int afs_UFSHandleLink(register struct vcache *avc, struct vrequest *areq) { register struct dcache *tdc; register char *tp, *rbuf; void *tfile; afs_size_t offset, len; afs_int32 tlen, alen; register afs_int32 code; /* two different formats, one for links protected 644, have a "." at the * end of the file name, which we turn into a null. Others, protected * 755, we add a null to the end of */ AFS_STATCNT(afs_UFSHandleLink); if (!avc->linkData) { tdc = afs_GetDCache(avc, (afs_size_t) 0, areq, &offset, &len, 0); afs_Trace3(afs_iclSetp, CM_TRACE_UFSLINK, ICL_TYPE_POINTER, avc, ICL_TYPE_POINTER, tdc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length)); if (!tdc) { if (AFS_IS_DISCONNECTED) return ENETDOWN; else return EIO; } /* otherwise we have the data loaded, go for it */ if (len > 1024) { afs_PutDCache(tdc); return EFAULT; } if (avc->f.m.Mode & 0111) alen = len + 1; /* regular link */ else alen = len; /* mt point */ rbuf = (char *)osi_AllocLargeSpace(AFS_LRALLOCSIZ); tlen = len; ObtainReadLock(&tdc->lock); #if defined(LINUX_USE_FH) tfile = osi_UFSOpen_fh(&tdc->f.fh, tdc->f.fh_type); #else tfile = osi_UFSOpen(tdc->f.inode); #endif code = afs_osi_Read(tfile, -1, rbuf, tlen); osi_UFSClose(tfile); ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); rbuf[alen - 1] = '\0'; alen = strlen(rbuf) + 1; tp = afs_osi_Alloc(alen); /* make room for terminating null */ memcpy(tp, rbuf, alen); osi_FreeLargeSpace(rbuf); if (code != tlen) { afs_osi_Free(tp, alen); return EIO; } avc->linkData = tp; } return 0; }
/* Generic write interface */ int afs_osi_Write(struct osi_file *afile, afs_int32 offset, void *aptr, afs_int32 asize) { afs_ucred_t *oldCred; unsigned int resid; afs_int32 code; AFS_STATCNT(osi_Write); if (!afile) osi_Panic("afs_osi_Write called with null param"); if (offset != -1) afile->offset = offset; /* Note the difference in the way the afile->offset is passed (see comments in gop_rdwr() in afs_aix_subr.c for comments) */ AFS_GUNLOCK(); #ifdef AFS_64BIT_KERNEL code = gop_rdwr(UIO_WRITE, afile->vnode, (caddr_t) aptr, asize, &afile->offset, AFS_UIOSYS, NULL, &resid); #else code = gop_rdwr(UIO_WRITE, afile->vnode, (caddr_t) aptr, asize, (off_t) & afile->offset, AFS_UIOSYS, NULL, &resid); #endif AFS_GLOCK(); if (code == 0) { if (resid) afs_Trace3(afs_iclSetp, CM_TRACE_WRITEFAILED, ICL_TYPE_INT32, asize, ICL_TYPE_INT32, resid, ICL_TYPE_INT32, code); code = asize - resid; afile->offset += code; } else { afs_Trace3(afs_iclSetp, CM_TRACE_WRITEFAILED, ICL_TYPE_INT32, asize, ICL_TYPE_INT32, resid, ICL_TYPE_INT32, code); if (code == ENOSPC) afs_warnuser ("\n\n\n*** Cache partition is FULL - Decrease cachesize!!! ***\n\n"); setuerror(code); if (code > 0) { code = -code; } } if (afile->proc) { (*afile->proc) (afile, code); } return code; }
void * afs_MemCacheOpen(afs_dcache_id_t *ainode) { struct memCacheEntry *mep; if (ainode->mem < 0 || ainode->mem > memMaxBlkNumber) { osi_Panic("afs_MemCacheOpen: invalid block #"); } mep = (memCache + ainode->mem); afs_Trace3(afs_iclSetp, CM_TRACE_MEMOPEN, ICL_TYPE_INT32, ainode->mem, ICL_TYPE_POINTER, mep, ICL_TYPE_POINTER, mep ? mep->data : 0); return (void *)mep; }
int afs_root(OSI_VFS_DECL(afsp), struct vnode **avpp) { afs_int32 code = 0; struct vrequest treq; struct vcache *tvp = 0; OSI_VFS_CONVERT(afsp); AFS_STATCNT(afs_root); if (afs_globalVp && (afs_globalVp->f.states & CStatd)) { tvp = afs_globalVp; } else { if (afs_globalVp) { afs_PutVCache(afs_globalVp); afs_globalVp = NULL; } if (!(code = afs_InitReq(&treq, get_user_struct()->u_cred)) && !(code = afs_CheckInit())) { tvp = afs_GetVCache(&afs_rootFid, &treq, NULL, NULL); /* we really want this to stay around */ if (tvp) { afs_globalVp = tvp; } else code = ENOENT; } } if (tvp) { VN_HOLD(AFSTOV(tvp)); AFSTOV(tvp)->v_flag |= VROOT; /* No-op on Ultrix 2.2 */ afs_globalVFS = afsp; *avpp = AFSTOV(tvp); } afs_Trace3(afs_iclSetp, CM_TRACE_GOPEN, ICL_TYPE_POINTER, *avpp, ICL_TYPE_INT32, 0, ICL_TYPE_INT32, code); return code; }
/* 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; } }
/* * afs_TruncateAllSegments * * Description: * Truncate a cache file. * * Parameters: * avc : Ptr to vcache entry to truncate. * alen : Number of bytes to make the file. * areq : Ptr to request structure. * * Environment: * Called with avc write-locked; in VFS40 systems, pvnLock is also * held. */ int afs_TruncateAllSegments(struct vcache *avc, afs_size_t alen, struct vrequest *areq, afs_ucred_t *acred) { struct dcache *tdc; afs_int32 code; afs_int32 index; afs_size_t newSize; int dcCount, dcPos; struct dcache **tdcArray = NULL; AFS_STATCNT(afs_TruncateAllSegments); avc->f.m.Date = osi_Time(); afs_Trace3(afs_iclSetp, CM_TRACE_TRUNCALL, ICL_TYPE_POINTER, avc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length), ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(alen)); if (alen >= avc->f.m.Length) { /* * Special speedup since Sun's vm extends the file this way; * we've never written to the file thus we can just set the new * length and avoid the needless calls below. * Also used for ftruncate calls which can extend the file. * To completely minimize the possible extra StoreMini RPC, we really * should keep the ExtendedPos as well and clear this flag if we * truncate below that value before we store the file back. */ avc->f.states |= CExtendedFile; avc->f.m.Length = alen; return 0; } #if (defined(AFS_SUN5_ENV)) /* Zero unused portion of last page */ osi_VM_PreTruncate(avc, alen, acred); #endif #if (defined(AFS_SUN5_ENV)) ObtainWriteLock(&avc->vlock, 546); avc->activeV++; /* Block new getpages */ ReleaseWriteLock(&avc->vlock); #endif ReleaseWriteLock(&avc->lock); AFS_GUNLOCK(); /* Flush pages beyond end-of-file. */ osi_VM_Truncate(avc, alen, acred); AFS_GLOCK(); ObtainWriteLock(&avc->lock, 79); avc->f.m.Length = alen; if (alen < avc->f.truncPos) avc->f.truncPos = alen; code = DVHash(&avc->f.fid); /* block out others from screwing with this table */ ObtainWriteLock(&afs_xdcache, 287); dcCount = 0; for (index = afs_dvhashTbl[code]; index != NULLIDX;) { if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) { 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++; afs_PutDCache(tdc); } index = afs_dvnextTbl[index]; } /* Now allocate space where we can save those dcache entries, and * do a second pass over them.. Since we're holding xdcache, it * shouldn't be changing. */ tdcArray = osi_Alloc(dcCount * sizeof(struct dcache *)); dcPos = 0; for (index = afs_dvhashTbl[code]; index != NULLIDX; index = afs_dvnextTbl[index]) { if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) { 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 */ if (dcPos < dcCount) { tdcArray[dcPos++] = tdc; } else { afs_PutDCache(tdc); } } else { afs_PutDCache(tdc); } } } ReleaseWriteLock(&afs_xdcache); /* Now we loop over the array of dcache entries and truncate them */ for (index = 0; index < dcPos; index++) { struct osi_file *tfile; tdc = tdcArray[index]; newSize = alen - AFS_CHUNKTOBASE(tdc->f.chunk); if (newSize < 0) newSize = 0; ObtainSharedLock(&tdc->lock, 672); 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, (afs_int32)newSize); afs_CFileClose(tfile); afs_AdjustSize(tdc, (afs_int32)newSize); if (alen < tdc->validPos) { if (alen < AFS_CHUNKTOBASE(tdc->f.chunk)) tdc->validPos = 0; else tdc->validPos = alen; } ConvertWToSLock(&tdc->lock); } ReleaseSharedLock(&tdc->lock); afs_PutDCache(tdc); } 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)) { avc->vstates &= ~VRevokeWait; afs_osi_Wakeup((char *)&avc->vstates); } ReleaseWriteLock(&avc->vlock); #endif return code; }
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) */
int afs_MemRead(struct vcache *avc, struct uio *auio, afs_ucred_t *acred, daddr_t albn, struct buf **abpp, int noLock) { afs_size_t totalLength; afs_size_t transferLength; afs_size_t filePos; afs_size_t offset, tlen; afs_size_t len = 0; afs_int32 trimlen; struct dcache *tdc = 0; afs_int32 error, trybusy = 1; afs_int32 code; struct vrequest *treq = NULL; #ifdef AFS_DARWIN80_ENV uio_t tuiop = NULL; #else struct uio tuio; struct uio *tuiop = &tuio; struct iovec *tvec; memset(&tuio, 0, sizeof(tuio)); #endif AFS_STATCNT(afs_MemRead); if (avc->vc_error) return EIO; /* check that we have the latest status info in the vnode cache */ if ((code = afs_CreateReq(&treq, acred))) return code; if (!noLock) { code = afs_VerifyVCache(avc, treq); if (code) { code = afs_CheckCode(code, treq, 8); /* failed to get it */ afs_DestroyReq(treq); return code; } } #ifndef AFS_VM_RDWR_ENV if (AFS_NFSXLATORREQ(acred)) { if (!afs_AccessOK (avc, PRSFS_READ, treq, CHECK_MODE_BITS | CMB_ALLOW_EXEC_AS_READ)) { code = afs_CheckCode(EACCES, treq, 9); afs_DestroyReq(treq); return code; } } #endif #ifndef AFS_DARWIN80_ENV tvec = (struct iovec *)osi_AllocSmallSpace(sizeof(struct iovec)); memset(tvec, 0, sizeof(struct iovec)); #endif totalLength = AFS_UIO_RESID(auio); filePos = AFS_UIO_OFFSET(auio); afs_Trace4(afs_iclSetp, CM_TRACE_READ, ICL_TYPE_POINTER, avc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_INT32, totalLength, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length)); error = 0; transferLength = 0; if (!noLock) ObtainReadLock(&avc->lock); #if defined(AFS_TEXT_ENV) && !defined(AFS_VM_RDWR_ENV) if (avc->flushDV.high == AFS_MAXDV && avc->flushDV.low == AFS_MAXDV) { hset(avc->flushDV, avc->f.m.DataVersion); } #endif /* * Locks held: * avc->lock(R) */ /* This bit is bogus. We're checking to see if the read goes past the * end of the file. If so, we should be zeroing out all of the buffers * that the client has passed into us (there is a danger that we may leak * kernel memory if we do not). However, this behaviour is disabled by * not setting len before this segment runs, and by setting len to 0 * immediately we enter it. In addition, we also need to check for a read * which partially goes off the end of the file in the while loop below. */ if (filePos >= avc->f.m.Length) { if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ len = 0; #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); } while (avc->f.m.Length > 0 && totalLength > 0) { /* read all of the cached info */ if (filePos >= avc->f.m.Length) break; /* all done */ if (noLock) { if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); } tdc = afs_FindDCache(avc, filePos); if (tdc) { ObtainReadLock(&tdc->lock); offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); len = tdc->f.chunkBytes - offset; } } else { int versionOk; /* a tricky question: does the presence of the DFFetching flag * mean that we're fetching the latest version of the file? No. * The server could update the file as soon as the fetch responsible * for the setting of the DFFetching flag completes. * * However, the presence of the DFFetching flag (visible under * a dcache read lock since it is set and cleared only under a * dcache write lock) means that we're fetching as good a version * as was known to this client at the time of the last call to * afs_VerifyVCache, since the latter updates the stat cache's * m.DataVersion field under a vcache write lock, and from the * time that the DFFetching flag goes on in afs_GetDCache (before * the fetch starts), to the time it goes off (after the fetch * completes), afs_GetDCache keeps at least a read lock on the * vcache entry. * * This means that if the DFFetching flag is set, we can use that * data for any reads that must come from the current version of * the file (current == m.DataVersion). * * Another way of looking at this same point is this: if we're * fetching some data and then try do an afs_VerifyVCache, the * VerifyVCache operation will not complete until after the * DFFetching flag is turned off and the dcache entry's f.versionNo * field is updated. * * Note, by the way, that if DFFetching is set, * m.DataVersion > f.versionNo (the latter is not updated until * after the fetch completes). */ if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); /* before reusing tdc */ } #ifdef STRUCT_TASK_STRUCT_HAS_CRED try_background: #endif tdc = afs_GetDCache(avc, filePos, treq, &offset, &len, 2); ObtainReadLock(&tdc->lock); /* now, first try to start transfer, if we'll need the data. If * data already coming, we don't need to do this, obviously. Type * 2 requests never return a null dcache entry, btw. */ if (!(tdc->dflags & DFFetching) && !hsame(avc->f.m.DataVersion, tdc->f.versionNo)) { /* have cache entry, it is not coming in now, * and we'll need new data */ tagain: #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (trybusy && (!afs_BBusy() || (afs_protocols & VICEP_ACCESS))) { #else if (trybusy && !afs_BBusy()) { #endif struct brequest *bp; /* daemon is not busy */ ObtainSharedLock(&tdc->mflock, 665); if (!(tdc->mflags & DFFetchReq)) { int dontwait = B_DONTWAIT; /* start the daemon (may already be running, however) */ UpgradeSToWLock(&tdc->mflock, 666); tdc->mflags |= DFFetchReq; #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (afs_protocols & VICEP_ACCESS) dontwait = 0; #endif bp = afs_BQueue(BOP_FETCH, avc, dontwait, 0, acred, (afs_size_t) filePos, (afs_size_t) 0, tdc, (void *)0, (void *)0); if (!bp) { tdc->mflags &= ~DFFetchReq; trybusy = 0; /* Avoid bkg daemon since they're too busy */ ReleaseWriteLock(&tdc->mflock); goto tagain; } ConvertWToSLock(&tdc->mflock); /* don't use bp pointer! */ } code = 0; ConvertSToRLock(&tdc->mflock); while (!code && tdc->mflags & DFFetchReq) { afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); /* don't need waiting flag on this one */ ReleaseReadLock(&tdc->mflock); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); ObtainReadLock(&tdc->mflock); } ReleaseReadLock(&tdc->mflock); if (code) { error = code; break; } } } /* now data may have started flowing in (if DFFetching is on). If * data is now streaming in, then wait for some interesting stuff. */ code = 0; while (!code && (tdc->dflags & DFFetching) && tdc->validPos <= filePos) { /* too early: wait for DFFetching flag to vanish, * or data to appear */ afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); } if (code) { error = code; break; } /* fetching flag gone, data is here, or we never tried * (BBusy for instance) */ len = tdc->validPos - filePos; versionOk = hsame(avc->f.m.DataVersion, tdc->f.versionNo) ? 1 : 0; if (tdc->dflags & DFFetching) { /* still fetching, some new data is here: * compute length and offset */ offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); } else { /* no longer fetching, verify data version * (avoid new GetDCache call) */ if (versionOk && len > 0) { offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); } else { /* don't have current data, so get it below */ afs_Trace3(afs_iclSetp, CM_TRACE_VERSIONNO, ICL_TYPE_INT64, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_HYPER, &avc->f.m.DataVersion, ICL_TYPE_HYPER, &tdc->f.versionNo); #if 0 #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (afs_protocols & VICEP_ACCESS) { printf("afs_read: DV mismatch? %d instead of %d for %u.%u.%u\n", tdc->f.versionNo.low, avc->f.m.DataVersion.low, avc->f.fid.Fid.Volume, avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique); printf("afs_read: validPos %llu filePos %llu totalLength %lld m.Length %llu noLock %d\n", tdc->validPos, filePos, totalLength, avc->f.m.Length, noLock); printf("afs_read: or len too low? %lld for %u.%u.%u\n", len, avc->f.fid.Fid.Volume, avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique); } #endif #endif ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); tdc = NULL; } } if (!tdc) { #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (afs_protocols & VICEP_ACCESS) { /* avoid foreground fetch */ if (!versionOk) { printf("afs_read: avoid forground %u.%u.%u\n", avc->f.fid.Fid.Volume, avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique); goto try_background; } #if 0 printf("afs_read: forground %u.%u.%u\n", avc->f.fid.Fid.Volume, avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique); #endif } #endif /* If we get here, it was not possible to start the * background daemon. With flag == 1 afs_GetDCache * does the FetchData rpc synchronously. */ ReleaseReadLock(&avc->lock); tdc = afs_GetDCache(avc, filePos, treq, &offset, &len, 1); ObtainReadLock(&avc->lock); if (tdc) ObtainReadLock(&tdc->lock); } } afs_Trace3(afs_iclSetp, CM_TRACE_VNODEREAD, ICL_TYPE_POINTER, tdc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(offset), ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(len)); if (!tdc) { error = EIO; break; } /* * Locks held: * avc->lock(R) * tdc->lock(R) */ if (len > totalLength) len = totalLength; /* will read len bytes */ if (len <= 0) { /* shouldn't get here if DFFetching is on */ /* read past the end of a chunk, may not be at next chunk yet, and yet * also not at eof, so may have to supply fake zeros */ len = AFS_CHUNKTOSIZE(tdc->f.chunk) - offset; /* bytes left in chunk addr space */ if (len > totalLength) len = totalLength; /* and still within xfr request */ tlen = avc->f.m.Length - offset; /* and still within file */ if (len > tlen) len = tlen; if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); if (code) { error = code; break; } } else { /* get the data from the mem cache */ /* mung uio structure to be right for this transfer */ #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); uio_setoffset(tuiop, offset); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); tuio.afsio_offset = offset; #endif code = afs_MemReadUIO(&tdc->f.inode, tuiop); if (code) { error = code; break; } } /* otherwise we've read some, fixup length, etc and continue with next seg */ len = len - AFS_UIO_RESID(tuiop); /* compute amount really transferred */ trimlen = len; afsio_skip(auio, trimlen); /* update input uio structure */ totalLength -= len; transferLength += len; filePos += len; if (len <= 0) break; /* surprise eof */ #ifdef AFS_DARWIN80_ENV if (tuiop) { uio_free(tuiop); tuiop = 0; } #endif } /* the whole while loop */ /* * Locks held: * avc->lock(R) * tdc->lock(R) if tdc */ /* if we make it here with tdc non-zero, then it is the last chunk we * dealt with, and we have to release it when we're done. We hold on * to it in case we need to do a prefetch. */ if (tdc) { ReleaseReadLock(&tdc->lock); /* * try to queue prefetch, if needed. If DataVersion is zero there * should not be any more: files with DV 0 never have been stored * on the fileserver, symbolic links and directories never require * more than a single chunk. */ if (!noLock && !(hiszero(avc->f.m.DataVersion)) && #ifndef AFS_VM_RDWR_ENV afs_preCache #else 1 #endif ) { afs_PrefetchChunk(avc, tdc, acred, treq); } afs_PutDCache(tdc); } if (!noLock) ReleaseReadLock(&avc->lock); #ifdef AFS_DARWIN80_ENV if (tuiop) uio_free(tuiop); #else osi_FreeSmallSpace(tvec); #endif error = afs_CheckCode(error, treq, 10); afs_DestroyReq(treq); return error; } /* called with the dcache entry triggering the fetch, the vcache entry involved, * and a vrequest for the read call. Marks the dcache entry as having already * triggered a prefetch, starts the prefetch going and sets the DFFetchReq * flag in the prefetched block, so that the next call to read knows to wait * for the daemon to start doing things. * * This function must be called with the vnode at least read-locked, and * no locks on the dcache, because it plays around with dcache entries. */ void afs_PrefetchChunk(struct vcache *avc, struct dcache *adc, afs_ucred_t *acred, struct vrequest *areq) { struct dcache *tdc; afs_size_t offset; afs_size_t j1, j2; /* junk vbls for GetDCache to trash */ offset = adc->f.chunk + 1; /* next chunk we'll need */ offset = AFS_CHUNKTOBASE(offset); /* base of next chunk */ ObtainReadLock(&adc->lock); ObtainSharedLock(&adc->mflock, 662); if (offset < avc->f.m.Length && !(adc->mflags & DFNextStarted) && !afs_BBusy()) { struct brequest *bp; UpgradeSToWLock(&adc->mflock, 663); adc->mflags |= DFNextStarted; /* we've tried to prefetch for this guy */ ReleaseWriteLock(&adc->mflock); ReleaseReadLock(&adc->lock); tdc = afs_GetDCache(avc, offset, areq, &j1, &j2, 2); /* type 2 never returns 0 */ /* * In disconnected mode, type 2 can return 0 because it doesn't * make any sense to allocate a dcache we can never fill */ if (tdc == NULL) return; ObtainSharedLock(&tdc->mflock, 651); if (!(tdc->mflags & DFFetchReq)) { /* ask the daemon to do the work */ UpgradeSToWLock(&tdc->mflock, 652); tdc->mflags |= DFFetchReq; /* guaranteed to be cleared by BKG or GetDCache */ /* last parm (1) tells bkg daemon to do an afs_PutDCache when it is done, * since we don't want to wait for it to finish before doing so ourselves. */ bp = afs_BQueue(BOP_FETCH, avc, B_DONTWAIT, 0, acred, (afs_size_t) offset, (afs_size_t) 1, tdc, (void *)0, (void *)0); if (!bp) { /* Bkg table full; just abort non-important prefetching to avoid deadlocks */ tdc->mflags &= ~DFFetchReq; ReleaseWriteLock(&tdc->mflock); afs_PutDCache(tdc); /* * DCLOCKXXX: This is a little sketchy, since someone else * could have already started a prefetch.. In practice, * this probably doesn't matter; at most it would cause an * extra slot in the BKG table to be used up when someone * prefetches this for the second time. */ ObtainReadLock(&adc->lock); ObtainWriteLock(&adc->mflock, 664); adc->mflags &= ~DFNextStarted; ReleaseWriteLock(&adc->mflock); ReleaseReadLock(&adc->lock); } else { ReleaseWriteLock(&tdc->mflock); } } else { ReleaseSharedLock(&tdc->mflock); afs_PutDCache(tdc); } } else { ReleaseSharedLock(&adc->mflock); ReleaseReadLock(&adc->lock); } } int afs_UFSRead(struct vcache *avc, struct uio *auio, afs_ucred_t *acred, daddr_t albn, struct buf **abpp, int noLock) { afs_size_t totalLength; afs_size_t transferLength; afs_size_t filePos; afs_size_t offset, tlen; afs_size_t len = 0; afs_int32 trimlen; struct dcache *tdc = 0; afs_int32 error; struct osi_file *tfile; afs_int32 code; int trybusy = 1; struct vrequest *treq = NULL; #ifdef AFS_DARWIN80_ENV uio_t tuiop=NULL; #else struct uio tuio; struct uio *tuiop = &tuio; struct iovec *tvec; memset(&tuio, 0, sizeof(tuio)); #endif AFS_STATCNT(afs_UFSRead); if (avc && avc->vc_error) return EIO; AFS_DISCON_LOCK(); /* check that we have the latest status info in the vnode cache */ if ((code = afs_CreateReq(&treq, acred))) return code; if (!noLock) { if (!avc) osi_Panic("null avc in afs_UFSRead"); else { code = afs_VerifyVCache(avc, treq); if (code) { code = afs_CheckCode(code, treq, 11); /* failed to get it */ afs_DestroyReq(treq); AFS_DISCON_UNLOCK(); return code; } } } #ifndef AFS_VM_RDWR_ENV if (AFS_NFSXLATORREQ(acred)) { if (!afs_AccessOK (avc, PRSFS_READ, treq, CHECK_MODE_BITS | CMB_ALLOW_EXEC_AS_READ)) { AFS_DISCON_UNLOCK(); code = afs_CheckCode(EACCES, treq, 12); afs_DestroyReq(treq); return code; } } #endif #ifndef AFS_DARWIN80_ENV tvec = (struct iovec *)osi_AllocSmallSpace(sizeof(struct iovec)); memset(tvec, 0, sizeof(struct iovec)); #endif totalLength = AFS_UIO_RESID(auio); filePos = AFS_UIO_OFFSET(auio); afs_Trace4(afs_iclSetp, CM_TRACE_READ, ICL_TYPE_POINTER, avc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_INT32, totalLength, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length)); error = 0; transferLength = 0; if (!noLock) ObtainReadLock(&avc->lock); #if defined(AFS_TEXT_ENV) && !defined(AFS_VM_RDWR_ENV) if (avc->flushDV.high == AFS_MAXDV && avc->flushDV.low == AFS_MAXDV) { hset(avc->flushDV, avc->f.m.DataVersion); } #endif /* This bit is bogus. We're checking to see if the read goes past the * end of the file. If so, we should be zeroing out all of the buffers * that the client has passed into us (there is a danger that we may leak * kernel memory if we do not). However, this behaviour is disabled by * not setting len before this segment runs, and by setting len to 0 * immediately we enter it. In addition, we also need to check for a read * which partially goes off the end of the file in the while loop below. */ if (filePos >= avc->f.m.Length) { if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ len = 0; #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); } while (avc->f.m.Length > 0 && totalLength > 0) { /* read all of the cached info */ if (filePos >= avc->f.m.Length) break; /* all done */ if (noLock) { if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); } tdc = afs_FindDCache(avc, filePos); if (tdc) { ObtainReadLock(&tdc->lock); offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); len = tdc->validPos - filePos; } } else { int versionOk; /* a tricky question: does the presence of the DFFetching flag * mean that we're fetching the latest version of the file? No. * The server could update the file as soon as the fetch responsible * for the setting of the DFFetching flag completes. * * However, the presence of the DFFetching flag (visible under * a dcache read lock since it is set and cleared only under a * dcache write lock) means that we're fetching as good a version * as was known to this client at the time of the last call to * afs_VerifyVCache, since the latter updates the stat cache's * m.DataVersion field under a vcache write lock, and from the * time that the DFFetching flag goes on in afs_GetDCache (before * the fetch starts), to the time it goes off (after the fetch * completes), afs_GetDCache keeps at least a read lock on the * vcache entry. * * This means that if the DFFetching flag is set, we can use that * data for any reads that must come from the current version of * the file (current == m.DataVersion). * * Another way of looking at this same point is this: if we're * fetching some data and then try do an afs_VerifyVCache, the * VerifyVCache operation will not complete until after the * DFFetching flag is turned off and the dcache entry's f.versionNo * field is updated. * * Note, by the way, that if DFFetching is set, * m.DataVersion > f.versionNo (the latter is not updated until * after the fetch completes). */ if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); /* before reusing tdc */ } #ifdef STRUCT_TASK_STRUCT_HAS_CRED try_background: #endif tdc = afs_GetDCache(avc, filePos, treq, &offset, &len, 2); if (!tdc) { error = ENETDOWN; break; } ObtainReadLock(&tdc->lock); /* now, first try to start transfer, if we'll need the data. If * data already coming, we don't need to do this, obviously. Type * 2 requests never return a null dcache entry, btw. */ if (!(tdc->dflags & DFFetching) && !hsame(avc->f.m.DataVersion, tdc->f.versionNo)) { /* have cache entry, it is not coming in now, and we'll need new data */ tagain: #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (trybusy && (!afs_BBusy() || (afs_protocols & VICEP_ACCESS))) { #else if (trybusy && !afs_BBusy()) { #endif struct brequest *bp; /* daemon is not busy */ ObtainSharedLock(&tdc->mflock, 667); if (!(tdc->mflags & DFFetchReq)) { int dontwait = B_DONTWAIT; UpgradeSToWLock(&tdc->mflock, 668); tdc->mflags |= DFFetchReq; #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (afs_protocols & VICEP_ACCESS) dontwait = 0; #endif bp = afs_BQueue(BOP_FETCH, avc, dontwait, 0, acred, (afs_size_t) filePos, (afs_size_t) 0, tdc, (void *)0, (void *)0); if (!bp) { /* Bkg table full; retry deadlocks */ tdc->mflags &= ~DFFetchReq; trybusy = 0; /* Avoid bkg daemon since they're too busy */ ReleaseWriteLock(&tdc->mflock); goto tagain; } ConvertWToSLock(&tdc->mflock); } code = 0; ConvertSToRLock(&tdc->mflock); while (!code && tdc->mflags & DFFetchReq) { afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); /* don't need waiting flag on this one */ ReleaseReadLock(&tdc->mflock); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); ObtainReadLock(&tdc->mflock); } ReleaseReadLock(&tdc->mflock); if (code) { error = code; break; } } } /* now data may have started flowing in (if DFFetching is on). If * data is now streaming in, then wait for some interesting stuff. */ code = 0; while (!code && (tdc->dflags & DFFetching) && tdc->validPos <= filePos) { /* too early: wait for DFFetching flag to vanish, * or data to appear */ afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); } if (code) { error = code; break; } /* fetching flag gone, data is here, or we never tried * (BBusy for instance) */ len = tdc->validPos - filePos; versionOk = hsame(avc->f.m.DataVersion, tdc->f.versionNo) ? 1 : 0; if (tdc->dflags & DFFetching) { /* still fetching, some new data is here: * compute length and offset */ offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); } else { /* no longer fetching, verify data version (avoid new * GetDCache call) */ if (versionOk && len > 0) { offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); } else { /* don't have current data, so get it below */ afs_Trace3(afs_iclSetp, CM_TRACE_VERSIONNO, ICL_TYPE_INT64, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_HYPER, &avc->f.m.DataVersion, ICL_TYPE_HYPER, &tdc->f.versionNo); #if 0 #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (afs_protocols & VICEP_ACCESS) { printf("afs_read: DV mismatch? %d instead of %d for %u.%u.%u\n", tdc->f.versionNo.low, avc->f.m.DataVersion.low, avc->f.fid.Fid.Volume, avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique); printf("afs_read: validPos %llu filePos %llu totalLength %d m.Length %llu noLock %d\n", tdc->validPos, filePos, totalLength, avc->f.m.Length, noLock); printf("afs_read: or len too low? %lld for %u.%u.%u\n", len, avc->f.fid.Fid.Volume, avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique); } #endif #endif ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); tdc = NULL; } } if (!tdc) { #ifdef STRUCT_TASK_STRUCT_HAS_CRED if (afs_protocols & VICEP_ACCESS) { /* avoid foreground fetch */ if (!versionOk) { printf("afs_read: avoid forground %u.%u.%u\n", avc->f.fid.Fid.Volume, avc->f.fid.Fid.Vnode, avc->f.fid.Fid.Unique); goto try_background; } } #endif /* If we get here, it was not possible to start the * background daemon. With flag == 1 afs_GetDCache * does the FetchData rpc synchronously. */ ReleaseReadLock(&avc->lock); tdc = afs_GetDCache(avc, filePos, treq, &offset, &len, 1); ObtainReadLock(&avc->lock); if (tdc) ObtainReadLock(&tdc->lock); } } if (!tdc) { error = EIO; break; } len = tdc->validPos - filePos; afs_Trace3(afs_iclSetp, CM_TRACE_VNODEREAD, ICL_TYPE_POINTER, tdc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(offset), ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(len)); if (len > totalLength) len = totalLength; /* will read len bytes */ if (len <= 0) { /* shouldn't get here if DFFetching is on */ afs_Trace4(afs_iclSetp, CM_TRACE_VNODEREAD2, ICL_TYPE_POINTER, tdc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(tdc->validPos), ICL_TYPE_INT32, tdc->f.chunkBytes, ICL_TYPE_INT32, tdc->dflags); /* read past the end of a chunk, may not be at next chunk yet, and yet * also not at eof, so may have to supply fake zeros */ len = AFS_CHUNKTOSIZE(tdc->f.chunk) - offset; /* bytes left in chunk addr space */ if (len > totalLength) len = totalLength; /* and still within xfr request */ tlen = avc->f.m.Length - offset; /* and still within file */ if (len > tlen) len = tlen; if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); if (code) { error = code; break; } } else { /* get the data from the file */ tfile = (struct osi_file *)osi_UFSOpen(&tdc->f.inode); #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); uio_setoffset(tuiop, offset); #else /* mung uio structure to be right for this transfer */ afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); tuio.afsio_offset = offset; #endif #if defined(AFS_AIX41_ENV) AFS_GUNLOCK(); code = VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, &tuio, NULL, NULL, NULL, afs_osi_credp); AFS_GLOCK(); #elif defined(AFS_AIX32_ENV) code = VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, &tuio, NULL, NULL); /* Flush all JFS pages now for big performance gain in big file cases * If we do something like this, must check to be sure that AFS file * isn't mmapped... see afs_gn_map() for why. */ /* if (tfile->vnode->v_gnode && tfile->vnode->v_gnode->gn_seg) { many different ways to do similar things: so far, the best performing one is #2, but #1 might match it if we straighten out the confusion regarding which pages to flush. It really does matter. 1. vm_flushp(tfile->vnode->v_gnode->gn_seg, 0, len/PAGESIZE - 1); 2. vm_releasep(tfile->vnode->v_gnode->gn_seg, offset/PAGESIZE, (len + PAGESIZE-1)/PAGESIZE); 3. vms_inactive(tfile->vnode->v_gnode->gn_seg) Doesn't work correctly 4. vms_delete(tfile->vnode->v_gnode->gn_seg) probably also fails tfile->vnode->v_gnode->gn_seg = NULL; 5. deletep 6. ipgrlse 7. ifreeseg Unfortunately, this seems to cause frequent "cache corruption" episodes. vm_releasep(tfile->vnode->v_gnode->gn_seg, offset/PAGESIZE, (len + PAGESIZE-1)/PAGESIZE); } */ #elif defined(AFS_AIX_ENV) code = VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, (off_t) & offset, &tuio, NULL, NULL, -1); #elif defined(AFS_SUN5_ENV) AFS_GUNLOCK(); #ifdef AFS_SUN510_ENV VOP_RWLOCK(tfile->vnode, 0, NULL); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp, NULL); VOP_RWUNLOCK(tfile->vnode, 0, NULL); #else VOP_RWLOCK(tfile->vnode, 0); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_RWUNLOCK(tfile->vnode, 0); #endif AFS_GLOCK(); #elif defined(AFS_SGI_ENV) AFS_GUNLOCK(); AFS_VOP_RWLOCK(tfile->vnode, VRWLOCK_READ); AFS_VOP_READ(tfile->vnode, &tuio, IO_ISLOCKED, afs_osi_credp, code); AFS_VOP_RWUNLOCK(tfile->vnode, VRWLOCK_READ); AFS_GLOCK(); #elif defined(AFS_HPUX100_ENV) AFS_GUNLOCK(); code = VOP_RDWR(tfile->vnode, &tuio, UIO_READ, 0, afs_osi_credp); AFS_GLOCK(); #elif defined(AFS_LINUX20_ENV) AFS_GUNLOCK(); code = osi_rdwr(tfile, &tuio, UIO_READ); AFS_GLOCK(); #elif defined(AFS_DARWIN80_ENV) AFS_GUNLOCK(); code = VNOP_READ(tfile->vnode, tuiop, 0, afs_osi_ctxtp); AFS_GLOCK(); #elif defined(AFS_DARWIN_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, current_proc()); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0, current_proc()); AFS_GLOCK(); #elif defined(AFS_FBSD80_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0); AFS_GLOCK(); #elif defined(AFS_FBSD_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, curthread); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0, curthread); AFS_GLOCK(); #elif defined(AFS_NBSD_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0); AFS_GLOCK(); #elif defined(AFS_XBSD_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, curproc); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0, curproc); AFS_GLOCK(); #else code = VOP_RDWR(tfile->vnode, &tuio, UIO_READ, 0, afs_osi_credp); #endif osi_UFSClose(tfile); if (code) { error = code; break; } } /* otherwise we've read some, fixup length, etc and continue with next seg */ len = len - AFS_UIO_RESID(tuiop); /* compute amount really transferred */ trimlen = len; afsio_skip(auio, trimlen); /* update input uio structure */ totalLength -= len; transferLength += len; filePos += len; if (len <= 0) break; /* surprise eof */ #ifdef AFS_DARWIN80_ENV if (tuiop) { uio_free(tuiop); tuiop = 0; } #endif } /* if we make it here with tdc non-zero, then it is the last chunk we * dealt with, and we have to release it when we're done. We hold on * to it in case we need to do a prefetch, obviously. */ if (tdc) { ReleaseReadLock(&tdc->lock); #if !defined(AFS_VM_RDWR_ENV) /* * try to queue prefetch, if needed. If DataVersion is zero there * should not be any more: files with DV 0 never have been stored * on the fileserver, symbolic links and directories never require * more than a single chunk. */ if (!noLock && !(hiszero(avc->f.m.DataVersion))) { if (!(tdc->mflags & DFNextStarted)) afs_PrefetchChunk(avc, tdc, acred, treq); } #endif afs_PutDCache(tdc); } if (!noLock) ReleaseReadLock(&avc->lock); #ifdef AFS_DARWIN80_ENV if (tuiop) uio_free(tuiop); #else osi_FreeSmallSpace(tvec); #endif AFS_DISCON_UNLOCK(); error = afs_CheckCode(error, treq, 13); afs_DestroyReq(treq); return error; }
/* this call, unlike osi_FlushText, is supposed to discard caches that may contain invalid information if a file is written remotely, but that may contain valid information that needs to be written back if the file is being written locally. It doesn't subsume osi_FlushText, since the latter function may be needed to flush caches that are invalidated by local writes. avc->pvnLock is already held, avc->lock is guaranteed not to be held (by us, of course). */ void osi_FlushPages(struct vcache *avc, afs_ucred_t *credp) { afs_hyper_t origDV; #if defined(AFS_CACHE_BYPASS) /* The optimization to check DV under read lock below is identical a * change in CITI cache bypass work. The problem CITI found in 1999 * was that this code and background daemon doing prefetching competed * for the vcache entry shared lock. It's not clear to me from the * tech report, but it looks like CITI fixed the general prefetch code * path as a bonus when experimenting on prefetch for cache bypass, see * citi-tr-01-3. */ #endif if (vType(avc) == VDIR) { /* not applicable to directories; they're never mapped or stored in * pages */ return; } ObtainReadLock(&avc->lock); /* If we've already purged this version, or if we're the ones * writing this version, don't flush it (could lose the * data we're writing). */ if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0) || ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) { ReleaseReadLock(&avc->lock); return; } ReleaseReadLock(&avc->lock); ObtainWriteLock(&avc->lock, 10); /* Check again */ if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0) || ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) { ReleaseWriteLock(&avc->lock); return; } /* At this point, you might think that we can skip trying to flush pages * if mapDV is zero, since a file with a zero DV will not have any data in * it. However, some platforms (notably Linux 2.6.22+) will keep a page * full of zeroes around for an empty file. So play it safe and always * flush pages. */ AFS_STATCNT(osi_FlushPages); hset(origDV, avc->f.m.DataVersion); afs_Trace3(afs_iclSetp, CM_TRACE_FLUSHPAGES, ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, origDV.low, ICL_TYPE_INT32, avc->f.m.Length); ReleaseWriteLock(&avc->lock); #ifndef AFS_FBSD70_ENV AFS_GUNLOCK(); #endif osi_VM_FlushPages(avc, credp); #ifndef AFS_FBSD70_ENV AFS_GLOCK(); #endif ObtainWriteLock(&avc->lock, 88); /* do this last, and to original version, since stores may occur * while executing above PUTPAGE call */ hset(avc->mapDV, origDV); ReleaseWriteLock(&avc->lock); }
/* this call, unlike osi_FlushText, is supposed to discard caches that may contain invalid information if a file is written remotely, but that may contain valid information that needs to be written back if the file is being written locally. It doesn't subsume osi_FlushText, since the latter function may be needed to flush caches that are invalidated by local writes. avc->pvnLock is already held, avc->lock is guaranteed not to be held (by us, of course). */ void osi_FlushPages(struct vcache *avc, afs_ucred_t *credp) { #ifdef AFS_FBSD70_ENV int vfslocked; #endif afs_hyper_t origDV; #if defined(AFS_CACHE_BYPASS) /* The optimization to check DV under read lock below is identical a * change in CITI cache bypass work. The problem CITI found in 1999 * was that this code and background daemon doing prefetching competed * for the vcache entry shared lock. It's not clear to me from the * tech report, but it looks like CITI fixed the general prefetch code * path as a bonus when experimenting on prefetch for cache bypass, see * citi-tr-01-3. */ #endif ObtainReadLock(&avc->lock); /* If we've already purged this version, or if we're the ones * writing this version, don't flush it (could lose the * data we're writing). */ if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0) || ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) { ReleaseReadLock(&avc->lock); return; } ReleaseReadLock(&avc->lock); ObtainWriteLock(&avc->lock, 10); /* Check again */ if ((hcmp((avc->f.m.DataVersion), (avc->mapDV)) <= 0) || ((avc->execsOrWriters > 0) && afs_DirtyPages(avc))) { ReleaseWriteLock(&avc->lock); return; } if (hiszero(avc->mapDV)) { hset(avc->mapDV, avc->f.m.DataVersion); ReleaseWriteLock(&avc->lock); return; } AFS_STATCNT(osi_FlushPages); hset(origDV, avc->f.m.DataVersion); afs_Trace3(afs_iclSetp, CM_TRACE_FLUSHPAGES, ICL_TYPE_POINTER, avc, ICL_TYPE_INT32, origDV.low, ICL_TYPE_INT32, avc->f.m.Length); ReleaseWriteLock(&avc->lock); #ifdef AFS_FBSD70_ENV vfslocked = VFS_LOCK_GIANT(AFSTOV(avc)->v_mount); #endif #ifndef AFS_FBSD70_ENV AFS_GUNLOCK(); #endif osi_VM_FlushPages(avc, credp); #ifndef AFS_FBSD70_ENV AFS_GLOCK(); #endif #ifdef AFS_FBSD70_ENV VFS_UNLOCK_GIANT(vfslocked); #endif ObtainWriteLock(&avc->lock, 88); /* do this last, and to original version, since stores may occur * while executing above PUTPAGE call */ hset(avc->mapDV, origDV); ReleaseWriteLock(&avc->lock); }
int afs_MemRead(register struct vcache *avc, struct uio *auio, struct AFS_UCRED *acred, daddr_t albn, struct buf **abpp, int noLock) { afs_size_t totalLength; afs_size_t transferLength; afs_size_t filePos; afs_size_t offset, len, tlen; afs_int32 trimlen; struct dcache *tdc = 0; afs_int32 error, trybusy = 1; #ifdef AFS_DARWIN80_ENV uio_t tuiop = NULL; #else struct uio tuio; struct uio *tuiop = &tuio; struct iovec *tvec; #endif afs_int32 code; struct vrequest treq; AFS_STATCNT(afs_MemRead); if (avc->vc_error) return EIO; /* check that we have the latest status info in the vnode cache */ if ((code = afs_InitReq(&treq, acred))) return code; if (!noLock) { code = afs_VerifyVCache(avc, &treq); if (code) { code = afs_CheckCode(code, &treq, 8); /* failed to get it */ return code; } } #ifndef AFS_VM_RDWR_ENV if (AFS_NFSXLATORREQ(acred)) { if (!afs_AccessOK (avc, PRSFS_READ, &treq, CHECK_MODE_BITS | CMB_ALLOW_EXEC_AS_READ)) { return afs_CheckCode(EACCES, &treq, 9); } } #endif #ifndef AFS_DARWIN80_ENV tvec = (struct iovec *)osi_AllocSmallSpace(sizeof(struct iovec)); #endif totalLength = AFS_UIO_RESID(auio); filePos = AFS_UIO_OFFSET(auio); afs_Trace4(afs_iclSetp, CM_TRACE_READ, ICL_TYPE_POINTER, avc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_INT32, totalLength, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length)); error = 0; transferLength = 0; if (!noLock) ObtainReadLock(&avc->lock); #if defined(AFS_TEXT_ENV) && !defined(AFS_VM_RDWR_ENV) if (avc->flushDV.high == AFS_MAXDV && avc->flushDV.low == AFS_MAXDV) { hset(avc->flushDV, avc->f.m.DataVersion); } #endif /* * Locks held: * avc->lock(R) */ if (filePos >= avc->f.m.Length) { if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ len = 0; #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); } while (avc->f.m.Length > 0 && totalLength > 0) { /* read all of the cached info */ if (filePos >= avc->f.m.Length) break; /* all done */ if (noLock) { if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); } tdc = afs_FindDCache(avc, filePos); if (tdc) { ObtainReadLock(&tdc->lock); offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); len = tdc->f.chunkBytes - offset; } } else { /* a tricky question: does the presence of the DFFetching flag * mean that we're fetching the latest version of the file? No. * The server could update the file as soon as the fetch responsible * for the setting of the DFFetching flag completes. * * However, the presence of the DFFetching flag (visible under * a dcache read lock since it is set and cleared only under a * dcache write lock) means that we're fetching as good a version * as was known to this client at the time of the last call to * afs_VerifyVCache, since the latter updates the stat cache's * m.DataVersion field under a vcache write lock, and from the * time that the DFFetching flag goes on in afs_GetDCache (before * the fetch starts), to the time it goes off (after the fetch * completes), afs_GetDCache keeps at least a read lock on the * vcache entry. * * This means that if the DFFetching flag is set, we can use that * data for any reads that must come from the current version of * the file (current == m.DataVersion). * * Another way of looking at this same point is this: if we're * fetching some data and then try do an afs_VerifyVCache, the * VerifyVCache operation will not complete until after the * DFFetching flag is turned off and the dcache entry's f.versionNo * field is updated. * * Note, by the way, that if DFFetching is set, * m.DataVersion > f.versionNo (the latter is not updated until * after the fetch completes). */ if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); /* before reusing tdc */ } tdc = afs_GetDCache(avc, filePos, &treq, &offset, &len, 2); ObtainReadLock(&tdc->lock); /* now, first try to start transfer, if we'll need the data. If * data already coming, we don't need to do this, obviously. Type * 2 requests never return a null dcache entry, btw. */ if (!(tdc->dflags & DFFetching) && !hsame(avc->f.m.DataVersion, tdc->f.versionNo)) { /* have cache entry, it is not coming in now, * and we'll need new data */ tagain: if (trybusy && !afs_BBusy()) { struct brequest *bp; /* daemon is not busy */ ObtainSharedLock(&tdc->mflock, 665); if (!(tdc->mflags & DFFetchReq)) { /* start the daemon (may already be running, however) */ UpgradeSToWLock(&tdc->mflock, 666); tdc->mflags |= DFFetchReq; bp = afs_BQueue(BOP_FETCH, avc, B_DONTWAIT, 0, acred, (afs_size_t) filePos, (afs_size_t) 0, tdc); if (!bp) { tdc->mflags &= ~DFFetchReq; trybusy = 0; /* Avoid bkg daemon since they're too busy */ ReleaseWriteLock(&tdc->mflock); goto tagain; } ConvertWToSLock(&tdc->mflock); /* don't use bp pointer! */ } code = 0; ConvertSToRLock(&tdc->mflock); while (!code && tdc->mflags & DFFetchReq) { afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); /* don't need waiting flag on this one */ ReleaseReadLock(&tdc->mflock); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); ObtainReadLock(&tdc->mflock); } ReleaseReadLock(&tdc->mflock); if (code) { error = code; break; } } } /* now data may have started flowing in (if DFFetching is on). If * data is now streaming in, then wait for some interesting stuff. */ code = 0; while (!code && (tdc->dflags & DFFetching) && tdc->validPos <= filePos) { /* too early: wait for DFFetching flag to vanish, * or data to appear */ afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); } if (code) { error = code; break; } /* fetching flag gone, data is here, or we never tried * (BBusy for instance) */ if (tdc->dflags & DFFetching) { /* still fetching, some new data is here: * compute length and offset */ offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); len = tdc->validPos - filePos; } else { /* no longer fetching, verify data version * (avoid new GetDCache call) */ if (hsame(avc->f.m.DataVersion, tdc->f.versionNo) && ((len = tdc->validPos - filePos) > 0)) { offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); } else { /* don't have current data, so get it below */ afs_Trace3(afs_iclSetp, CM_TRACE_VERSIONNO, ICL_TYPE_INT64, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_HYPER, &avc->f.m.DataVersion, ICL_TYPE_HYPER, &tdc->f.versionNo); ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); tdc = NULL; } } if (!tdc) { /* If we get, it was not possible to start the * background daemon. With flag == 1 afs_GetDCache * does the FetchData rpc synchronously. */ ReleaseReadLock(&avc->lock); tdc = afs_GetDCache(avc, filePos, &treq, &offset, &len, 1); ObtainReadLock(&avc->lock); if (tdc) ObtainReadLock(&tdc->lock); } } afs_Trace3(afs_iclSetp, CM_TRACE_VNODEREAD, ICL_TYPE_POINTER, tdc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(offset), ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(len)); if (!tdc) { error = EIO; break; } /* * Locks held: * avc->lock(R) * tdc->lock(R) */ if (len > totalLength) len = totalLength; /* will read len bytes */ if (len <= 0) { /* shouldn't get here if DFFetching is on */ /* read past the end of a chunk, may not be at next chunk yet, and yet * also not at eof, so may have to supply fake zeros */ len = AFS_CHUNKTOSIZE(tdc->f.chunk) - offset; /* bytes left in chunk addr space */ if (len > totalLength) len = totalLength; /* and still within xfr request */ tlen = avc->f.m.Length - offset; /* and still within file */ if (len > tlen) len = tlen; if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); if (code) { error = code; break; } } else { /* get the data from the mem cache */ /* mung uio structure to be right for this transfer */ #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); uio_setoffset(tuiop, offset); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); tuio.afsio_offset = offset; #endif code = afs_MemReadUIO(tdc->f.inode, tuiop); if (code) { error = code; break; } } /* otherwise we've read some, fixup length, etc and continue with next seg */ len = len - AFS_UIO_RESID(tuiop); /* compute amount really transferred */ trimlen = len; afsio_skip(auio, trimlen); /* update input uio structure */ totalLength -= len; transferLength += len; filePos += len; if (len <= 0) break; /* surprise eof */ #ifdef AFS_DARWIN80_ENV if (tuiop) { uio_free(tuiop); tuiop = 0; } #endif } /* the whole while loop */ /* * Locks held: * avc->lock(R) * tdc->lock(R) if tdc */ /* if we make it here with tdc non-zero, then it is the last chunk we * dealt with, and we have to release it when we're done. We hold on * to it in case we need to do a prefetch. */ if (tdc) { ReleaseReadLock(&tdc->lock); /* try to queue prefetch, if needed */ if (!noLock && #ifndef AFS_VM_RDWR_ENV afs_preCache #else 1 #endif ) { afs_PrefetchChunk(avc, tdc, acred, &treq); } afs_PutDCache(tdc); } if (!noLock) ReleaseReadLock(&avc->lock); #ifdef AFS_DARWIN80_ENV if (tuiop) uio_free(tuiop); #else osi_FreeSmallSpace(tvec); #endif error = afs_CheckCode(error, &treq, 10); return error; }
int afs_UFSRead(register struct vcache *avc, struct uio *auio, struct AFS_UCRED *acred, daddr_t albn, struct buf **abpp, int noLock) { afs_size_t totalLength; afs_size_t transferLength; afs_size_t filePos; afs_size_t offset, len, tlen; afs_int32 trimlen; struct dcache *tdc = 0; afs_int32 error; #ifdef AFS_DARWIN80_ENV uio_t tuiop=NULL; #else struct uio tuio; struct uio *tuiop = &tuio; struct iovec *tvec; #endif struct osi_file *tfile; afs_int32 code; int trybusy = 1; struct vrequest treq; AFS_STATCNT(afs_UFSRead); if (avc && avc->vc_error) return EIO; AFS_DISCON_LOCK(); /* check that we have the latest status info in the vnode cache */ if ((code = afs_InitReq(&treq, acred))) return code; if (!noLock) { if (!avc) osi_Panic("null avc in afs_UFSRead"); else { code = afs_VerifyVCache(avc, &treq); if (code) { code = afs_CheckCode(code, &treq, 11); /* failed to get it */ AFS_DISCON_UNLOCK(); return code; } } } #ifndef AFS_VM_RDWR_ENV if (AFS_NFSXLATORREQ(acred)) { if (!afs_AccessOK (avc, PRSFS_READ, &treq, CHECK_MODE_BITS | CMB_ALLOW_EXEC_AS_READ)) { AFS_DISCON_UNLOCK(); return afs_CheckCode(EACCES, &treq, 12); } } #endif #ifndef AFS_DARWIN80_ENV tvec = (struct iovec *)osi_AllocSmallSpace(sizeof(struct iovec)); #endif totalLength = AFS_UIO_RESID(auio); filePos = AFS_UIO_OFFSET(auio); afs_Trace4(afs_iclSetp, CM_TRACE_READ, ICL_TYPE_POINTER, avc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_INT32, totalLength, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length)); error = 0; transferLength = 0; if (!noLock) ObtainReadLock(&avc->lock); #if defined(AFS_TEXT_ENV) && !defined(AFS_VM_RDWR_ENV) if (avc->flushDV.high == AFS_MAXDV && avc->flushDV.low == AFS_MAXDV) { hset(avc->flushDV, avc->f.m.DataVersion); } #endif if (filePos >= avc->f.m.Length) { if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ len = 0; #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); } while (avc->f.m.Length > 0 && totalLength > 0) { /* read all of the cached info */ if (filePos >= avc->f.m.Length) break; /* all done */ if (noLock) { if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); } tdc = afs_FindDCache(avc, filePos); if (tdc) { ObtainReadLock(&tdc->lock); offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); len = tdc->validPos - filePos; } } else { /* a tricky question: does the presence of the DFFetching flag * mean that we're fetching the latest version of the file? No. * The server could update the file as soon as the fetch responsible * for the setting of the DFFetching flag completes. * * However, the presence of the DFFetching flag (visible under * a dcache read lock since it is set and cleared only under a * dcache write lock) means that we're fetching as good a version * as was known to this client at the time of the last call to * afs_VerifyVCache, since the latter updates the stat cache's * m.DataVersion field under a vcache write lock, and from the * time that the DFFetching flag goes on in afs_GetDCache (before * the fetch starts), to the time it goes off (after the fetch * completes), afs_GetDCache keeps at least a read lock on the * vcache entry. * * This means that if the DFFetching flag is set, we can use that * data for any reads that must come from the current version of * the file (current == m.DataVersion). * * Another way of looking at this same point is this: if we're * fetching some data and then try do an afs_VerifyVCache, the * VerifyVCache operation will not complete until after the * DFFetching flag is turned off and the dcache entry's f.versionNo * field is updated. * * Note, by the way, that if DFFetching is set, * m.DataVersion > f.versionNo (the latter is not updated until * after the fetch completes). */ if (tdc) { ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); /* before reusing tdc */ } tdc = afs_GetDCache(avc, filePos, &treq, &offset, &len, 2); #ifdef AFS_DISCON_ENV if (!tdc) { printf("Network down in afs_read"); error = ENETDOWN; break; } #endif /* AFS_DISCON_ENV */ ObtainReadLock(&tdc->lock); /* now, first try to start transfer, if we'll need the data. If * data already coming, we don't need to do this, obviously. Type * 2 requests never return a null dcache entry, btw. */ if (!(tdc->dflags & DFFetching) && !hsame(avc->f.m.DataVersion, tdc->f.versionNo)) { /* have cache entry, it is not coming in now, and we'll need new data */ tagain: if (trybusy && !afs_BBusy()) { struct brequest *bp; /* daemon is not busy */ ObtainSharedLock(&tdc->mflock, 667); if (!(tdc->mflags & DFFetchReq)) { UpgradeSToWLock(&tdc->mflock, 668); tdc->mflags |= DFFetchReq; bp = afs_BQueue(BOP_FETCH, avc, B_DONTWAIT, 0, acred, (afs_size_t) filePos, (afs_size_t) 0, tdc); if (!bp) { /* Bkg table full; retry deadlocks */ tdc->mflags &= ~DFFetchReq; trybusy = 0; /* Avoid bkg daemon since they're too busy */ ReleaseWriteLock(&tdc->mflock); goto tagain; } ConvertWToSLock(&tdc->mflock); } code = 0; ConvertSToRLock(&tdc->mflock); while (!code && tdc->mflags & DFFetchReq) { afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); /* don't need waiting flag on this one */ ReleaseReadLock(&tdc->mflock); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); ObtainReadLock(&tdc->mflock); } ReleaseReadLock(&tdc->mflock); if (code) { error = code; break; } } } /* now data may have started flowing in (if DFFetching is on). If * data is now streaming in, then wait for some interesting stuff. */ code = 0; while (!code && (tdc->dflags & DFFetching) && tdc->validPos <= filePos) { /* too early: wait for DFFetching flag to vanish, * or data to appear */ afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAIT, ICL_TYPE_STRING, __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32, tdc->dflags); ReleaseReadLock(&tdc->lock); ReleaseReadLock(&avc->lock); code = afs_osi_SleepSig(&tdc->validPos); ObtainReadLock(&avc->lock); ObtainReadLock(&tdc->lock); } if (code) { error = code; break; } /* fetching flag gone, data is here, or we never tried * (BBusy for instance) */ if (tdc->dflags & DFFetching) { /* still fetching, some new data is here: * compute length and offset */ offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); len = tdc->validPos - filePos; } else { /* no longer fetching, verify data version (avoid new * GetDCache call) */ if (hsame(avc->f.m.DataVersion, tdc->f.versionNo) && ((len = tdc->validPos - filePos) > 0)) { offset = filePos - AFS_CHUNKTOBASE(tdc->f.chunk); } else { /* don't have current data, so get it below */ afs_Trace3(afs_iclSetp, CM_TRACE_VERSIONNO, ICL_TYPE_INT64, ICL_HANDLE_OFFSET(filePos), ICL_TYPE_HYPER, &avc->f.m.DataVersion, ICL_TYPE_HYPER, &tdc->f.versionNo); ReleaseReadLock(&tdc->lock); afs_PutDCache(tdc); tdc = NULL; } } if (!tdc) { /* If we get, it was not possible to start the * background daemon. With flag == 1 afs_GetDCache * does the FetchData rpc synchronously. */ ReleaseReadLock(&avc->lock); tdc = afs_GetDCache(avc, filePos, &treq, &offset, &len, 1); ObtainReadLock(&avc->lock); if (tdc) ObtainReadLock(&tdc->lock); } } if (!tdc) { error = EIO; break; } len = tdc->validPos - filePos; afs_Trace3(afs_iclSetp, CM_TRACE_VNODEREAD, ICL_TYPE_POINTER, tdc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(offset), ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(len)); if (len > totalLength) len = totalLength; /* will read len bytes */ if (len <= 0) { /* shouldn't get here if DFFetching is on */ afs_Trace4(afs_iclSetp, CM_TRACE_VNODEREAD2, ICL_TYPE_POINTER, tdc, ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(tdc->validPos), ICL_TYPE_INT32, tdc->f.chunkBytes, ICL_TYPE_INT32, tdc->dflags); /* read past the end of a chunk, may not be at next chunk yet, and yet * also not at eof, so may have to supply fake zeros */ len = AFS_CHUNKTOSIZE(tdc->f.chunk) - offset; /* bytes left in chunk addr space */ if (len > totalLength) len = totalLength; /* and still within xfr request */ tlen = avc->f.m.Length - offset; /* and still within file */ if (len > tlen) len = tlen; if (len > AFS_ZEROS) len = sizeof(afs_zeros); /* and in 0 buffer */ #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); #else afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); #endif AFS_UIOMOVE(afs_zeros, trimlen, UIO_READ, tuiop, code); if (code) { error = code; break; } } else { /* get the data from the file */ #ifdef IHINT if (tfile = tdc->ihint) { if (tdc->f.inode != tfile->inum) { afs_warn("afs_UFSRead: %x hint mismatch tdc %d inum %d\n", tdc, tdc->f.inode, tfile->inum); osi_UFSClose(tfile); tdc->ihint = tfile = 0; nihints--; } } if (tfile != 0) { usedihint++; } else #endif /* IHINT */ #if defined(LINUX_USE_FH) tfile = (struct osi_file *)osi_UFSOpen_fh(&tdc->f.fh, tdc->f.fh_type); #else tfile = (struct osi_file *)osi_UFSOpen(tdc->f.inode); #endif #ifdef AFS_DARWIN80_ENV trimlen = len; tuiop = afsio_darwin_partialcopy(auio, trimlen); uio_setoffset(tuiop, offset); #else /* mung uio structure to be right for this transfer */ afsio_copy(auio, &tuio, tvec); trimlen = len; afsio_trim(&tuio, trimlen); tuio.afsio_offset = offset; #endif #if defined(AFS_AIX41_ENV) AFS_GUNLOCK(); code = VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, &tuio, NULL, NULL, NULL, afs_osi_credp); AFS_GLOCK(); #elif defined(AFS_AIX32_ENV) code = VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, &tuio, NULL, NULL); /* Flush all JFS pages now for big performance gain in big file cases * If we do something like this, must check to be sure that AFS file * isn't mmapped... see afs_gn_map() for why. */ /* if (tfile->vnode->v_gnode && tfile->vnode->v_gnode->gn_seg) { many different ways to do similar things: so far, the best performing one is #2, but #1 might match it if we straighten out the confusion regarding which pages to flush. It really does matter. 1. vm_flushp(tfile->vnode->v_gnode->gn_seg, 0, len/PAGESIZE - 1); 2. vm_releasep(tfile->vnode->v_gnode->gn_seg, offset/PAGESIZE, (len + PAGESIZE-1)/PAGESIZE); 3. vms_inactive(tfile->vnode->v_gnode->gn_seg) Doesn't work correctly 4. vms_delete(tfile->vnode->v_gnode->gn_seg) probably also fails tfile->vnode->v_gnode->gn_seg = NULL; 5. deletep 6. ipgrlse 7. ifreeseg Unfortunately, this seems to cause frequent "cache corruption" episodes. vm_releasep(tfile->vnode->v_gnode->gn_seg, offset/PAGESIZE, (len + PAGESIZE-1)/PAGESIZE); } */ #elif defined(AFS_AIX_ENV) code = VNOP_RDWR(tfile->vnode, UIO_READ, FREAD, (off_t) & offset, &tuio, NULL, NULL, -1); #elif defined(AFS_SUN5_ENV) AFS_GUNLOCK(); #ifdef AFS_SUN510_ENV { caller_context_t ct; VOP_RWLOCK(tfile->vnode, 0, &ct); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp, &ct); VOP_RWUNLOCK(tfile->vnode, 0, &ct); } #else VOP_RWLOCK(tfile->vnode, 0); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_RWUNLOCK(tfile->vnode, 0); #endif AFS_GLOCK(); #elif defined(AFS_SGI_ENV) AFS_GUNLOCK(); AFS_VOP_RWLOCK(tfile->vnode, VRWLOCK_READ); AFS_VOP_READ(tfile->vnode, &tuio, IO_ISLOCKED, afs_osi_credp, code); AFS_VOP_RWUNLOCK(tfile->vnode, VRWLOCK_READ); AFS_GLOCK(); #elif defined(AFS_OSF_ENV) tuio.uio_rw = UIO_READ; AFS_GUNLOCK(); VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp, code); AFS_GLOCK(); #elif defined(AFS_HPUX100_ENV) AFS_GUNLOCK(); code = VOP_RDWR(tfile->vnode, &tuio, UIO_READ, 0, afs_osi_credp); AFS_GLOCK(); #elif defined(AFS_LINUX20_ENV) AFS_GUNLOCK(); code = osi_rdwr(tfile, &tuio, UIO_READ); AFS_GLOCK(); #elif defined(AFS_DARWIN80_ENV) AFS_GUNLOCK(); code = VNOP_READ(tfile->vnode, tuiop, 0, afs_osi_ctxtp); AFS_GLOCK(); #elif defined(AFS_DARWIN_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, current_proc()); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0, current_proc()); AFS_GLOCK(); #elif defined(AFS_FBSD80_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0); AFS_GLOCK(); #elif defined(AFS_FBSD50_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, curthread); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0, curthread); AFS_GLOCK(); #elif defined(AFS_XBSD_ENV) AFS_GUNLOCK(); VOP_LOCK(tfile->vnode, LK_EXCLUSIVE, curproc); code = VOP_READ(tfile->vnode, &tuio, 0, afs_osi_credp); VOP_UNLOCK(tfile->vnode, 0, curproc); AFS_GLOCK(); #else code = VOP_RDWR(tfile->vnode, &tuio, UIO_READ, 0, afs_osi_credp); #endif #ifdef IHINT if (!tdc->ihint && nihints < maxIHint) { tdc->ihint = tfile; nihints++; } else #endif /* IHINT */ osi_UFSClose(tfile); if (code) { error = code; break; } } /* otherwise we've read some, fixup length, etc and continue with next seg */ len = len - AFS_UIO_RESID(tuiop); /* compute amount really transferred */ trimlen = len; afsio_skip(auio, trimlen); /* update input uio structure */ totalLength -= len; transferLength += len; filePos += len; if (len <= 0) break; /* surprise eof */ #ifdef AFS_DARWIN80_ENV if (tuiop) { uio_free(tuiop); tuiop = 0; } #endif } /* if we make it here with tdc non-zero, then it is the last chunk we * dealt with, and we have to release it when we're done. We hold on * to it in case we need to do a prefetch, obviously. */ if (tdc) { ReleaseReadLock(&tdc->lock); #if !defined(AFS_VM_RDWR_ENV) /* try to queue prefetch, if needed */ if (!noLock) { if (!(tdc->mflags & DFNextStarted)) afs_PrefetchChunk(avc, tdc, acred, &treq); } #endif afs_PutDCache(tdc); } if (!noLock) ReleaseReadLock(&avc->lock); #ifdef AFS_DARWIN80_ENV if (tuiop) uio_free(tuiop); #else osi_FreeSmallSpace(tvec); #endif AFS_DISCON_UNLOCK(); error = afs_CheckCode(error, &treq, 13); return error; }
afs_link(struct vcache *avc, OSI_VC_DECL(adp), char *aname, afs_ucred_t *acred) #endif { struct vrequest treq; struct dcache *tdc; afs_int32 code; struct afs_conn *tc; afs_size_t offset, len; struct AFSFetchStatus OutFidStatus, OutDirStatus; struct AFSVolSync tsync; struct afs_fakestat_state vfakestate, dfakestate; struct rx_connection *rxconn; XSTATS_DECLS; OSI_VC_CONVERT(adp); AFS_STATCNT(afs_link); afs_Trace3(afs_iclSetp, CM_TRACE_LINK, ICL_TYPE_POINTER, adp, ICL_TYPE_POINTER, avc, ICL_TYPE_STRING, aname); /* create a hard link; new entry is aname in dir adp */ if ((code = afs_InitReq(&treq, acred))) goto done2; afs_InitFakeStat(&vfakestate); afs_InitFakeStat(&dfakestate); AFS_DISCON_LOCK(); code = afs_EvalFakeStat(&avc, &vfakestate, &treq); if (code) goto done; code = afs_EvalFakeStat(&adp, &dfakestate, &treq); if (code) goto done; if (avc->f.fid.Cell != adp->f.fid.Cell || avc->f.fid.Fid.Volume != adp->f.fid.Fid.Volume) { code = EXDEV; goto done; } if (strlen(aname) > AFSNAMEMAX) { code = ENAMETOOLONG; goto done; } 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) { code = ENETDOWN; goto done; } tdc = afs_GetDCache(adp, (afs_size_t) 0, &treq, &offset, &len, 1); /* test for error below */ ObtainWriteLock(&adp->lock, 145); do { tc = afs_Conn(&adp->f.fid, &treq, SHARED_LOCK, &rxconn); if (tc) { XSTATS_START_TIME(AFS_STATS_FS_RPCIDX_LINK); RX_AFS_GUNLOCK(); code = RXAFS_Link(rxconn, (struct AFSFid *)&adp->f.fid.Fid, aname, (struct AFSFid *)&avc->f.fid.Fid, &OutFidStatus, &OutDirStatus, &tsync); RX_AFS_GLOCK(); XSTATS_END_TIME; } else code = -1; } while (afs_Analyze (tc, rxconn, code, &adp->f.fid, &treq, AFS_STATS_FS_RPCIDX_LINK, SHARED_LOCK, NULL)); if (code) { if (tdc) afs_PutDCache(tdc); if (code < 0) { ObtainWriteLock(&afs_xcbhash, 492); afs_DequeueCallback(adp); adp->f.states &= ~CStatd; ReleaseWriteLock(&afs_xcbhash); osi_dnlc_purgedp(adp); } ReleaseWriteLock(&adp->lock); goto done; } if (tdc) ObtainWriteLock(&tdc->lock, 635); if (afs_LocalHero(adp, tdc, &OutDirStatus, 1)) { /* we can do it locally */ ObtainWriteLock(&afs_xdcache, 290); code = afs_dir_Create(tdc, aname, &avc->f.fid.Fid); ReleaseWriteLock(&afs_xdcache); if (code) { ZapDCE(tdc); /* surprise error -- invalid value */ DZap(tdc); } } if (tdc) { ReleaseWriteLock(&tdc->lock); afs_PutDCache(tdc); /* drop ref count */ } ReleaseWriteLock(&adp->lock); ObtainWriteLock(&avc->lock, 146); /* correct link count */ /* we could lock both dir and file; since we get the new fid * status back, you'd think we could put it in the cache status * entry at that point. Note that if we don't lock the file over * the rpc call, we have no guarantee that the status info * returned in ustat is the most recent to store in the file's * cache entry */ ObtainWriteLock(&afs_xcbhash, 493); afs_DequeueCallback(avc); avc->f.states &= ~CStatd; /* don't really know new link count */ ReleaseWriteLock(&afs_xcbhash); if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR)) osi_dnlc_purgedp(avc); ReleaseWriteLock(&avc->lock); code = 0; done: code = afs_CheckCode(code, &treq, 24); afs_PutFakeStat(&vfakestate); afs_PutFakeStat(&dfakestate); AFS_DISCON_UNLOCK(); done2: return code; }
/* 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); } }
static int ClearCallBack(struct rx_connection *a_conn, struct AFSFid *a_fid) { struct vcache *tvc; int i; struct VenusFid localFid; struct volume *tv; #ifdef AFS_DARWIN80_ENV vnode_t vp; #endif AFS_STATCNT(ClearCallBack); AFS_ASSERT_GLOCK(); /* * XXXX Don't hold any server locks here because of callback protocol XXX */ localFid.Cell = 0; localFid.Fid.Volume = a_fid->Volume; localFid.Fid.Vnode = a_fid->Vnode; localFid.Fid.Unique = a_fid->Unique; /* * Volume ID of zero means don't do anything. */ if (a_fid->Volume != 0) { if (a_fid->Vnode == 0) { struct afs_q *tq, *uq; /* * Clear callback for the whole volume. Zip through the * hash chain, nullifying entries whose volume ID matches. */ loop1: ObtainReadLock(&afs_xvcache); i = VCHashV(&localFid); for (tq = afs_vhashTV[i].prev; tq != &afs_vhashTV[i]; tq = uq) { uq = QPrev(tq); tvc = QTOVH(tq); if (tvc->f.fid.Fid.Volume == a_fid->Volume) { tvc->callback = NULL; if (!localFid.Cell) localFid.Cell = tvc->f.fid.Cell; tvc->dchint = NULL; /* invalidate hints */ if (tvc->f.states & CVInit) { ReleaseReadLock(&afs_xvcache); afs_osi_Sleep(&tvc->f.states); goto loop1; } #if defined(AFS_SGI_ENV) || defined(AFS_SUN5_ENV) || defined(AFS_HPUX_ENV) || defined(AFS_LINUX20_ENV) AFS_FAST_HOLD(tvc); #else #ifdef AFS_DARWIN80_ENV if (tvc->f.states & CDeadVnode) { if (!(tvc->f.states & CBulkFetching)) { ReleaseReadLock(&afs_xvcache); afs_osi_Sleep(&tvc->f.states); goto loop1; } } vp = AFSTOV(tvc); if (vnode_get(vp)) continue; if (vnode_ref(vp)) { AFS_GUNLOCK(); vnode_put(vp); AFS_GLOCK(); continue; } if (tvc->f.states & (CBulkFetching|CDeadVnode)) { AFS_GUNLOCK(); vnode_recycle(AFSTOV(tvc)); AFS_GLOCK(); } #else AFS_FAST_HOLD(tvc); #endif #endif ReleaseReadLock(&afs_xvcache); ObtainWriteLock(&afs_xcbhash, 449); afs_DequeueCallback(tvc); tvc->f.states &= ~(CStatd | CUnique | CBulkFetching); afs_allCBs++; if (tvc->f.fid.Fid.Vnode & 1) afs_oddCBs++; else afs_evenCBs++; ReleaseWriteLock(&afs_xcbhash); if ((tvc->f.fid.Fid.Vnode & 1 || (vType(tvc) == VDIR))) osi_dnlc_purgedp(tvc); afs_Trace3(afs_iclSetp, CM_TRACE_CALLBACK, ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, tvc->f.states, ICL_TYPE_INT32, a_fid->Volume); #ifdef AFS_DARWIN80_ENV vnode_put(AFSTOV(tvc)); #endif ObtainReadLock(&afs_xvcache); uq = QPrev(tq); AFS_FAST_RELE(tvc); } else if ((tvc->f.states & CMValid) && (tvc->mvid->Fid.Volume == a_fid->Volume)) { tvc->f.states &= ~CMValid; if (!localFid.Cell) localFid.Cell = tvc->mvid->Cell; } } ReleaseReadLock(&afs_xvcache); /* * XXXX Don't hold any locks here XXXX */ tv = afs_FindVolume(&localFid, 0); if (tv) { afs_ResetVolumeInfo(tv); afs_PutVolume(tv, 0); /* invalidate mtpoint? */ } } /*Clear callbacks for whole volume */ else { /* * Clear callbacks just for the one file. */ struct vcache *uvc; afs_allCBs++; if (a_fid->Vnode & 1) afs_oddCBs++; /*Could do this on volume basis, too */ else afs_evenCBs++; /*A particular fid was specified */ loop2: ObtainReadLock(&afs_xvcache); i = VCHash(&localFid); for (tvc = afs_vhashT[i]; tvc; tvc = uvc) { uvc = tvc->hnext; if (tvc->f.fid.Fid.Vnode == a_fid->Vnode && tvc->f.fid.Fid.Volume == a_fid->Volume && tvc->f.fid.Fid.Unique == a_fid->Unique) { tvc->callback = NULL; tvc->dchint = NULL; /* invalidate hints */ if (tvc->f.states & CVInit) { ReleaseReadLock(&afs_xvcache); afs_osi_Sleep(&tvc->f.states); goto loop2; } #if defined(AFS_SGI_ENV) || defined(AFS_SUN5_ENV) || defined(AFS_HPUX_ENV) || defined(AFS_LINUX20_ENV) AFS_FAST_HOLD(tvc); #else #ifdef AFS_DARWIN80_ENV if (tvc->f.states & CDeadVnode) { if (!(tvc->f.states & CBulkFetching)) { ReleaseReadLock(&afs_xvcache); afs_osi_Sleep(&tvc->f.states); goto loop2; } } vp = AFSTOV(tvc); if (vnode_get(vp)) continue; if (vnode_ref(vp)) { AFS_GUNLOCK(); vnode_put(vp); AFS_GLOCK(); continue; } if (tvc->f.states & (CBulkFetching|CDeadVnode)) { AFS_GUNLOCK(); vnode_recycle(AFSTOV(tvc)); AFS_GLOCK(); } #else AFS_FAST_HOLD(tvc); #endif #endif ReleaseReadLock(&afs_xvcache); ObtainWriteLock(&afs_xcbhash, 450); afs_DequeueCallback(tvc); tvc->f.states &= ~(CStatd | CUnique | CBulkFetching); ReleaseWriteLock(&afs_xcbhash); if ((tvc->f.fid.Fid.Vnode & 1 || (vType(tvc) == VDIR))) osi_dnlc_purgedp(tvc); afs_Trace3(afs_iclSetp, CM_TRACE_CALLBACK, ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, tvc->f.states, ICL_TYPE_LONG, 0); #ifdef CBDEBUG lastCallBack_vnode = afid->Vnode; lastCallBack_dv = tvc->mstat.DataVersion.low; osi_GetuTime(&lastCallBack_time); #endif /* CBDEBUG */ #ifdef AFS_DARWIN80_ENV vnode_put(AFSTOV(tvc)); #endif ObtainReadLock(&afs_xvcache); uvc = tvc->hnext; AFS_FAST_RELE(tvc); } } /*Walk through hash table */ ReleaseReadLock(&afs_xvcache); } /*Clear callbacks for one file */ } /*Fid has non-zero volume ID */ /* * Always return a predictable value. */ return (0); } /*ClearCallBack */