if ( tsp ) /* an existing server - ref count increased */
{
lock_ObtainMutex(&tsp->mx);
- tsp->ipRank = (USHORT)dwRank;
+ tsp->adminRank = (USHORT)dwRank;
_InterlockedOr(&tsp->flags, CM_SERVERFLAG_PREF_SET);
- tsp->adminRank = tsp->ipRank;
+ cm_RankServer(tsp);
lock_ReleaseMutex(&tsp->mx);
/* set preferences for an existing vlserver */
{
tsp = cm_NewServer(&saddr, CM_SERVER_VLDB, NULL, NULL, CM_FLAG_NOPROBE); /* refcount = 1 */
lock_ObtainMutex(&tsp->mx);
- tsp->ipRank = (USHORT)dwRank;
+ tsp->adminRank = (USHORT)dwRank;
_InterlockedOr(&tsp->flags, CM_SERVERFLAG_PREF_SET);
- tsp->adminRank = tsp->ipRank;
+ cm_RankServer(tsp);
lock_ReleaseMutex(&tsp->mx);
}
}
if ( tsp ) /* an existing server - ref count increased */
{
lock_ObtainMutex(&tsp->mx);
- tsp->ipRank = (USHORT)dwRank;
+ tsp->adminRank = (USHORT)dwRank;
_InterlockedOr(&tsp->flags, CM_SERVERFLAG_PREF_SET);
- tsp->adminRank = tsp->ipRank;
+ cm_RankServer(tsp);
lock_ReleaseMutex(&tsp->mx);
/* find volumes which might have RO copy
{
tsp = cm_NewServer(&saddr, CM_SERVER_FILE, NULL, NULL, CM_FLAG_NOPROBE); /* refcount = 1 */
lock_ObtainMutex(&tsp->mx);
- tsp->ipRank = (USHORT)dwRank;
+ tsp->adminRank = (USHORT)dwRank;
_InterlockedOr(&tsp->flags, CM_SERVERFLAG_PREF_SET);
- tsp->adminRank = tsp->ipRank;
+ cm_RankServer(tsp);
lock_ReleaseMutex(&tsp->mx);
}
}
*
* At the present time the return value is ignored by the caller.
*/
-long cm_AddCellProc(void *rockp, struct sockaddr_in *addrp, char *hostnamep, unsigned short ipRank)
+long cm_AddCellProc(void *rockp, struct sockaddr_in *addrp, char *hostnamep, unsigned short adminRank)
{
cm_server_t *tsp;
cm_serverRef_t *tsrp;
else
tsp = cm_NewServer(addrp, CM_SERVER_VLDB, cellp, NULL, probe ? 0 : CM_FLAG_NOPROBE);
- if (ipRank)
- tsp->ipRank = ipRank;
+ if (adminRank)
+ tsp->adminRank = adminRank;
/* Insert the vlserver into a sorted list, sorted by server rank */
tsrp = cm_NewServerRef(tsp, 0);
DWORD dwType, dwSize;
DWORD dwCells, dwServers, dwForceDNS;
DWORD dwIndex, dwRank, dwPort;
- unsigned short ipRank;
+ unsigned short adminRank;
unsigned short vlPort;
LONG code;
FILETIME ftLastWriteTime;
code = RegQueryValueEx(hkServerName, "Rank", NULL, &dwType,
(BYTE *) &dwRank, &dwSize);
if (code == ERROR_SUCCESS && dwType == REG_DWORD) {
- ipRank = (unsigned short)(dwRank <= 65535 ? dwRank : 65535);
+ adminRank = (unsigned short)(dwRank <= 65535 ? dwRank : 65535);
} else {
- ipRank = 0;
+ adminRank = 0;
}
dwSize = sizeof(szAddr);
vlSockAddr.sin_family = AF_INET;
/* sin_port supplied by connection code */
if (procp)
- (*procp)(rockp, &vlSockAddr, s, ipRank);
+ (*procp)(rockp, &vlSockAddr, s, adminRank);
} else if (szAddr[0]) {
afs_uint32 ip_addr;
unsigned int c1, c2, c3, c4;
vlSockAddr.sin_family = AF_INET;
/* sin_port supplied by connection code */
if (procp)
- (*procp)(rockp, &vlSockAddr, s, ipRank);
+ (*procp)(rockp, &vlSockAddr, s, adminRank);
}
}
int rc;
int cellHostAddrs[AFSMAXCELLHOSTS];
char cellHostNames[AFSMAXCELLHOSTS][MAXHOSTCHARS];
- unsigned short ipRanks[AFSMAXCELLHOSTS];
+ unsigned short adminRanks[AFSMAXCELLHOSTS];
unsigned short ports[AFSMAXCELLHOSTS]; /* network byte order */
int numServers;
int i;
return -1;
rc = getAFSServer("afs3-vlserver", "udp", cellNamep, htons(7003),
- cellHostAddrs, cellHostNames, ports, ipRanks, &numServers, ttl);
+ cellHostAddrs, cellHostNames, ports, adminRanks, &numServers, ttl);
if (rc == 0 && numServers > 0) { /* found the cell */
for (i = 0; i < numServers; i++) {
memcpy(&vlSockAddr.sin_addr.s_addr, &cellHostAddrs[i],
vlSockAddr.sin_port = ports[i];
vlSockAddr.sin_family = AF_INET;
if (procp)
- (*procp)(rockp, &vlSockAddr, cellHostNames[i], ipRanks[i]);
+ (*procp)(rockp, &vlSockAddr, cellHostNames[i], adminRanks[i]);
}
if (newCellNamep) {
if(FAILED(StringCchCopy(newCellNamep, CELL_MAXNAMELEN, cellNamep)))
};
void processReplyBuffer_AFSDB(SOCKET commSock, PDNS_HDR replyBuff, int *cellHostAddrs, char cellHostNames[][MAXHOSTCHARS],
- unsigned short ports[], unsigned short ipRanks[], int *numServers, int *ttl)
+ unsigned short ports[], unsigned short adminRanks[], int *numServers, int *ttl)
/*PAFS_SRV_LIST (srvList)*/
{
u_char *ptr = (u_char *) replyBuff;
memcpy(&cellHostAddrs[srvCount], &addr.s_addr, sizeof(addr.s_addr));
strncpy(cellHostNames[srvCount], hostName, CELL_MAXNAMELEN);
cellHostNames[srvCount][CELL_MAXNAMELEN-1] = '\0';
- ipRanks[srvCount] = 0;
+ adminRanks[srvCount] = 0;
ports[srvCount] = htons(7003);
srvCount++;
}
unsigned short afsdbPort, /* network byte order */
int *cellHostAddrs, char cellHostNames[][MAXHOSTCHARS],
unsigned short ports[], /* network byte order */
- unsigned short ipRanks[],
+ unsigned short adminRanks[],
int *numServers, int *ttl)
{
#ifndef DNSAPI_ENV
/*printReplyBuffer_AFSDB(pDNShdr);*/
if (pDNShdr)
- processReplyBuffer_AFSDB(commSock, pDNShdr, cellHostAddrs, cellHostNames, ports, ipRanks, numServers, ttl);
+ processReplyBuffer_AFSDB(commSock, pDNShdr, cellHostAddrs, cellHostNames, ports, adminRanks, numServers, ttl);
closesocket(commSock);
if (*numServers == 0)
if (pDnsIter->wType == DNS_TYPE_SRV) {
StringCbCopyA(cellHostNames[*numServers], sizeof(cellHostNames[*numServers]),
pDnsIter->Data.SRV.pNameTarget);
- ipRanks[*numServers] = pDnsIter->Data.SRV.wPriority;
+ adminRanks[*numServers] = pDnsIter->Data.SRV.wPriority;
ports[*numServers] = htons(pDnsIter->Data.SRV.wPort);
(*numServers)++;
if (pDnsIter->wType == DNS_TYPE_AFSDB && pDnsIter->Data.Afsdb.wPreference == 1) {
StringCbCopyA(cellHostNames[*numServers], sizeof(cellHostNames[*numServers]),
pDnsIter->Data.Afsdb.pNameExchange);
- ipRanks[*numServers] = 0;
+ adminRanks[*numServers] = 0;
ports[*numServers] = afsdbPort;
(*numServers)++;
int *cellHostAddrs,
cm_unichar_t cellHostNames[][MAXHOSTCHARS],
unsigned short ports[], /* network byte order */
- unsigned short ipRanks[],
+ unsigned short adminRanks[],
int *numServers, int *ttl)
{
#ifdef DNSAPI_ENV
if (pDnsIter->wType == DNS_TYPE_SRV) {
StringCbCopyW(cellHostNames[*numServers], sizeof(cellHostNames[*numServers]),
pDnsIter->Data.SRV.pNameTarget);
- ipRanks[*numServers] = pDnsIter->Data.SRV.wPriority;
+ adminRanks[*numServers] = pDnsIter->Data.SRV.wPriority;
ports[*numServers] = htons(pDnsIter->Data.SRV.wPort);
(*numServers)++;
if (pDnsIter->wType == DNS_TYPE_AFSDB && pDnsIter->Data.Afsdb.wPreference == 1) {
StringCbCopyW(cellHostNames[*numServers], sizeof(cellHostNames[*numServers]),
pDnsIter->Data.Afsdb.pNameExchange);
- ipRanks[*numServers] = 0;
+ adminRanks[*numServers] = 0;
ports[*numServers] = afsdbPort;
(*numServers)++;
if ( tsp ) /* an existing server - ref count increased */
{
lock_ObtainMutex(&tsp->mx);
- tsp->ipRank = rank;
+ tsp->adminRank = rank;
_InterlockedOr(&tsp->flags, CM_SERVERFLAG_PREF_SET);
- tsp->adminRank = tsp->ipRank;
lock_ReleaseMutex(&tsp->mx);
switch (type) {
{
tsp = cm_NewServer(&tmp, type, NULL, NULL, CM_FLAG_NOPROBE); /* refcount = 1 */
lock_ObtainMutex(&tsp->mx);
- tsp->ipRank = rank;
+ tsp->adminRank = rank;
_InterlockedOr(&tsp->flags, CM_SERVERFLAG_PREF_SET);
- tsp->adminRank = tsp->ipRank;
lock_ReleaseMutex(&tsp->mx);
- tsp->ipRank = rank;
}
cm_PutServer(tsp); /* decrease refcount */
}
continue; /* ignore vlservers */
srvout->host = tsp->addr.sin_addr;
- srvout->rank = tsp->ipRank;
+ srvout->rank = tsp->activeRank;
srvout++;
spout->num_servers++;
noServers--;
afs_int32 code = 0; /* start with "success" */
struct rx_debugPeer tpeer;
afs_uint16 port;
- afs_uint16 newRank;
+ afs_uint64 newRank;
+ afs_uint64 perfRank = 0;
+ afs_uint64 rtt;
+ double log_rtt;
+ int isDown = (tsp->flags & CM_SERVERFLAG_DOWN);
switch(tsp->type) {
case CM_SERVER_VLDB:
return -1;
}
- code = rx_GetLocalPeers(tsp->addr.sin_addr.s_addr, port, &tpeer);
+ cm_SetServerIPRank(tsp);
- /*check if rx_GetLocalPeers succeeded and if there is data for tsp */
- if(code == 0 && (tpeer.rtt == 0 && tpeer.rtt_dev == 0))
- code = -1;
+ if (isDown) {
+ newRank = 0xFFFF;
+ } else {
+ /*
+ * There are three potential components to the ranking:
+ * 1. Any administrative set preference whether it be
+ * via "fs setserverprefs", registry or dns.
+ *
+ * 2. Network subnet mask comparison.
+ *
+ * 3. Performance data.
+ *
+ * If there is an administrative rank, that is the
+ * the primary factor. If not the primary factor
+ * is the network ranking.
+ */
+
+ code = rx_GetLocalPeers(tsp->addr.sin_addr.s_addr, port, &tpeer);
+ if (code == 0) {
+ if (tpeer.rtt) {
+ /* rtt is ms/8 */
+ rtt = tpeer.rtt;
+ log_rtt = log(tpeer.rtt);
+ perfRank += (6000 * log_rtt / 5000) * 5000;
+
+ if (tsp->type == CM_SERVER_FILE) {
+ /* give an edge to servers with high congestion windows */
+ perfRank -= (tpeer.cwind - 1)* 15;
+ }
+ }
+ }
- if(code == 0) {
- if((tsp->flags & CM_SERVERFLAG_PREF_SET))
- newRank = tsp->adminRank +
- ((int)(623 * log(tpeer.rtt) / 10) * 10 + 5);
- else /* rank has not been set by admin, derive rank from rtt */
- newRank = (int)(7200 * log(tpeer.rtt) / 5000) * 5000 + 5000;
+ if (tsp->adminRank) {
+ newRank = tsp->adminRank * 0.8;
+ newRank += tsp->ipRank * 0.2;
+ } else {
+ newRank = tsp->ipRank;
+ }
+ if (perfRank) {
+ newRank *= 0.9;
+ newRank += perfRank * 0.1;
+ }
+ newRank += (rand() & 0x000f); /* randomize */
- newRank += (rand() & 0x000f); /* randomize */
+ if (newRank > 0xFFFF)
+ osi_Log1(afsd_logp, "new server rank %I64u exceeds 0xFFFF", newRank);
- if (abs(newRank - tsp->ipRank) > 0xf) {
- tsp->ipRank = newRank;
+ /*
+ * If the ranking changes by more than the randomization
+ * factor, update the server reference lists.
+ */
+ if (abs(newRank - tsp->activeRank) > 0xf) {
+ tsp->activeRank = newRank;
lock_ReleaseMutex(&tsp->mx);
switch (tsp->type) {
lock_ReleaseMutex(&serverp->mx);
}
-void cm_SetServerPrefs(cm_server_t * serverp)
+void cm_SetServerIPRank(cm_server_t * serverp)
{
unsigned long serverAddr; /* in host byte order */
unsigned long myAddr, myNet, mySubnet;/* in host byte order */
unsigned long netMask;
int i;
long code;
- int writeLock = 0;
lock_ObtainRead(&cm_syscfgLock);
if (cm_LanAdapterChangeDetected) {
lock_ConvertRToW(&cm_syscfgLock);
- writeLock = 1;
if (cm_LanAdapterChangeDetected) {
/* get network related info */
cm_noIPAddr = CM_MAXINTERFACE_ADDR;
{
if ( (serverAddr & cm_SubnetMask[i]) == mySubnet)
{
- if ( serverAddr == myAddr )
+ if ( serverAddr == myAddr ) {
serverp->ipRank = min(serverp->ipRank,
CM_IPRANK_TOP);/* same machine */
- else serverp->ipRank = min(serverp->ipRank,
- CM_IPRANK_HI); /* same subnet */
- }
- else serverp->ipRank = min(serverp->ipRank,CM_IPRANK_MED);
- /* same net */
+ } else {
+ serverp->ipRank = min(serverp->ipRank,
+ CM_IPRANK_HI); /* same subnet */
+ }
+ } else {
+ serverp->ipRank = min(serverp->ipRank, CM_IPRANK_MED); /* same net */
+ }
}
} /* and of for loop */
-
- /* random between 0..15*/
- serverp->ipRank += (rand() % 0x000f);
lock_ReleaseRead(&cm_syscfgLock);
}
lock_InitializeMutex(&tsp->mx, "cm_server_t mutex", LOCK_HIERARCHY_SERVER);
tsp->addr = *socketp;
- cm_SetServerPrefs(tsp);
+ cm_SetServerIPRank(tsp);
tsp->allNextp = cm_allServersp;
cm_allServersp = tsp;
void cm_InsertServerList(cm_serverRef_t** list, cm_serverRef_t* element)
{
cm_serverRef_t *current;
- unsigned short ipRank;
+ unsigned short rank;
lock_ObtainWrite(&cm_serverLock);
/*
goto done;
}
- ipRank = element->server->ipRank;
+ rank = element->server->activeRank;
/* insertion at the beginning of the list */
- if ((*list)->server->ipRank > ipRank)
+ if ((*list)->server->activeRank > rank)
{
element->next = *list;
*list = element;
/* find appropriate place to insert */
for ( current = *list; current->next; current = current->next)
{
- if ( current->next->server->ipRank > ipRank )
+ if ( current->next->server->activeRank > rank )
break;
}
element->next = current->next;
}
/* count the number of servers with the lowest rank */
- lowestRank = tsrp->server->ipRank;
+ lowestRank = tsrp->server->activeRank;
for ( count=1, tsrp=tsrp->next; tsrp; tsrp=tsrp->next)
{
- if ( tsrp->server->ipRank != lowestRank)
+ if ( tsrp->server->activeRank != lowestRank)
break;
else
count++;
"flags=0x%x waitCount=%u rank=%u downTime=\"%s\" refCount=%u\r\n",
cookie, tsp, tsp->cellp ? tsp->cellp->name : "", hoststr,
ntohs(tsp->addr.sin_port), uuidstr, type,
- tsp->capabilities, tsp->flags, tsp->waitCount, tsp->ipRank,
+ tsp->capabilities, tsp->flags, tsp->waitCount, tsp->activeRank,
(tsp->flags & CM_SERVERFLAG_DOWN) ? down : "up",
tsp->refCount);
WriteFile(outputFile, output, (DWORD)strlen(output), &zilch, NULL);
struct cm_cell *cellp; /* cell containing this server */
afs_int32 refCount; /* Interlocked with cm_serverLock */
osi_mutex_t mx;
- unsigned short ipRank; /* server priority */
+ unsigned short ipRank; /* network class rank */
+ unsigned short adminRank; /* set if admin sets a rank
+ * (fs setserverpref or registry or dns)
+ */
+ unsigned short activeRank; /* Computed rank combining ipRank, adminRank,
+ * and performance data.
+ */
cm_server_vols_t * vols; /* by mx */
time_t downTime; /* by mx */
afsUUID uuid; /* by mx */
- unsigned short adminRank; /* only set if admin sets a rank */
} cm_server_t;
enum repstate {srv_not_busy, srv_busy, srv_offline, srv_deleted};
#define CM_IPRANK_HI 20000 /* on same subnet */
#define CM_IPRANK_MED 30000 /* on same network */
#define CM_IPRANK_LOW 40000 /* on different networks */
+#define CM_IPRANK_DOWN 65535 /* unavailable */
/* the maximum number of network interfaces that this client has */
extern void cm_RankUpServers();
-extern void cm_SetServerPrefs(cm_server_t * serverp);
+extern void cm_SetServerIPRank(cm_server_t * serverp);
extern void cm_InsertServerList(cm_serverRef_t** list,cm_serverRef_t* element);
/*
* Randomize RO list
*
- * If the first n servers have the same ipRank, then we
+ * If the first n servers have the same rank, then we
* randomly pick one among them and move it to the beginning.
* We don't bother to re-order the whole list because
* the rest of the list is used only if the first server is
git://git.openafs.org / openafs.git / commitdiff