int
afs_StoreAllSegments(struct vcache *avc, struct vrequest *areq,
int sync)
{
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;
unsigned int i, j, minj, moredata, high, off;
afs_size_t maxStoredLength; /* highest offset we've written to server. */
int safety, marineronce = 0;
AFS_STATCNT(afs_StoreAllSegments);
hset(oldDV, avc->f.m.DataVersion);
hset(newDV, avc->f.m.DataVersion);
hash = DVHash(&avc->f.fid);
foreign = (avc->f.states & CForeign);
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)
/* In the aix vm implementation we need to do the vm_writep even
* 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_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
* system enough to cause a pageout, and this vnode would be
* locked when the pageout occurs. We can prevent this problem
* by making sure all dirty pages are already flushed. We don't
* do this when diskless because reading a diskless (i.e.
* memory-resident) chunk doesn't require using new VM, and we
* also don't want to dump more dirty data into a diskless cache,
* since they're smaller, and we might exceed its available
* space.
*/
#if defined(AFS_SUN5_ENV)
if (sync & AFS_VMSYNC_INVAL) /* invalidate VM pages */
osi_VM_TryToSmush(avc, CRED(), 1);
else
#endif
osi_VM_StoreAllSegments(avc);
}
if (AFS_IS_DISCONNECTED && !AFS_IN_SYNC) {
/* This will probably make someone sad ... */
/*printf("Net down in afs_StoreSegments\n");*/
return ENETDOWN;
}
ConvertWToSLock(&avc->lock);
/*
* Subsequent code expects a sorted list, and it expects all the
* chunks in the list to be contiguous, so we need a sort and a
* while loop in here, too - but this will work for a first pass...
* 92.10.05 - OK, there's a sort in here now. It's kind of a modified
* bin sort, I guess. Chunk numbers start with 0
*
* - 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
* 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
* are still on disk.
*/
origCBs = afs_allCBs;
maxStoredLength = 0;
minj = 0;
do {
memset(dcList, 0, NCHUNKSATONCE * sizeof(struct dcache *));
high = 0;
moredata = FALSE;
/* lock and start over from beginning of hash chain
* in order to avoid a race condition. */
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_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;
j++;
/* DCLOCKXXX: chunkBytes is protected by tdc->lock which we
* can't grab here, due to lock ordering with afs_xdcache.
* So, disable this shortcut for now. -- kolya 2001-10-13
*/
/* shortcut: big win for little files */
/* tlen -= tdc->f.chunkBytes;
* if (tlen <= 0)
* break;
*/
} else {
moredata = TRUE;
afs_PutDCache(tdc);
if (j == NCHUNKSATONCE)
break;
}
} else {
afs_PutDCache(tdc);
}
}
index = afs_dvnextTbl[index];
}
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) {
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.
*/
for (j = 0; j <= high; j++) {
tdc = dcList[j];
if (tdc) {
osi_Assert(tdc->f.chunkBytes == 0);
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
}
}
}
/* if (j) */
minj += NCHUNKSATONCE;
} while (!code && moredata);
done:
UpgradeSToWLock(&avc->lock, 29);
/* send a trivial truncation store if did nothing else */
if (code == 0) {
/*
* 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)
|| ((avc->f.states & CExtendedFile)
&& (maxStoredLength < avc->f.m.Length))) {
code = afs_StoreMini(avc, areq);
if (code == 0)
hadd32(newDV, 1); /* just bumped here, too */
}
avc->f.states &= ~CExtendedFile;
}
/*
* Finally, turn off DWriting, turn on DFEntryMod,
* update f.versionNo.
* A lot of this could be integrated into the loop above
*/
if (!code) {
afs_hyper_t h_unset;
hones(h_unset);
minj = 0;
do {
moredata = FALSE;
memset(dcList, 0,
NCHUNKSATONCE * sizeof(struct dcache *));
/* overkill, but it gets the lock in case GetDSlot needs it */
ObtainWriteLock(&afs_xdcache, 285);
for (j = 0, safety = 0, index = afs_dvhashTbl[hash];
index != NULLIDX && safety < afs_cacheFiles + 2;
index = afs_dvnextTbl[index]) {
if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
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)
&& tdc->f.chunk >= minj) {
off = tdc->f.chunk - minj;
if (off < NCHUNKSATONCE) {
/* this is the file, and the correct chunk range */
if (j >= NCHUNKSATONCE)
osi_Panic
("Too many dcache entries in range\n");
dcList[j++] = tdc;
} else {
moredata = TRUE;
afs_PutDCache(tdc);
if (j == NCHUNKSATONCE)
break;
}
} else {
afs_PutDCache(tdc);
}
}
}
ReleaseWriteLock(&afs_xdcache);
for (i = 0; i < j; i++) {
/* Iterate over the dcache entries we collected above */
tdc = dcList[i];
ObtainSharedLock(&tdc->lock, 677);
/* was code here to clear IFDataMod, but it should only be done
* in storedcache and storealldcache.
*/
/* Only increase DV if we had up-to-date data to start with.
* Otherwise, we could be falsely upgrading an old chunk
* (that we never read) into one labelled with the current
* DV #. Also note that we check that no intervening stores
* occurred, otherwise we might mislabel cache information
* for a chunk that we didn't store this time
*/
/* Don't update the version number if it's not yet set. */
if (!hsame(tdc->f.versionNo, h_unset)
&& hcmp(tdc->f.versionNo, oldDV) >= 0) {
if ((!(afs_dvhack || foreign)
&& hsame(avc->f.m.DataVersion, newDV))
|| ((afs_dvhack || foreign)
&& (origCBs == afs_allCBs))) {
/* no error, this is the DV */
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);
}
}
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
}
minj += NCHUNKSATONCE;
} while (moredata);
}
if (code) {
/*
* Invalidate chunks after an error for ccores files since
* afs_inactive won't be called for these and they won't be
* invalidated. Also discard data if it's a permanent error from the
* fileserver.
*/
if (areq->permWriteError || (avc->f.states & CCore)) {
afs_InvalidateAllSegments(avc);
}
}
afs_Trace3(afs_iclSetp, CM_TRACE_STOREALLDONE, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, avc->f.m.Length, ICL_TYPE_INT32, code);
/* would like a Trace5, but it doesn't exist... */
afs_Trace3(afs_iclSetp, CM_TRACE_AVCLOCKER, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, avc->lock.wait_states, ICL_TYPE_INT32,
avc->lock.excl_locked);
afs_Trace4(afs_iclSetp, CM_TRACE_AVCLOCKEE, ICL_TYPE_POINTER, avc,
ICL_TYPE_INT32, avc->lock.wait_states, ICL_TYPE_INT32,
avc->lock.readers_reading, ICL_TYPE_INT32,
avc->lock.num_waiting);
/*
* Finally, if updated DataVersion matches newDV, we did all of the
* stores. If mapDV indicates that the page cache was flushed up
* to when we started the store, then we can relabel them as flushed
* as recently as newDV.
* Turn off CDirty bit because the stored data is now in sync with server.
*/
if (code == 0 && hcmp(avc->mapDV, oldDV) >= 0) {
if ((!(afs_dvhack || foreign) && hsame(avc->f.m.DataVersion, newDV))
|| ((afs_dvhack || foreign) && (origCBs == afs_allCBs))) {
hset(avc->mapDV, newDV);
avc->f.states &= ~CDirty;
}
}
osi_FreeLargeSpace(dcList);
/* If not the final write a temporary error is ok. */
if (code && !areq->permWriteError && !(sync & AFS_LASTSTORE))
code = 0;
return code;
} /*afs_StoreAllSegments (new 03/02/94) */
/* question: does afs_create need to set CDirty in the adp or the avc?
* I think we can get away without it, but I'm not sure. Note that
* afs_setattr is called in here for truncation.
*/
#ifdef AFS_SGI64_ENV
int
afs_create(OSI_VC_DECL(adp), char *aname, struct vattr *attrs, int flags,
int amode, struct vcache **avcp, afs_ucred_t *acred)
#else /* AFS_SGI64_ENV */
int
afs_create(OSI_VC_DECL(adp), char *aname, struct vattr *attrs,
enum vcexcl aexcl, int amode, struct vcache **avcp,
afs_ucred_t *acred)
#endif /* AFS_SGI64_ENV */
{
afs_int32 origCBs, origZaps, finalZaps;
struct vrequest *treq = NULL;
afs_int32 code;
struct afs_conn *tc;
struct VenusFid newFid;
struct AFSStoreStatus InStatus;
struct AFSFetchStatus *OutFidStatus, *OutDirStatus;
struct AFSVolSync tsync;
struct AFSCallBack CallBack;
afs_int32 now;
struct dcache *tdc;
afs_size_t offset, len;
struct server *hostp = 0;
struct vcache *tvc;
struct volume *volp = 0;
struct afs_fakestat_state fakestate;
struct rx_connection *rxconn;
XSTATS_DECLS;
OSI_VC_CONVERT(adp);
AFS_STATCNT(afs_create);
OutFidStatus = osi_AllocSmallSpace(sizeof(struct AFSFetchStatus));
OutDirStatus = osi_AllocSmallSpace(sizeof(struct AFSFetchStatus));
memset(&InStatus, 0, sizeof(InStatus));
if ((code = afs_CreateReq(&treq, acred)))
goto done2;
afs_Trace3(afs_iclSetp, CM_TRACE_CREATE, ICL_TYPE_POINTER, adp,
ICL_TYPE_STRING, aname, ICL_TYPE_INT32, amode);
afs_InitFakeStat(&fakestate);
#ifdef AFS_SGI65_ENV
/* If avcp is passed not null, it's the old reference to this file.
* We can use this to avoid create races. For now, just decrement
* the reference count on it.
*/
if (*avcp) {
AFS_RELE(AFSTOV(*avcp));
*avcp = NULL;
}
#endif
if (strlen(aname) > AFSNAMEMAX) {
code = ENAMETOOLONG;
goto done3;
}
if (!afs_ENameOK(aname)) {
code = EINVAL;
goto done3;
}
switch (attrs->va_type) {
case VBLK:
case VCHR:
#if !defined(AFS_SUN5_ENV)
case VSOCK:
#endif
case VFIFO:
/* We don't support special devices or FIFOs */
code = EINVAL;
goto done3;
default:
;
}
AFS_DISCON_LOCK();
code = afs_EvalFakeStat(&adp, &fakestate, treq);
if (code)
goto done;
tagain:
code = afs_VerifyVCache(adp, treq);
if (code)
goto done;
/** If the volume is read-only, return error without making an RPC to the
* fileserver
*/
if (adp->f.states & CRO) {
code = EROFS;
goto done;
}
if (AFS_IS_DISCONNECTED && !AFS_IS_DISCON_RW) {
code = ENETDOWN;
goto done;
}
tdc = afs_GetDCache(adp, (afs_size_t) 0, treq, &offset, &len, 1);
ObtainWriteLock(&adp->lock, 135);
if (tdc)
ObtainSharedLock(&tdc->lock, 630);
/*
* Make sure that the data in the cache is current. We may have
* received a callback while we were waiting for the write lock.
*/
if (!(adp->f.states & CStatd)
|| (tdc && !hsame(adp->f.m.DataVersion, tdc->f.versionNo))) {
ReleaseWriteLock(&adp->lock);
if (tdc) {
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
}
goto tagain;
}
if (tdc) {
/* see if file already exists. If it does, we only set
* the size attributes (to handle O_TRUNC) */
code = afs_dir_Lookup(tdc, aname, &newFid.Fid); /* use dnlc first xxx */
if (code == 0) {
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
ReleaseWriteLock(&adp->lock);
#ifdef AFS_SGI64_ENV
if (flags & VEXCL) {
#else
if (aexcl != NONEXCL) {
#endif
code = EEXIST; /* file exists in excl mode open */
goto done;
}
/* found the file, so use it */
newFid.Cell = adp->f.fid.Cell;
newFid.Fid.Volume = adp->f.fid.Fid.Volume;
tvc = NULL;
if (newFid.Fid.Unique == 0) {
tvc = afs_LookupVCache(&newFid, treq, NULL, adp, aname);
}
if (!tvc) /* lookup failed or wasn't called */
tvc = afs_GetVCache(&newFid, treq, NULL, NULL);
if (tvc) {
/* if the thing exists, we need the right access to open it.
* we must check that here, since no other checks are
* made by the open system call */
len = attrs->va_size; /* only do the truncate */
/*
* We used to check always for READ access before; the
* problem is that we will fail if the existing file
* has mode -w-w-w, which is wrong.
*/
if ((amode & VREAD)
&& !afs_AccessOK(tvc, PRSFS_READ, treq, CHECK_MODE_BITS)) {
afs_PutVCache(tvc);
code = EACCES;
goto done;
}
#if defined(AFS_DARWIN80_ENV)
if ((amode & VWRITE) || VATTR_IS_ACTIVE(attrs, va_data_size))
#elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
if ((amode & VWRITE) || (attrs->va_mask & AT_SIZE))
#else
if ((amode & VWRITE) || len != 0xffffffff)
#endif
{
/* needed for write access check */
tvc->f.parent.vnode = adp->f.fid.Fid.Vnode;
tvc->f.parent.unique = adp->f.fid.Fid.Unique;
/* need write mode for these guys */
if (!afs_AccessOK
(tvc, PRSFS_WRITE, treq, CHECK_MODE_BITS)) {
afs_PutVCache(tvc);
code = EACCES;
goto done;
}
}
#if defined(AFS_DARWIN80_ENV)
if (VATTR_IS_ACTIVE(attrs, va_data_size))
#elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
if (attrs->va_mask & AT_SIZE)
#else
if (len != 0xffffffff)
#endif
{
if (vType(tvc) != VREG) {
afs_PutVCache(tvc);
code = EISDIR;
goto done;
}
/* do a truncate */
#if defined(AFS_DARWIN80_ENV)
VATTR_INIT(attrs);
VATTR_SET_SUPPORTED(attrs, va_data_size);
VATTR_SET_ACTIVE(attrs, va_data_size);
#elif defined(UKERNEL)
attrs->va_mask = ATTR_SIZE;
#elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
attrs->va_mask = AT_SIZE;
#else
VATTR_NULL(attrs);
#endif
attrs->va_size = len;
ObtainWriteLock(&tvc->lock, 136);
tvc->f.states |= CCreating;
ReleaseWriteLock(&tvc->lock);
#if defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
#if defined(AFS_SGI64_ENV)
code =
afs_setattr(VNODE_TO_FIRST_BHV((vnode_t *) tvc),
attrs, 0, acred);
#else
code = afs_setattr(tvc, attrs, 0, acred);
#endif /* AFS_SGI64_ENV */
#else /* SUN5 || SGI */
code = afs_setattr(tvc, attrs, acred);
#endif /* SUN5 || SGI */
ObtainWriteLock(&tvc->lock, 137);
tvc->f.states &= ~CCreating;
ReleaseWriteLock(&tvc->lock);
if (code) {
afs_PutVCache(tvc);
goto done;
}
}
*avcp = tvc;
} else
code = ENOENT; /* shouldn't get here */
/* make sure vrefCount bumped only if code == 0 */
goto done;
}
}
/* if we create the file, we don't do any access checks, since
* that's how O_CREAT is supposed to work */
if (adp->f.states & CForeign) {
origCBs = afs_allCBs;
origZaps = afs_allZaps;
} else {
origCBs = afs_evenCBs; /* if changes, we don't really have a callback */
origZaps = afs_evenZaps; /* number of even numbered vnodes discarded */
}
InStatus.Mask = AFS_SETMODTIME | AFS_SETMODE | AFS_SETGROUP;
InStatus.ClientModTime = osi_Time();
InStatus.Group = (afs_int32) afs_cr_gid(acred);
if (AFS_NFSXLATORREQ(acred)) {
/*
* XXX The following is mainly used to fix a bug in the HP-UX
* nfs client where they create files with mode of 0 without
* doing any setattr later on to fix it. * XXX
*/
#if defined(AFS_AIX_ENV)
if (attrs->va_mode != -1) {
#else
#if defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
if (attrs->va_mask & AT_MODE) {
#else
if (attrs->va_mode != ((unsigned short)-1)) {
#endif
#endif
if (!attrs->va_mode)
attrs->va_mode = 0x1b6; /* XXX default mode: rw-rw-rw XXX */
}
}
if (!AFS_IS_DISCONNECTED) {
/* If not disconnected, connect to the server.*/
InStatus.UnixModeBits = attrs->va_mode & 0xffff; /* only care about protection bits */
do {
tc = afs_Conn(&adp->f.fid, treq, SHARED_LOCK, &rxconn);
if (tc) {
hostp = tc->srvr->server; /* remember for callback processing */
now = osi_Time();
XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_CREATEFILE);
RX_AFS_GUNLOCK();
code =
RXAFS_CreateFile(rxconn, (struct AFSFid *)&adp->f.fid.Fid,
aname, &InStatus, (struct AFSFid *)
&newFid.Fid, OutFidStatus, OutDirStatus,
&CallBack, &tsync);
RX_AFS_GLOCK();
XSTATS_END_TIME;
CallBack.ExpirationTime += now;
} else
code = -1;
} while (afs_Analyze
(tc, rxconn, code, &adp->f.fid, treq, AFS_STATS_FS_RPCIDX_CREATEFILE,
SHARED_LOCK, NULL));
if ((code == EEXIST || code == UAEEXIST) &&
#ifdef AFS_SGI64_ENV
!(flags & VEXCL)
#else /* AFS_SGI64_ENV */
aexcl == NONEXCL
#endif
) {
/* if we get an EEXIST in nonexcl mode, just do a lookup */
if (tdc) {
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
}
ReleaseWriteLock(&adp->lock);
#if defined(AFS_SGI64_ENV)
code = afs_lookup(VNODE_TO_FIRST_BHV((vnode_t *) adp), aname, avcp,
NULL, 0, NULL, acred);
#elif defined(AFS_SUN5_ENV) || defined(AFS_SGI_ENV)
code = afs_lookup(adp, aname, avcp, NULL, 0, NULL, acred);
#elif defined(UKERNEL)
code = afs_lookup(adp, aname, avcp, acred, 0);
#elif !defined(AFS_DARWIN_ENV)
code = afs_lookup(adp, aname, avcp, acred);
#endif
goto done;
}
if (code) {
if (code < 0) {
ObtainWriteLock(&afs_xcbhash, 488);
afs_DequeueCallback(adp);
adp->f.states &= ~CStatd;
ReleaseWriteLock(&afs_xcbhash);
osi_dnlc_purgedp(adp);
}
ReleaseWriteLock(&adp->lock);
if (tdc) {
ReleaseSharedLock(&tdc->lock);
afs_PutDCache(tdc);
}
goto done;
}
} else {
/* Generate a fake FID for disconnected mode. */
newFid.Cell = adp->f.fid.Cell;
newFid.Fid.Volume = adp->f.fid.Fid.Volume;
afs_GenFakeFid(&newFid, VREG, 1);
} /* if (!AFS_IS_DISCON_RW) */
/* otherwise, we should see if we can make the change to the dir locally */
if (tdc)
UpgradeSToWLock(&tdc->lock, 631);
if (AFS_IS_DISCON_RW || afs_LocalHero(adp, tdc, OutDirStatus, 1)) {
/* we can do it locally */
ObtainWriteLock(&afs_xdcache, 291);
code = afs_dir_Create(tdc, aname, &newFid.Fid);
ReleaseWriteLock(&afs_xdcache);
if (code) {
ZapDCE(tdc);
DZap(tdc);
}
}
if (tdc) {
ReleaseWriteLock(&tdc->lock);
afs_PutDCache(tdc);
}
if (AFS_IS_DISCON_RW)
adp->f.m.LinkCount++;
newFid.Cell = adp->f.fid.Cell;
newFid.Fid.Volume = adp->f.fid.Fid.Volume;
ReleaseWriteLock(&adp->lock);
volp = afs_FindVolume(&newFid, READ_LOCK);
/* New tricky optimistic callback handling algorithm for file creation works
* as follows. We create the file essentially with no locks set at all. File
* server may thus handle operations from others cache managers as well as from
* this very own cache manager that reference the file in question before
* we managed to create the cache entry. However, if anyone else changes
* any of the status information for a file, we'll see afs_evenCBs increase
* (files always have even fids). If someone on this workstation manages
* to do something to the file, they'll end up having to create a cache
* entry for the new file. Either we'll find it once we've got the afs_xvcache
* lock set, or it was also *deleted* the vnode before we got there, in which case
* we will find evenZaps has changed, too. Thus, we only assume we have the right
* status information if no callbacks or vnode removals have occurred to even
* numbered files from the time the call started until the time that we got the xvcache
* lock set. Of course, this also assumes that any call that modifies a file first
* gets a write lock on the file's vnode, but if that weren't true, the whole cache manager
* would fail, since no call would be able to update the local vnode status after modifying
* a file on a file server. */
ObtainWriteLock(&afs_xvcache, 138);
if (adp->f.states & CForeign)
finalZaps = afs_allZaps; /* do this before calling newvcache */
else
finalZaps = afs_evenZaps; /* do this before calling newvcache */
/* don't need to call RemoveVCB, since only path leaving a callback is the
* one where we pass through afs_NewVCache. Can't have queued a VCB unless
* we created and freed an entry between file creation time and here, and the
* freeing of the vnode will change evenZaps. Don't need to update the VLRU
* queue, since the find will only succeed in the event of a create race, and
* then the vcache will be at the front of the VLRU queue anyway... */
if (!(tvc = afs_FindVCache(&newFid, 0, DO_STATS))) {
tvc = afs_NewVCache(&newFid, hostp);
if (tvc) {
int finalCBs;
ObtainWriteLock(&tvc->lock, 139);
ObtainWriteLock(&afs_xcbhash, 489);
finalCBs = afs_evenCBs;
/* add the callback in */
if (adp->f.states & CForeign) {
tvc->f.states |= CForeign;
finalCBs = afs_allCBs;
}
if (origCBs == finalCBs && origZaps == finalZaps) {
tvc->f.states |= CStatd; /* we've fake entire thing, so don't stat */
tvc->f.states &= ~CBulkFetching;
if (!AFS_IS_DISCON_RW) {
tvc->cbExpires = CallBack.ExpirationTime;
afs_QueueCallback(tvc, CBHash(CallBack.ExpirationTime), volp);
}
} else {
afs_DequeueCallback(tvc);
tvc->f.states &= ~(CStatd | CUnique);
tvc->callback = 0;
if (tvc->f.fid.Fid.Vnode & 1 || (vType(tvc) == VDIR))
osi_dnlc_purgedp(tvc);
}
ReleaseWriteLock(&afs_xcbhash);
if (AFS_IS_DISCON_RW) {
afs_DisconAddDirty(tvc, VDisconCreate, 0);
afs_GenDisconStatus(adp, tvc, &newFid, attrs, treq, VREG);
} else {
afs_ProcessFS(tvc, OutFidStatus, treq);
}
tvc->f.parent.vnode = adp->f.fid.Fid.Vnode;
tvc->f.parent.unique = adp->f.fid.Fid.Unique;
#if !defined(UKERNEL)
if (volp && (volp->states & VPartVisible))
tvc->f.states |= CPartVisible;
#endif
ReleaseWriteLock(&tvc->lock);
*avcp = tvc;
code = 0;
} else
code = ENOENT;
} else {
/* otherwise cache entry already exists, someone else must
* have created it. Comments used to say: "don't need write
* lock to *clear* these flags" but we should do it anyway.
* Code used to clear stat bit and callback, but I don't see
* the point -- we didn't have a create race, somebody else just
* snuck into NewVCache before we got here, probably a racing
* lookup.
*/
*avcp = tvc;
code = 0;
}
ReleaseWriteLock(&afs_xvcache);
done:
AFS_DISCON_UNLOCK();
done3:
if (volp)
afs_PutVolume(volp, READ_LOCK);
if (code == 0) {
if (afs_mariner)
afs_AddMarinerName(aname, *avcp);
/* return the new status in vattr */
afs_CopyOutAttrs(*avcp, attrs);
if (afs_mariner)
afs_MarinerLog("store$Creating", *avcp);
}
afs_PutFakeStat(&fakestate);
code = afs_CheckCode(code, treq, 20);
afs_DestroyReq(treq);
done2:
osi_FreeSmallSpace(OutFidStatus);
osi_FreeSmallSpace(OutDirStatus);
return code;
}
/*
* Check to see if we can track the change locally: requires that
* we have sufficiently recent info in data cache. If so, we
* know the new DataVersion number, and place it correctly in both the
* data and stat cache entries. This routine returns 1 if we should
* do the operation locally, and 0 otherwise.
*
* This routine must be called with the stat cache entry write-locked,
* and dcache entry write-locked.
*/
int
afs_LocalHero(struct vcache *avc, struct dcache *adc,
AFSFetchStatus * astat, int aincr)
{
afs_int32 ok;
afs_hyper_t avers;
AFS_STATCNT(afs_LocalHero);
hset64(avers, astat->dataVersionHigh, astat->DataVersion);
/* avers *is* the version number now, no matter what */
if (adc) {
/* does what's in the dcache *now* match what's in the vcache *now*,
* and do we have a valid callback? if not, our local copy is not "ok" */
ok = (hsame(avc->f.m.DataVersion, adc->f.versionNo) && avc->callback
&& (avc->f.states & CStatd) && avc->cbExpires >= osi_Time());
} else {
ok = 0;
}
if (ok) {
/* check that the DV on the server is what we expect it to be */
afs_hyper_t newDV;
hset(newDV, adc->f.versionNo);
hadd32(newDV, aincr);
if (!hsame(avers, newDV)) {
ok = 0;
}
}
#if defined(AFS_SGI_ENV)
osi_Assert(avc->v.v_type == VDIR);
#endif
/* The bulk status code used the length as a sequence number. */
/* Don't update the vcache entry unless the stats are current. */
if (avc->f.states & CStatd) {
hset(avc->f.m.DataVersion, avers);
#ifdef AFS_64BIT_CLIENT
FillInt64(avc->f.m.Length, astat->Length_hi, astat->Length);
#else /* AFS_64BIT_CLIENT */
avc->f.m.Length = astat->Length;
#endif /* AFS_64BIT_CLIENT */
avc->f.m.Date = astat->ClientModTime;
}
if (ok) {
/* we've been tracking things correctly */
adc->dflags |= DFEntryMod;
adc->f.versionNo = avers;
return 1;
} else {
if (adc) {
ZapDCE(adc);
DZap(adc);
}
if (avc->f.states & CStatd) {
osi_dnlc_purgedp(avc);
}
return 0;
}
}
/* rxi_ReadProc -- internal version.
*
* LOCKS USED -- called at netpri
*/
int
rxi_ReadProc(struct rx_call *call, char *buf,
int nbytes)
{
struct rx_packet *cp = call->currentPacket;
struct rx_packet *rp;
int requestCount;
unsigned int t;
/* XXXX took out clock_NewTime from here. Was it needed? */
requestCount = nbytes;
/* Free any packets from the last call to ReadvProc/WritevProc */
if (queue_IsNotEmpty(&call->iovq)) {
#ifdef RXDEBUG_PACKET
call->iovqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &call->iovq);
}
do {
if (call->nLeft == 0) {
/* Get next packet */
MUTEX_ENTER(&call->lock);
for (;;) {
if (call->error || (call->mode != RX_MODE_RECEIVING)) {
if (call->error) {
call->mode = RX_MODE_ERROR;
MUTEX_EXIT(&call->lock);
return 0;
}
if (call->mode == RX_MODE_SENDING) {
MUTEX_EXIT(&call->lock);
rxi_FlushWrite(call);
MUTEX_ENTER(&call->lock);
continue;
}
}
if (queue_IsNotEmpty(&call->rq)) {
/* Check that next packet available is next in sequence */
rp = queue_First(&call->rq, rx_packet);
if (rp->header.seq == call->rnext) {
afs_int32 error;
struct rx_connection *conn = call->conn;
queue_Remove(rp);
#ifdef RX_TRACK_PACKETS
rp->flags &= ~RX_PKTFLAG_RQ;
#endif
#ifdef RXDEBUG_PACKET
call->rqc--;
#endif /* RXDEBUG_PACKET */
/* RXS_CheckPacket called to undo RXS_PreparePacket's
* work. It may reduce the length of the packet by up
* to conn->maxTrailerSize, to reflect the length of the
* data + the header. */
if ((error =
RXS_CheckPacket(conn->securityObject, call,
rp))) {
/* Used to merely shut down the call, but now we
* shut down the whole connection since this may
* indicate an attempt to hijack it */
MUTEX_EXIT(&call->lock);
rxi_ConnectionError(conn, error);
MUTEX_ENTER(&conn->conn_data_lock);
rp = rxi_SendConnectionAbort(conn, rp, 0, 0);
MUTEX_EXIT(&conn->conn_data_lock);
rxi_FreePacket(rp);
return 0;
}
call->rnext++;
cp = call->currentPacket = rp;
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags |= RX_PKTFLAG_CP;
#endif
call->curvec = 1; /* 0th vec is always header */
/* begin at the beginning [ more or less ], continue
* on until the end, then stop. */
call->curpos =
(char *)cp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
call->curlen =
cp->wirevec[1].iov_len -
call->conn->securityHeaderSize;
/* Notice that this code works correctly if the data
* size is 0 (which it may be--no reply arguments from
* server, for example). This relies heavily on the
* fact that the code below immediately frees the packet
* (no yields, etc.). If it didn't, this would be a
* problem because a value of zero for call->nLeft
* normally means that there is no read packet */
call->nLeft = cp->length;
hadd32(call->bytesRcvd, cp->length);
/* Send a hard ack for every rxi_HardAckRate+1 packets
* consumed. Otherwise schedule an event to send
* the hard ack later on.
*/
call->nHardAcks++;
if (!(call->flags & RX_CALL_RECEIVE_DONE)) {
if (call->nHardAcks > (u_short) rxi_HardAckRate) {
rxevent_Cancel(call->delayedAckEvent, call,
RX_CALL_REFCOUNT_DELAY);
rxi_SendAck(call, 0, 0, RX_ACK_DELAY, 0);
} else {
struct clock when, now;
clock_GetTime(&now);
when = now;
/* Delay to consolidate ack packets */
clock_Add(&when, &rx_hardAckDelay);
if (!call->delayedAckEvent
|| clock_Gt(&call->delayedAckEvent->
eventTime, &when)) {
rxevent_Cancel(call->delayedAckEvent,
call,
RX_CALL_REFCOUNT_DELAY);
MUTEX_ENTER(&rx_refcnt_mutex);
CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
MUTEX_EXIT(&rx_refcnt_mutex);
call->delayedAckEvent =
rxevent_PostNow(&when, &now,
rxi_SendDelayedAck, call,
0);
}
}
}
break;
}
}
/*
* If we reach this point either we have no packets in the
* receive queue or the next packet in the queue is not the
* one we are looking for. There is nothing else for us to
* do but wait for another packet to arrive.
*/
/* Are there ever going to be any more packets? */
if (call->flags & RX_CALL_RECEIVE_DONE) {
MUTEX_EXIT(&call->lock);
return requestCount - nbytes;
}
/* Wait for in-sequence packet */
call->flags |= RX_CALL_READER_WAIT;
clock_NewTime();
call->startWait = clock_Sec();
while (call->flags & RX_CALL_READER_WAIT) {
#ifdef RX_ENABLE_LOCKS
CV_WAIT(&call->cv_rq, &call->lock);
#else
osi_rxSleep(&call->rq);
#endif
}
cp = call->currentPacket;
call->startWait = 0;
#ifdef RX_ENABLE_LOCKS
if (call->error) {
MUTEX_EXIT(&call->lock);
return 0;
}
#endif /* RX_ENABLE_LOCKS */
}
MUTEX_EXIT(&call->lock);
} else
/* osi_Assert(cp); */
/* MTUXXX this should be replaced by some error-recovery code before shipping */
/* yes, the following block is allowed to be the ELSE clause (or not) */
/* It's possible for call->nLeft to be smaller than any particular
* iov_len. Usually, recvmsg doesn't change the iov_len, since it
* reflects the size of the buffer. We have to keep track of the
* number of bytes read in the length field of the packet struct. On
* the final portion of a received packet, it's almost certain that
* call->nLeft will be smaller than the final buffer. */
while (nbytes && cp) {
t = MIN((int)call->curlen, nbytes);
t = MIN(t, (int)call->nLeft);
memcpy(buf, call->curpos, t);
buf += t;
nbytes -= t;
call->curpos += t;
call->curlen -= t;
call->nLeft -= t;
if (!call->nLeft) {
/* out of packet. Get another one. */
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
} else if (!call->curlen) {
/* need to get another struct iov */
if (++call->curvec >= cp->niovecs) {
/* current packet is exhausted, get ready for another */
/* don't worry about curvec and stuff, they get set somewhere else */
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
call->nLeft = 0;
} else {
call->curpos =
(char *)cp->wirevec[call->curvec].iov_base;
call->curlen = cp->wirevec[call->curvec].iov_len;
}
}
}
if (!nbytes) {
/* user buffer is full, return */
return requestCount;
}
} while (nbytes);
return requestCount;
}
int
rxi_WriteProc(struct rx_call *call, char *buf,
int nbytes)
{
struct rx_connection *conn = call->conn;
struct rx_packet *cp = call->currentPacket;
unsigned int t;
int requestCount = nbytes;
/* Free any packets from the last call to ReadvProc/WritevProc */
if (queue_IsNotEmpty(&call->iovq)) {
#ifdef RXDEBUG_PACKET
call->iovqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &call->iovq);
}
if (call->mode != RX_MODE_SENDING) {
if ((conn->type == RX_SERVER_CONNECTION)
&& (call->mode == RX_MODE_RECEIVING)) {
call->mode = RX_MODE_SENDING;
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
call->nLeft = 0;
call->nFree = 0;
}
} else {
return 0;
}
}
/* Loop condition is checked at end, so that a write of 0 bytes
* will force a packet to be created--specially for the case where
* there are 0 bytes on the stream, but we must send a packet
* anyway. */
do {
if (call->nFree == 0) {
MUTEX_ENTER(&call->lock);
#ifdef AFS_GLOBAL_RXLOCK_KERNEL
rxi_WaitforTQBusy(call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
cp = call->currentPacket;
if (call->error)
call->mode = RX_MODE_ERROR;
if (!call->error && cp) {
/* Clear the current packet now so that if
* we are forced to wait and drop the lock
* the packet we are planning on using
* cannot be freed.
*/
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
call->currentPacket = (struct rx_packet *)0;
clock_NewTime(); /* Bogus: need new time package */
/* The 0, below, specifies that it is not the last packet:
* there will be others. PrepareSendPacket may
* alter the packet length by up to
* conn->securityMaxTrailerSize */
hadd32(call->bytesSent, cp->length);
rxi_PrepareSendPacket(call, cp, 0);
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_TQ;
#endif
queue_Append(&call->tq, cp);
#ifdef RXDEBUG_PACKET
call->tqc++;
#endif /* RXDEBUG_PACKET */
cp = (struct rx_packet *)0;
if (!
(call->
flags & (RX_CALL_FAST_RECOVER |
RX_CALL_FAST_RECOVER_WAIT))) {
rxi_Start(0, call, 0, 0);
}
} else if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
cp = call->currentPacket = (struct rx_packet *)0;
}
/* Wait for transmit window to open up */
while (!call->error
&& call->tnext + 1 > call->tfirst + (2 * call->twind)) {
clock_NewTime();
call->startWait = clock_Sec();
#ifdef RX_ENABLE_LOCKS
CV_WAIT(&call->cv_twind, &call->lock);
#else
call->flags |= RX_CALL_WAIT_WINDOW_ALLOC;
osi_rxSleep(&call->twind);
#endif
call->startWait = 0;
#ifdef RX_ENABLE_LOCKS
if (call->error) {
call->mode = RX_MODE_ERROR;
MUTEX_EXIT(&call->lock);
return 0;
}
#endif /* RX_ENABLE_LOCKS */
}
if ((cp = rxi_AllocSendPacket(call, nbytes))) {
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_CP;
#endif
call->currentPacket = cp;
call->nFree = cp->length;
call->curvec = 1; /* 0th vec is always header */
/* begin at the beginning [ more or less ], continue
* on until the end, then stop. */
call->curpos =
(char *)cp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
call->curlen =
cp->wirevec[1].iov_len - call->conn->securityHeaderSize;
}
if (call->error) {
call->mode = RX_MODE_ERROR;
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
call->currentPacket = NULL;
}
MUTEX_EXIT(&call->lock);
return 0;
}
MUTEX_EXIT(&call->lock);
}
if (cp && (int)call->nFree < nbytes) {
/* Try to extend the current buffer */
int len, mud;
len = cp->length;
mud = rx_MaxUserDataSize(call);
if (mud > len) {
int want;
want = MIN(nbytes - (int)call->nFree, mud - len);
rxi_AllocDataBuf(cp, want, RX_PACKET_CLASS_SEND_CBUF);
if (cp->length > (unsigned)mud)
cp->length = mud;
call->nFree += (cp->length - len);
}
}
/* If the remaining bytes fit in the buffer, then store them
* and return. Don't ship a buffer that's full immediately to
* the peer--we don't know if it's the last buffer yet */
if (!cp) {
call->nFree = 0;
}
while (nbytes && call->nFree) {
t = MIN((int)call->curlen, nbytes);
t = MIN((int)call->nFree, t);
memcpy(call->curpos, buf, t);
buf += t;
nbytes -= t;
call->curpos += t;
call->curlen -= (u_short)t;
call->nFree -= (u_short)t;
if (!call->curlen) {
/* need to get another struct iov */
if (++call->curvec >= cp->niovecs) {
/* current packet is full, extend or send it */
call->nFree = 0;
} else {
call->curpos = (char *)cp->wirevec[call->curvec].iov_base;
call->curlen = cp->wirevec[call->curvec].iov_len;
}
}
} /* while bytes to send and room to send them */
/* might be out of space now */
if (!nbytes) {
return requestCount;
} else; /* more data to send, so get another packet and keep going */
} while (nbytes);
return requestCount - nbytes;
}
/* rxi_FillReadVec
*
* Uses packets in the receive queue to fill in as much of the
* current iovec as possible. Does not block if it runs out
* of packets to complete the iovec. Return true if an ack packet
* was sent, otherwise return false */
int
rxi_FillReadVec(struct rx_call *call, afs_uint32 serial)
{
int didConsume = 0;
int didHardAck = 0;
unsigned int t;
struct rx_packet *rp;
struct rx_packet *curp;
struct iovec *call_iov;
struct iovec *cur_iov = NULL;
curp = call->currentPacket;
if (curp) {
cur_iov = &curp->wirevec[call->curvec];
}
call_iov = &call->iov[call->iovNext];
while (!call->error && call->iovNBytes && call->iovNext < call->iovMax) {
if (call->nLeft == 0) {
/* Get next packet */
if (queue_IsNotEmpty(&call->rq)) {
/* Check that next packet available is next in sequence */
rp = queue_First(&call->rq, rx_packet);
if (rp->header.seq == call->rnext) {
afs_int32 error;
struct rx_connection *conn = call->conn;
queue_Remove(rp);
#ifdef RX_TRACK_PACKETS
rp->flags &= ~RX_PKTFLAG_RQ;
#endif
#ifdef RXDEBUG_PACKET
call->rqc--;
#endif /* RXDEBUG_PACKET */
/* RXS_CheckPacket called to undo RXS_PreparePacket's
* work. It may reduce the length of the packet by up
* to conn->maxTrailerSize, to reflect the length of the
* data + the header. */
if ((error =
RXS_CheckPacket(conn->securityObject, call, rp))) {
/* Used to merely shut down the call, but now we
* shut down the whole connection since this may
* indicate an attempt to hijack it */
MUTEX_EXIT(&call->lock);
rxi_ConnectionError(conn, error);
MUTEX_ENTER(&conn->conn_data_lock);
rp = rxi_SendConnectionAbort(conn, rp, 0, 0);
MUTEX_EXIT(&conn->conn_data_lock);
rxi_FreePacket(rp);
MUTEX_ENTER(&call->lock);
return 1;
}
call->rnext++;
curp = call->currentPacket = rp;
#ifdef RX_TRACK_PACKETS
call->currentPacket->flags |= RX_PKTFLAG_CP;
#endif
call->curvec = 1; /* 0th vec is always header */
cur_iov = &curp->wirevec[1];
/* begin at the beginning [ more or less ], continue
* on until the end, then stop. */
call->curpos =
(char *)curp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
call->curlen =
curp->wirevec[1].iov_len -
call->conn->securityHeaderSize;
/* Notice that this code works correctly if the data
* size is 0 (which it may be--no reply arguments from
* server, for example). This relies heavily on the
* fact that the code below immediately frees the packet
* (no yields, etc.). If it didn't, this would be a
* problem because a value of zero for call->nLeft
* normally means that there is no read packet */
call->nLeft = curp->length;
hadd32(call->bytesRcvd, curp->length);
/* Send a hard ack for every rxi_HardAckRate+1 packets
* consumed. Otherwise schedule an event to send
* the hard ack later on.
*/
call->nHardAcks++;
didConsume = 1;
continue;
}
}
break;
}
/* It's possible for call->nLeft to be smaller than any particular
* iov_len. Usually, recvmsg doesn't change the iov_len, since it
* reflects the size of the buffer. We have to keep track of the
* number of bytes read in the length field of the packet struct. On
* the final portion of a received packet, it's almost certain that
* call->nLeft will be smaller than the final buffer. */
while (call->iovNBytes && call->iovNext < call->iovMax && curp) {
t = MIN((int)call->curlen, call->iovNBytes);
t = MIN(t, (int)call->nLeft);
call_iov->iov_base = call->curpos;
call_iov->iov_len = t;
call_iov++;
call->iovNext++;
call->iovNBytes -= t;
call->curpos += t;
call->curlen -= t;
call->nLeft -= t;
if (!call->nLeft) {
/* out of packet. Get another one. */
#ifdef RX_TRACK_PACKETS
curp->flags &= ~RX_PKTFLAG_CP;
curp->flags |= RX_PKTFLAG_IOVQ;
#endif
queue_Append(&call->iovq, curp);
#ifdef RXDEBUG_PACKET
call->iovqc++;
#endif /* RXDEBUG_PACKET */
curp = call->currentPacket = (struct rx_packet *)0;
} else if (!call->curlen) {
/* need to get another struct iov */
if (++call->curvec >= curp->niovecs) {
/* current packet is exhausted, get ready for another */
/* don't worry about curvec and stuff, they get set somewhere else */
#ifdef RX_TRACK_PACKETS
curp->flags &= ~RX_PKTFLAG_CP;
curp->flags |= RX_PKTFLAG_IOVQ;
#endif
queue_Append(&call->iovq, curp);
#ifdef RXDEBUG_PACKET
call->iovqc++;
#endif /* RXDEBUG_PACKET */
curp = call->currentPacket = (struct rx_packet *)0;
call->nLeft = 0;
} else {
cur_iov++;
call->curpos = (char *)cur_iov->iov_base;
call->curlen = cur_iov->iov_len;
}
}
}
}
/* If we consumed any packets then check whether we need to
* send a hard ack. */
if (didConsume && (!(call->flags & RX_CALL_RECEIVE_DONE))) {
if (call->nHardAcks > (u_short) rxi_HardAckRate) {
rxevent_Cancel(call->delayedAckEvent, call,
RX_CALL_REFCOUNT_DELAY);
rxi_SendAck(call, 0, serial, RX_ACK_DELAY, 0);
didHardAck = 1;
} else {
struct clock when, now;
clock_GetTime(&now);
when = now;
/* Delay to consolidate ack packets */
clock_Add(&when, &rx_hardAckDelay);
if (!call->delayedAckEvent
|| clock_Gt(&call->delayedAckEvent->eventTime, &when)) {
rxevent_Cancel(call->delayedAckEvent, call,
RX_CALL_REFCOUNT_DELAY);
MUTEX_ENTER(&rx_refcnt_mutex);
CALL_HOLD(call, RX_CALL_REFCOUNT_DELAY);
MUTEX_EXIT(&rx_refcnt_mutex);
call->delayedAckEvent =
rxevent_PostNow(&when, &now, rxi_SendDelayedAck, call, 0);
}
}
}
return didHardAck;
}
/* Flush any buffered data to the stream, switch to read mode
* (clients) or to EOF mode (servers)
*
* LOCKS HELD: called at netpri.
*/
void
rxi_FlushWrite(struct rx_call *call)
{
struct rx_packet *cp = NULL;
/* Free any packets from the last call to ReadvProc/WritevProc */
if (queue_IsNotEmpty(&call->iovq)) {
#ifdef RXDEBUG_PACKET
call->iovqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &call->iovq);
}
if (call->mode == RX_MODE_SENDING) {
call->mode =
(call->conn->type ==
RX_CLIENT_CONNECTION ? RX_MODE_RECEIVING : RX_MODE_EOF);
#ifdef RX_KERNEL_TRACE
{
int glockOwner = ISAFS_GLOCK();
if (!glockOwner)
AFS_GLOCK();
afs_Trace3(afs_iclSetp, CM_TRACE_WASHERE, ICL_TYPE_STRING,
__FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
call);
if (!glockOwner)
AFS_GUNLOCK();
}
#endif
MUTEX_ENTER(&call->lock);
#ifdef AFS_GLOBAL_RXLOCK_KERNEL
rxi_WaitforTQBusy(call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
if (call->error)
call->mode = RX_MODE_ERROR;
cp = call->currentPacket;
if (cp) {
/* cp->length is only supposed to be the user's data */
/* cp->length was already set to (then-current)
* MaxUserDataSize or less. */
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
cp->length -= call->nFree;
call->currentPacket = (struct rx_packet *)0;
call->nFree = 0;
} else {
cp = rxi_AllocSendPacket(call, 0);
if (!cp) {
/* Mode can no longer be MODE_SENDING */
return;
}
cp->length = 0;
cp->niovecs = 2; /* header + space for rxkad stuff */
call->nFree = 0;
}
/* The 1 specifies that this is the last packet */
hadd32(call->bytesSent, cp->length);
rxi_PrepareSendPacket(call, cp, 1);
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_TQ;
#endif
queue_Append(&call->tq, cp);
#ifdef RXDEBUG_PACKET
call->tqc++;
#endif /* RXDEBUG_PACKET */
if (!
(call->
flags & (RX_CALL_FAST_RECOVER | RX_CALL_FAST_RECOVER_WAIT))) {
rxi_Start(0, call, 0, 0);
}
MUTEX_EXIT(&call->lock);
}
}
/* rxi_WritevProc -- internal version.
*
* Send buffers allocated in rxi_WritevAlloc.
*
* LOCKS USED -- called at netpri.
*/
int
rxi_WritevProc(struct rx_call *call, struct iovec *iov, int nio, int nbytes)
{
struct rx_packet *cp = NULL;
#ifdef RX_TRACK_PACKETS
struct rx_packet *p, *np;
#endif
int nextio;
int requestCount;
struct rx_queue tmpq;
#ifdef RXDEBUG_PACKET
u_short tmpqc;
#endif
requestCount = nbytes;
nextio = 0;
MUTEX_ENTER(&call->lock);
if (call->error) {
call->mode = RX_MODE_ERROR;
} else if (call->mode != RX_MODE_SENDING) {
call->error = RX_PROTOCOL_ERROR;
}
#ifdef AFS_GLOBAL_RXLOCK_KERNEL
rxi_WaitforTQBusy(call);
#endif /* AFS_GLOBAL_RXLOCK_KERNEL */
cp = call->currentPacket;
if (call->error) {
call->mode = RX_MODE_ERROR;
MUTEX_EXIT(&call->lock);
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
cp->flags |= RX_PKTFLAG_IOVQ;
#endif
queue_Prepend(&call->iovq, cp);
#ifdef RXDEBUG_PACKET
call->iovqc++;
#endif /* RXDEBUG_PACKET */
call->currentPacket = (struct rx_packet *)0;
}
#ifdef RXDEBUG_PACKET
call->iovqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &call->iovq);
return 0;
}
/* Loop through the I/O vector adjusting packet pointers.
* Place full packets back onto the iovq once they are ready
* to send. Set RX_PROTOCOL_ERROR if any problems are found in
* the iovec. We put the loop condition at the end to ensure that
* a zero length write will push a short packet. */
nextio = 0;
queue_Init(&tmpq);
#ifdef RXDEBUG_PACKET
tmpqc = 0;
#endif /* RXDEBUG_PACKET */
do {
if (call->nFree == 0 && cp) {
clock_NewTime(); /* Bogus: need new time package */
/* The 0, below, specifies that it is not the last packet:
* there will be others. PrepareSendPacket may
* alter the packet length by up to
* conn->securityMaxTrailerSize */
hadd32(call->bytesSent, cp->length);
rxi_PrepareSendPacket(call, cp, 0);
queue_Append(&tmpq, cp);
#ifdef RXDEBUG_PACKET
tmpqc++;
#endif /* RXDEBUG_PACKET */
cp = call->currentPacket = (struct rx_packet *)0;
/* The head of the iovq is now the current packet */
if (nbytes) {
if (queue_IsEmpty(&call->iovq)) {
MUTEX_EXIT(&call->lock);
call->error = RX_PROTOCOL_ERROR;
#ifdef RXDEBUG_PACKET
tmpqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &tmpq);
return 0;
}
cp = queue_First(&call->iovq, rx_packet);
queue_Remove(cp);
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_IOVQ;
#endif
#ifdef RXDEBUG_PACKET
call->iovqc--;
#endif /* RXDEBUG_PACKET */
#ifdef RX_TRACK_PACKETS
cp->flags |= RX_PKTFLAG_CP;
#endif
call->currentPacket = cp;
call->nFree = cp->length;
call->curvec = 1;
call->curpos =
(char *)cp->wirevec[1].iov_base +
call->conn->securityHeaderSize;
call->curlen =
cp->wirevec[1].iov_len - call->conn->securityHeaderSize;
}
}
if (nbytes) {
/* The next iovec should point to the current position */
if (iov[nextio].iov_base != call->curpos
|| iov[nextio].iov_len > (int)call->curlen) {
call->error = RX_PROTOCOL_ERROR;
MUTEX_EXIT(&call->lock);
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
queue_Prepend(&tmpq, cp);
#ifdef RXDEBUG_PACKET
tmpqc++;
#endif /* RXDEBUG_PACKET */
cp = call->currentPacket = (struct rx_packet *)0;
}
#ifdef RXDEBUG_PACKET
tmpqc -=
#endif /* RXDEBUG_PACKET */
rxi_FreePackets(0, &tmpq);
return 0;
}
nbytes -= iov[nextio].iov_len;
call->curpos += iov[nextio].iov_len;
call->curlen -= iov[nextio].iov_len;
call->nFree -= iov[nextio].iov_len;
nextio++;
if (call->curlen == 0) {
if (++call->curvec > cp->niovecs) {
call->nFree = 0;
} else {
call->curpos = (char *)cp->wirevec[call->curvec].iov_base;
call->curlen = cp->wirevec[call->curvec].iov_len;
}
}
}
} while (nbytes && nextio < nio);
/* Move the packets from the temporary queue onto the transmit queue.
* We may end up with more than call->twind packets on the queue. */
#ifdef RX_TRACK_PACKETS
for (queue_Scan(&tmpq, p, np, rx_packet))
{
p->flags |= RX_PKTFLAG_TQ;
}
#endif
if (call->error)
call->mode = RX_MODE_ERROR;
queue_SpliceAppend(&call->tq, &tmpq);
if (!(call->flags & (RX_CALL_FAST_RECOVER | RX_CALL_FAST_RECOVER_WAIT))) {
rxi_Start(0, call, 0, 0);
}
/* Wait for the length of the transmit queue to fall below call->twind */
while (!call->error && call->tnext + 1 > call->tfirst + (2 * call->twind)) {
clock_NewTime();
call->startWait = clock_Sec();
#ifdef RX_ENABLE_LOCKS
CV_WAIT(&call->cv_twind, &call->lock);
#else
call->flags |= RX_CALL_WAIT_WINDOW_ALLOC;
osi_rxSleep(&call->twind);
#endif
call->startWait = 0;
}
/* cp is no longer valid since we may have given up the lock */
cp = call->currentPacket;
if (call->error) {
call->mode = RX_MODE_ERROR;
call->currentPacket = NULL;
MUTEX_EXIT(&call->lock);
if (cp) {
#ifdef RX_TRACK_PACKETS
cp->flags &= ~RX_PKTFLAG_CP;
#endif
rxi_FreePacket(cp);
}
return 0;
}
MUTEX_EXIT(&call->lock);
return requestCount - nbytes;
}