[j]) {
- ReleaseSharedLock(&(dcList[j]->lock));
- afs_PutDCache(dcList[j]);
- /* Releasing entry */
- dcList[j] = NULL;
- }
- }
- }
-
- afs_Trace2(afs_iclSetp, CM_TRACE_STOREALLDCDONE,
- ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, code);
- bytes = 0;
- }
- }
-
+ code =
+ afs_CacheStoreVCache(dcList, avc, areq, sync,
+ minj, high, moredata,
+ &newDV, &maxStoredLength);
/* Release any zero-length dcache entries in our interval
* that we locked but didn't store back above.
*/
minj += NCHUNKSATONCE;
} while (!code && moredata);
+ done:
UpgradeSToWLock(&avc->lock, 29);
/* send a trivial truncation store if did nothing else */
* Call StoreMini if we haven't written enough data to extend the
* file at the fileserver to the client's notion of the file length.
*/
- if ((avc->f.truncPos != AFS_NOTRUNC)
+ if ((avc->f.truncPos != AFS_NOTRUNC)
|| ((avc->f.states & CExtendedFile)
&& (maxStoredLength < avc->f.m.Length))) {
code = afs_StoreMini(avc, areq);
/*
* Finally, turn off DWriting, turn on DFEntryMod,
* update f.versionNo.
- * A lot of this could be integrated into the loop above
+ * A lot of this could be integrated into the loop above
*/
if (!code) {
afs_hyper_t h_unset;
do {
moredata = FALSE;
- memset((char *)dcList, 0,
+ memset(dcList, 0,
NCHUNKSATONCE * sizeof(struct dcache *));
/* overkill, but it gets the lock in case GetDSlot needs it */
- MObtainWriteLock(&afs_xdcache, 285);
+ ObtainWriteLock(&afs_xdcache, 285);
for (j = 0, safety = 0, index = afs_dvhashTbl[hash];
- index != NULLIDX && safety < afs_cacheFiles + 2;) {
+ index != NULLIDX && safety < afs_cacheFiles + 2;
+ index = afs_dvnextTbl[index]) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, 0);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ /* This is okay; since manipulating the dcaches at this
+ * point is best-effort. We only get a dcache here to
+ * increment the dv and turn off DWriting. If we were
+ * supposed to do that for a dcache, but could not
+ * due to an I/O error, it just means the dv won't
+ * be updated so we don't be able to use that cached
+ * chunk in the future. That's inefficient, but not
+ * an error. */
+ continue;
+ }
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid)
afs_PutDCache(tdc);
}
}
-
- index = afs_dvnextTbl[index];
}
- MReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&afs_xdcache);
for (i = 0; i < j; i++) {
/* Iterate over the dcache entries we collected above */
UpgradeSToWLock(&tdc->lock, 678);
hset(tdc->f.versionNo, avc->f.m.DataVersion);
tdc->dflags |= DFEntryMod;
+ /* DWriting may not have gotten cleared above, if all
+ * we did was a StoreMini */
+ tdc->f.states &= ~DWriting;
ConvertWToSLock(&tdc->lock);
}
}
* invalidated. Also discard data if it's a permanent error from the
* fileserver.
*/
- if (areq->permWriteError || (avc->f.states & (CCore1 | CCore))) {
+ if (areq->permWriteError || (avc->f.states & CCore)) {
afs_InvalidateAllSegments(avc);
}
}
/* Blow away pages; for now, only for Solaris */
#if (defined(AFS_SUN5_ENV))
if (WriteLocked(&avc->lock))
- osi_ReleaseVM(avc, (struct AFS_UCRED *)0);
+ osi_ReleaseVM(avc, (afs_ucred_t *)0);
#endif
/*
* Block out others from screwing with this table; is a read lock
* sufficient?
*/
- MObtainWriteLock(&afs_xdcache, 286);
+ ObtainWriteLock(&afs_xdcache, 286);
dcListMax = 0;
for (index = afs_dvhashTbl[hash]; index != NULLIDX;) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, 0);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ /* In the case of fatal errors during stores, we MUST
+ * invalidate all of the relevant chunks. Otherwise, the chunks
+ * will be left with the 'new' data that was never successfully
+ * written to the server, but the DV in the dcache is still the
+ * old DV. So, we may indefintely serve serve applications data
+ * that is not actually in the file on the fileserver. If we
+ * cannot afs_GetValidDSlot the appropriate entries, currently
+ * there is no way to ensure the dcache is invalidated. So for
+ * now, to avoid risking serving bad data from the cache, panic
+ * instead. */
+ osi_Panic("afs_InvalidateAllSegments tdc count");
+ }
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid))
dcListMax++;
for (index = afs_dvhashTbl[hash]; index != NULLIDX;) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, 0);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ /* We cannot proceed after getting this error; we risk serving
+ * incorrect data to applications. So panic instead. See the
+ * above comment next to the previous afs_GetValidDSlot call
+ * for details. */
+ osi_Panic("afs_InvalidateAllSegments tdc store");
+ }
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
/* same file? we'll zap it */
}
index = afs_dvnextTbl[index];
}
- MReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&afs_xdcache);
for (i = 0; i < dcListCount; i++) {
tdc = dcList[i];
return 0;
}
-/*!
- *
+/*!
+ *
* Extend a cache file
*
* \param avc pointer to vcache to extend data for
* \note avc must be write locked. May release and reobtain avc and GLOCK
*/
int
-afs_ExtendSegments(struct vcache *avc, afs_size_t alen, struct vrequest *areq) {
+afs_ExtendSegments(struct vcache *avc, afs_size_t alen, struct vrequest *areq)
+{
afs_size_t offset, toAdd;
struct osi_file *tfile;
afs_int32 code = 0;
struct dcache *tdc;
void *zeros;
- zeros = (void *) afs_osi_Alloc(AFS_PAGESIZE);
+ zeros = afs_osi_Alloc(AFS_PAGESIZE);
if (zeros == NULL)
return ENOMEM;
memset(zeros, 0, AFS_PAGESIZE);
towrite = (avc->f.m.Length + toAdd) - tdc->validPos;
if (towrite > AFS_PAGESIZE) towrite = AFS_PAGESIZE;
- code = afs_CFileWrite(tfile,
- tdc->validPos - AFS_CHUNKTOBASE(tdc->f.chunk),
+ code = afs_CFileWrite(tfile,
+ tdc->validPos - AFS_CHUNKTOBASE(tdc->f.chunk),
zeros, towrite);
tdc->validPos += towrite;
}
* held.
*/
int
-afs_TruncateAllSegments(register struct vcache *avc, afs_size_t alen,
- struct vrequest *areq, struct AFS_UCRED *acred)
+afs_TruncateAllSegments(struct vcache *avc, afs_size_t alen,
+ struct vrequest *areq, afs_ucred_t *acred)
{
- register struct dcache *tdc;
- register afs_int32 code;
- register afs_int32 index;
- afs_int32 newSize;
+ struct dcache *tdc;
+ afs_int32 code;
+ afs_int32 index;
+ afs_size_t newSize;
int dcCount, dcPos;
- struct dcache **tdcArray;
+ struct dcache **tdcArray = NULL;
AFS_STATCNT(afs_TruncateAllSegments);
avc->f.m.Date = osi_Time();
code = DVHash(&avc->f.fid);
/* block out others from screwing with this table */
- MObtainWriteLock(&afs_xdcache, 287);
+ ObtainWriteLock(&afs_xdcache, 287);
dcCount = 0;
for (index = afs_dvhashTbl[code]; index != NULLIDX;) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, 0);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ ReleaseWriteLock(&afs_xdcache);
+ code = EIO;
+ goto done;
+ }
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid))
dcCount++;
tdcArray = osi_Alloc(dcCount * sizeof(struct dcache *));
dcPos = 0;
- for (index = afs_dvhashTbl[code]; index != NULLIDX;) {
+ for (index = afs_dvhashTbl[code]; index != NULLIDX; index = afs_dvnextTbl[index]) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
- tdc = afs_GetDSlot(index, 0);
+ tdc = afs_GetValidDSlot(index);
+ if (!tdc) {
+ /* make sure we put back all of the tdcArray members before
+ * bailing out */
+ /* remember, the last valid tdc is at dcPos-1, so start at
+ * dcPos-1, not at dcPos itself. */
+ for (dcPos = dcPos - 1; dcPos >= 0; dcPos--) {
+ tdc = tdcArray[dcPos];
+ afs_PutDCache(tdc);
+ }
+ code = EIO;
+ goto done;
+ }
ReleaseReadLock(&tdc->tlock);
if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
/* same file, and modified, we'll store it back */
afs_PutDCache(tdc);
}
}
- index = afs_dvnextTbl[index];
}
- MReleaseWriteLock(&afs_xdcache);
+ ReleaseWriteLock(&afs_xdcache);
/* Now we loop over the array of dcache entries and truncate them */
for (index = 0; index < dcPos; index++) {
if (newSize < 0)
newSize = 0;
ObtainSharedLock(&tdc->lock, 672);
- if (newSize < tdc->f.chunkBytes) {
+ if (newSize < tdc->f.chunkBytes && newSize < MAX_AFS_UINT32) {
UpgradeSToWLock(&tdc->lock, 673);
+ tdc->f.states |= DWriting;
tfile = afs_CFileOpen(&tdc->f.inode);
- afs_CFileTruncate(tfile, newSize);
+ afs_CFileTruncate(tfile, (afs_int32)newSize);
afs_CFileClose(tfile);
- afs_AdjustSize(tdc, newSize);
+ afs_AdjustSize(tdc, (afs_int32)newSize);
if (alen < tdc->validPos) {
if (alen < AFS_CHUNKTOBASE(tdc->f.chunk))
tdc->validPos = 0;
afs_PutDCache(tdc);
}
- osi_Free(tdcArray, dcCount * sizeof(struct dcache *));
+ code = 0;
+ done:
+ if (tdcArray) {
+ osi_Free(tdcArray, dcCount * sizeof(struct dcache *));
+ }
#if (defined(AFS_SUN5_ENV))
ObtainWriteLock(&avc->vlock, 547);
if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
ReleaseWriteLock(&avc->vlock);
#endif
- return 0;
+ return code;
}
git://git.openafs.org / openafs.git / blobdiff