// . returns false on error
// . called from PageHosts.cpp!!!
bool Syncdb::syncHost ( long syncHostId ) {
Host *sh = g_hostdb.getHost ( syncHostId );
if ( ! sh ) return log("sync: bad host id %li",syncHostId);
// get its group
//Host *hosts = g_hostdb.getGroup ( sh->m_groupId );
Host *hosts = g_hostdb.getShard ( sh->m_shardNum );
// get the best twin for it to sync from
for ( long i = 0 ; i < g_hostdb.getNumHostsPerShard() ; i++ ) {
// get host
Host *h = &hosts[i];
// skip if dead
if ( g_hostdb.isDead ( h ) ) continue;
// skip if permanent out of sync
if ( h->m_isPermanentOutOfSync ) continue;
// not itself! it must be dead... wtf!?
if ( h == sh ) continue;
// save it
long tmp = syncHostId;
// log it
log("sync: sending sync request to host id #%li",h->m_hostId);
// shortcut
UdpServer *us = &g_udpServer;
// use that guy
if ( us->sendRequest ( (char *)&tmp ,
4 ,
0x55 , // SYNCDB REQUEST
0 , // ip
0 , // port
h->m_hostId , // hostId
NULL , // retSlot
NULL , // state
gotReplyWrapper55 , // wrapper
15 , // timeout
-1 , // backoff
-1 , // maxWait
NULL , // replyBuf
0 , // replyBufSize
MAX_NICENESS ))
// success
return true;
// note it
log("sync: had error sending sync request to host id #%li: %s",
h->m_hostId,mstrerror(g_errno));
// error!
return false;
}
// none to sync from
return log("sync: could not find adequate twin to sync from!");
}
// . returns false if blocked, true otherwise.
// . returns true and sets g_errno on error
// . before we can spider for a SpiderRequest we must be granted the lock
// . each group shares the same doledb and each host in the group competes
// for spidering all those urls.
// . that way if a host goes down is load is taken over
bool Msg12::getLocks ( int64_t uh48, // probDocId ,
char *url ,
DOLEDBKEY *doledbKey,
collnum_t collnum,
int32_t sameIpWaitTime,
int32_t maxSpidersOutPerIp,
int32_t firstIp,
void *state ,
void (* callback)(void *state) ) {
// ensure not in use. not msg12 replies outstanding.
if ( m_numRequests != m_numReplies ) { char *xx=NULL;*xx=0; }
// no longer use this
char *xx=NULL;*xx=0;
// do not use locks for injections
//if ( m_sreq->m_isInjecting ) return true;
// get # of hosts in each mirror group
int32_t hpg = g_hostdb.getNumHostsPerShard();
// reset
m_numRequests = 0;
m_numReplies = 0;
m_grants = 0;
m_removing = false;
m_confirming = false;
// make sure is really docid
//if ( probDocId & ~DOCID_MASK ) { char *xx=NULL;*xx=0; }
// . mask out the lower bits that may change if there is a collision
// . in this way a url has the same m_probDocId as the same url
// in the index. i.e. if we add a new spider request for url X and
// url X is already indexed, then they will share the same lock
// even though the indexed url X may have a different actual docid
// than its probable docid.
// . we now use probable docids instead of uh48 because query reindex
// in PageReindex.cpp adds docid based spider requests and we
// only know the docid, not the uh48 because it is creating
// SpiderRequests from docid-only search results. having to look
// up the msg20 summary for like 1M search results is too painful!
//m_lockKey = g_titledb.getFirstProbableDocId(probDocId);
// . use this for locking now, and let the docid-only requests just use
// the docid
m_lockKeyUh48 = makeLockTableKey ( uh48 , firstIp );
m_url = url;
m_callback = callback;
m_state = state;
m_hasLock = false;
m_origUh48 = uh48;
// support ability to spider multiple urls from same ip
m_doledbKey = *doledbKey;
m_collnum = collnum;
m_sameIpWaitTime = sameIpWaitTime;
m_maxSpidersOutPerIp = maxSpidersOutPerIp;
m_firstIp = firstIp;
// sanity check, just 6 bytes! (48 bits)
if ( uh48 & 0xffff000000000000LL ) { char *xx=NULL;*xx=0; }
if ( m_lockKeyUh48 & 0xffff000000000000LL ) { char *xx=NULL;*xx=0; }
// cache time
int32_t ct = 120;
// if docid based assume it was a query reindex and keep it short!
// otherwise we end up waiting 120 seconds for a query reindex to
// go through on a docid we just spidered. TODO: use m_urlIsDocId
// MDW: check this out
if ( url && is_digit(url[0]) ) ct = 2;
// . this seems to be messing us up and preventing us from adding new
// requests into doledb when only spidering a few IPs.
// . make it random in the case of twin contention
ct = rand() % 10;
// . check our cache to avoid repetitive asking
// . use -1 for maxAge to indicate no max age
// . returns -1 if not in cache
// . use maxage of two minutes, 120 seconds
int32_t lockTime ;
lockTime = g_spiderLoop.m_lockCache.getLong(0,m_lockKeyUh48,ct,true);
// if it was in the cache and less than 2 minutes old then return
// true now with m_hasLock set to false.
if ( lockTime >= 0 ) {
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: cached missed lock for %s "
"lockkey=%" PRIu64, m_url,m_lockKeyUh48);
return true;
}
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: sending lock request for %s "
"lockkey=%" PRIu64, m_url,m_lockKeyUh48);
// now the locking group is based on the probable docid
//m_lockGroupId = g_hostdb.getGroupIdFromDocId(m_lockKey);
// ptr to list of hosts in the group
//Host *hosts = g_hostdb.getGroup ( m_lockGroupId );
// the same group (shard) that has the spiderRequest/Reply is
// the one responsible for locking.
Host *hosts = g_hostdb.getMyShard();
// shortcut
UdpServer *us = &g_udpServer;
static int32_t s_lockSequence = 0;
// remember the lock sequence # in case we have to call remove locks
m_lockSequence = s_lockSequence++;
LockRequest *lr = &m_lockRequest;
lr->m_lockKeyUh48 = m_lockKeyUh48;
lr->m_firstIp = m_firstIp;
lr->m_removeLock = 0;
lr->m_lockSequence = m_lockSequence;
lr->m_collnum = collnum;
// reset counts
m_numRequests = 0;
m_numReplies = 0;
// point to start of the 12 byte request buffer
char *request = (char *)lr;//m_lockKey;
int32_t requestSize = sizeof(LockRequest);//12;
// loop over hosts in that shard
for ( int32_t i = 0 ; i < hpg ; i++ ) {
// get a host
Host *h = &hosts[i];
// skip if dead! no need to get a reply from dead guys
if ( g_hostdb.isDead (h) ) continue;
// note it
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: sent lock "
"request #%" PRId32" for lockkey=%" PRIu64" %s to "
"hid=%" PRId32,m_numRequests,m_lockKeyUh48,
m_url,h->m_hostId);
// send request to him
if ( ! us->sendRequest ( request ,
requestSize ,
0x12 , // msgType
h->m_ip ,
h->m_port ,
h->m_hostId ,
NULL , // retSlotPtrPtr
this , // state data
gotLockReplyWrapper ,
udpserver_sendrequest_infinite_timeout ) )
// udpserver returns false and sets g_errno on error
return true;
// count them
m_numRequests++;
}
// block?
if ( m_numRequests > 0 ) return false;
// i guess nothing... hmmm... all dead?
//char *xx=NULL; *xx=0;
// m_hasLock should be false... all lock hosts seem dead... wait
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: all lock hosts seem dead for %s "
"lockkey=%" PRIu64, m_url,m_lockKeyUh48);
return true;
}
bool Msg12::confirmLockAcquisition ( ) {
// ensure not in use. not msg12 replies outstanding.
if ( m_numRequests != m_numReplies ) { char *xx=NULL;*xx=0; }
// no longer use this
char *xx=NULL;*xx=0;
// we are now removing
m_confirming = true;
// make that the request
// . point to start of the 12 byte request buffer
// . m_lockSequence should still be valid
ConfirmRequest *cq = &m_confirmRequest;
char *request = (char *)cq;
int32_t requestSize = sizeof(ConfirmRequest);
// sanity
if ( requestSize == sizeof(LockRequest)){ char *xx=NULL;*xx=0; }
// set it
cq->m_collnum = m_collnum;
cq->m_doledbKey = m_doledbKey;
cq->m_firstIp = m_firstIp;
cq->m_lockKeyUh48 = m_lockKeyUh48;
cq->m_maxSpidersOutPerIp = m_maxSpidersOutPerIp;
// . use the locking group from when we sent the lock request
// . get ptr to list of hosts in the group
//Host *hosts = g_hostdb.getGroup ( m_lockGroupId );
// the same group (shard) that has the spiderRequest/Reply is
// the one responsible for locking.
Host *hosts = g_hostdb.getMyShard();
// this must select the same shard that is going to spider it!
// i.e. our shard! because we check our local lock table to see
// if a doled url is locked before spidering it ourselves.
//Host *hosts = g_hostdb.getMyShard();
// shortcut
UdpServer *us = &g_udpServer;
// get # of hosts in each mirror group
int32_t hpg = g_hostdb.getNumHostsPerShard();
// reset counts
m_numRequests = 0;
m_numReplies = 0;
// note it
if ( g_conf.m_logDebugSpider )
log("spider: confirming lock for uh48=%" PRIu64" firstip=%s",
m_lockKeyUh48,iptoa(m_firstIp));
// loop over hosts in that shard
for ( int32_t i = 0 ; i < hpg ; i++ ) {
// get a host
Host *h = &hosts[i];
// skip if dead! no need to get a reply from dead guys
if ( g_hostdb.isDead ( h ) ) continue;
// send request to him
if ( ! us->sendRequest ( request ,
// a size of 2 should mean confirm
requestSize ,
0x12 , // msgType
h->m_ip ,
h->m_port ,
h->m_hostId ,
NULL , // retSlotPtrPtr
this , // state data
gotLockReplyWrapper ,
udpserver_sendrequest_infinite_timeout ) )
// udpserver returns false and sets g_errno on error
return true;
// count them
m_numRequests++;
}
// block?
if ( m_numRequests > 0 ) return false;
// did not block
return true;
}
bool Msg12::removeAllLocks ( ) {
// ensure not in use. not msg12 replies outstanding.
if ( m_numRequests != m_numReplies ) { char *xx=NULL;*xx=0; }
// no longer use this
char *xx=NULL;*xx=0;
// skip if injecting
//if ( m_sreq->m_isInjecting ) return true;
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: removing all locks for %s %" PRIu64,
m_url,m_lockKeyUh48);
// we are now removing
m_removing = true;
LockRequest *lr = &m_lockRequest;
lr->m_lockKeyUh48 = m_lockKeyUh48;
lr->m_lockSequence = m_lockSequence;
lr->m_firstIp = m_firstIp;
lr->m_removeLock = 1;
// reset counts
m_numRequests = 0;
m_numReplies = 0;
// make that the request
// . point to start of the 12 byte request buffer
// . m_lockSequence should still be valid
char *request = (char *)lr;//m_lockKey;
int32_t requestSize = sizeof(LockRequest);//12;
// now the locking group is based on the probable docid
//uint32_t groupId = g_hostdb.getGroupIdFromDocId(m_lockKeyUh48);
// ptr to list of hosts in the group
//Host *hosts = g_hostdb.getGroup ( groupId );
Host *hosts = g_hostdb.getMyShard();
// this must select the same group that is going to spider it!
// i.e. our group! because we check our local lock table to see
// if a doled url is locked before spidering it ourselves.
//Host *hosts = g_hostdb.getMyGroup();
// shortcut
UdpServer *us = &g_udpServer;
// set the hi bit though for this one
//m_lockKey |= 0x8000000000000000LL;
// get # of hosts in each mirror group
int32_t hpg = g_hostdb.getNumHostsPerShard();
// loop over hosts in that shard
for ( int32_t i = 0 ; i < hpg ; i++ ) {
// get a host
Host *h = &hosts[i];
// skip if dead! no need to get a reply from dead guys
if ( g_hostdb.isDead ( h ) ) continue;
// send request to him
if ( ! us->sendRequest ( request ,
requestSize ,
0x12 , // msgType
h->m_ip ,
h->m_port ,
h->m_hostId ,
NULL , // retSlotPtrPtr
this , // state data
gotLockReplyWrapper ,
udpserver_sendrequest_infinite_timeout ) )
// udpserver returns false and sets g_errno on error
return true;
// count them
m_numRequests++;
}
// block?
if ( m_numRequests > 0 ) return false;
// did not block
return true;
}
// returns true if all done, false if waiting for more replies
bool Msg12::gotLockReply ( UdpSlot *slot ) {
// no longer use this
char *xx=NULL;*xx=0;
// got reply
m_numReplies++;
// don't let udpserver free the request, it's our m_request[]
slot->m_sendBufAlloc = NULL;
// check for a hammer reply
char *reply = slot->m_readBuf;
int32_t replySize = slot->m_readBufSize;
// if error, treat as a not grant
if ( g_errno ) {
bool logIt = true;
// note it
if ( g_conf.m_logDebugSpider )
log("spider: got msg12 reply error = %s",
mstrerror(g_errno));
// if we got an ETRYAGAIN when trying to confirm our lock
// that means doledb was saving/dumping to disk and we
// could not remove the record from doledb and add an
// entry to the waiting tree, so we need to keep trying
if ( g_errno == ETRYAGAIN && m_confirming ) {
// c ount it again
m_numRequests++;
// use what we were using
char *request = (char *)&m_confirmRequest;
int32_t requestSize = sizeof(ConfirmRequest);
Host *h = g_hostdb.getHost(slot->m_hostId);
// send request to him
UdpServer *us = &g_udpServer;
if ( ! us->sendRequest ( request ,
requestSize ,
0x12 , // msgType
h->m_ip ,
h->m_port ,
h->m_hostId ,
NULL , // retSlotPtrPt
this , // state data
gotLockReplyWrapper ,
udpserver_sendrequest_infinite_timeout ) )
return false;
// error?
// don't spam the log!
static int32_t s_last = 0;
int32_t now = getTimeLocal();
if ( now - s_last >= 1 ) {
s_last = now;
log("spider: error re-sending confirm "
"request: %s", mstrerror(g_errno));
}
}
// only log every 10 seconds for ETRYAGAIN
if ( g_errno == ETRYAGAIN ) {
static time_t s_lastTime = 0;
time_t now = getTimeLocal();
logIt = false;
if ( now - s_lastTime >= 3 ) {
logIt = true;
s_lastTime = now;
}
}
if ( logIt )
log ( "sploop: host had error getting lock url=%s"
": %s" ,
m_url,mstrerror(g_errno) );
}
// grant or not
if ( replySize == 1 && ! g_errno && *reply == 1 ) m_grants++;
// wait for all to get back
if ( m_numReplies < m_numRequests ) return false;
// all done if we were removing
if ( m_removing ) {
// note it
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: done removing all locks "
"(replies=%" PRId32") for %s",
m_numReplies,m_url);//m_sreq->m_url);
// we are done
m_gettingLocks = false;
return true;
}
// all done if we were confirming
if ( m_confirming ) {
// note it
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: done confirming all locks "
"for %s uh48=%" PRId64,m_url,m_origUh48);//m_sreq->m_url);
// we are done
m_gettingLocks = false;
// . keep processing
// . if the collection was nuked from under us the spiderUrl2
// will return true and set g_errno
if ( ! m_callback ) return g_spiderLoop.spiderUrl2();
// if we had a callback let our parent call it
return true;
}
// if got ALL locks, spider it
if ( m_grants == m_numReplies ) {
// note it
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: got lock for docid=lockkey=%" PRIu64,
m_lockKeyUh48);
// flag this
m_hasLock = true;
// we are done
//m_gettingLocks = false;
///////
//
// now tell our group (shard) to remove from doledb
// and re-add to waiting tree. the evalIpLoop() function
// should skip this probable docid because it is in the
// LOCK TABLE!
//
// This logic should allow us to spider multiple urls
// from the same IP at the same time.
//
///////
// returns false if would block
if ( ! confirmLockAcquisition ( ) ) return false;
// . we did it without blocking, maybe cuz we are a single node
// . ok, they are all back, resume loop
// . if the collection was nuked from under us the spiderUrl2
// will return true and set g_errno
if ( ! m_callback ) g_spiderLoop.spiderUrl2 ( );
// all done
return true;
}
// note it
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: missed lock for %s lockkey=%" PRIu64" "
"(grants=%" PRId32")", m_url,m_lockKeyUh48,m_grants);
// . if it was locked by another then add to our lock cache so we do
// not try to lock it again
// . if grants is not 0 then one host granted us the lock, but not
// all hosts, so we should probably keep trying on it until it is
// locked up by one host
if ( m_grants == 0 ) {
int32_t now = getTimeGlobal();
g_spiderLoop.m_lockCache.addLong(0,m_lockKeyUh48,now,NULL);
}
// reset again
m_numRequests = 0;
m_numReplies = 0;
// no need to remove them if none were granted because another
// host in our group might have it 100% locked.
if ( m_grants == 0 ) {
// no longer in locks operation mode
m_gettingLocks = false;
// ok, they are all back, resume loop
//if ( ! m_callback ) g_spiderLoop.spiderUrl2 ( );
// all done
return true;
}
// note that
if ( g_conf.m_logDebugSpider )
logf(LOG_DEBUG,"spider: sending request to all in shard to "
"remove lock uh48=%" PRIu64". grants=%" PRId32,
m_lockKeyUh48,(int32_t)m_grants);
// remove all locks we tried to get, BUT only if from our hostid!
// no no! that doesn't quite work right... we might be the ones
// locking it! i.e. another one of our spiders has it locked...
if ( ! removeAllLocks ( ) ) return false; // true;
// if did not block, how'd that happen?
log("sploop: did not block in removeAllLocks: %s",mstrerror(g_errno));
return true;
}
// . THIS Msg0 class must be alloc'd, i.e. not on the stack, etc.
// . if list is stored locally this tries to get it locally
// . otherwise tries to get the list from the network
// . returns false if blocked, true otherwise
// . sets g_errno on error
// . NOTE: i was having problems with queries being cached too long, you
// see the cache here is a NETWORK cache, so when the machines that owns
// the list updates it on disk it can't flush our cache... so use a small
// maxCacheAge of like , 30 seconds or so...
bool Msg0::getList ( int64_t hostId , // host to ask (-1 if none)
int32_t ip , // info on hostId
int16_t port ,
int32_t maxCacheAge , // max cached age in seconds
bool addToCache , // add net recv'd list to cache?
char rdbId , // specifies the rdb
collnum_t collnum ,
RdbList *list ,
const char *startKey ,
const char *endKey ,
int32_t minRecSizes , // use -1 for no max
void *state ,
void (* callback)(void *state ),//, RdbList *list ) ,
int32_t niceness ,
bool doErrorCorrection ,
bool includeTree ,
bool doMerge ,
int32_t firstHostId ,
int32_t startFileNum ,
int32_t numFiles ,
int64_t timeout ,
int64_t syncPoint ,
int32_t preferLocalReads ,
Msg5 *msg5 ,
bool isRealMerge ,
bool allowPageCache ,
bool forceLocalIndexdb ,
bool noSplit ,
int32_t forceParitySplit ) {
logTrace( g_conf.m_logTraceMsg0, "BEGIN. hostId: %" PRId64", rdbId: %d", hostId, (int)rdbId );
// warning
if ( collnum < 0 ) log(LOG_LOGIC,"net: NULL collection. msg0.");
// reset the list they passed us
list->reset();
// get keySize of rdb
m_ks = getKeySizeFromRdbId ( rdbId );
// if( g_conf.m_logTraceMsg0 )
// {
// log("%s:%s:%d: rdbId. [%d]", __FILE__,__func__,__LINE__, (int)rdbId);
// log("%s:%s:%d: m_ks.. [%d]", __FILE__,__func__,__LINE__, (int)m_ks);
// log("%s:%s:%d: hostId [%" PRId64"]", __FILE__,__func__,__LINE__, hostId);
// }
// if startKey > endKey, don't read anything
//if ( startKey > endKey ) return true;
if ( KEYCMP(startKey,endKey,m_ks)>0 ) { char *xx=NULL;*xx=0; }//rettrue
// . reset hostid if it is dead
// . this is causing UOR queries to take forever when we have a dead
if ( hostId >= 0 && g_hostdb.isDead ( hostId ) ) hostId = -1;
// no longer accept negative minrecsize
if ( minRecSizes < 0 ) {
g_errno = EBADENGINEER;
logTrace( g_conf.m_logTraceMsg0, "END" );
log(LOG_LOGIC, "net: msg0: Negative minRecSizes no longer supported.");
char *xx=NULL;*xx=0;
}
// remember these
m_state = state;
m_callback = callback;
m_list = list;
m_hostId = hostId;
m_niceness = niceness;
m_addToCache = addToCache;
// . these define our request 100%
KEYSET(m_startKey,startKey,m_ks);
KEYSET(m_endKey,endKey,m_ks);
m_minRecSizes = minRecSizes;
m_rdbId = rdbId;
m_collnum = collnum;// = coll;
m_isRealMerge = isRealMerge;
m_allowPageCache = allowPageCache;
// . group to ask is based on the first key
// . we only do 1 group per call right now
// . groupMask must turn on higher bits first (count downwards kinda)
// . titledb and spiderdb use special masks to get groupId
// if diffbot.cpp is reading spiderdb from each shard we have to
// get groupid from hostid here lest we core in getGroupId() below.
// it does that for dumping spiderdb to the client browser. they
// can download the whole enchilada.
if ( hostId >= 0 && m_rdbId == RDB_SPIDERDB )
m_shardNum = 0;
// did they force it? core until i figure out what this is
else if ( forceParitySplit >= 0 )
//m_groupId = g_hostdb.getGroupId ( forceParitySplit );
m_shardNum = forceParitySplit;
else
//m_groupId = getGroupId ( m_rdbId , startKey , ! noSplit );
m_shardNum = getShardNum ( m_rdbId , startKey );
// if we are looking up a termlist in posdb that is split by termid and
// not the usual docid then we have to set this posdb key bit that tells
// us that ...
if ( noSplit && m_rdbId == RDB_POSDB )
m_shardNum = g_hostdb.getShardNumByTermId ( startKey );
// how is this used?
if ( forceLocalIndexdb ) m_shardNum = getMyShardNum();
// if( g_conf.m_logTraceMsg0 ) log("%s:%s:%d: shardNum [%" PRId32"]", __FILE__,__func__, __LINE__, m_shardNum);
// . store these parameters
// . get a handle to the rdb in case we can satisfy locally
// . returns NULL and sets g_errno on error
QUICKPOLL((m_niceness));
Rdb *rdb = getRdbFromId ( m_rdbId );
if ( ! rdb ) return true;
// we need the fixedDataSize
m_fixedDataSize = rdb->getFixedDataSize();
m_useHalfKeys = rdb->useHalfKeys();
// . debug msg
// . Msg2 does this when checking for a cached compound list.
// compound lists do not actually exist, they are merges of smaller
// UOR'd lists.
if ( maxCacheAge != 0 && ! addToCache && (numFiles > 0 || includeTree)) {
log( LOG_LOGIC, "net: msg0: Weird. check but don't add... rdbid=%" PRId32".", ( int32_t ) m_rdbId );
}
// set this here since we may not call msg5 if list not local
//m_list->setFixedDataSize ( m_fixedDataSize );
// . now that we do load balancing we don't want to do a disk lookup
// even if local if we are merging or dumping
// . UNLESS g_conf.m_preferLocalReads is true
if ( preferLocalReads == -1 )
preferLocalReads = g_conf.m_preferLocalReads;
// . always prefer local for full split clusterdb
// . and keep the tfndb/titledb lookups in the same stripe
// . so basically we can't do biased caches if fully split
//if ( g_conf.m_fullSplit ) preferLocalReads = true;
preferLocalReads = true;
// it it stored locally?
bool isLocal = ( m_hostId == -1 && //g_hostdb.m_groupId == m_groupId );
m_shardNum == getMyShardNum() );
// only do local lookups if this is true
if ( ! preferLocalReads ) isLocal = false;
/*
int64_t singleDocIdQuery = 0LL;
if ( rdbId == RDB_POSDB ) {
int64_t d1 = g_posdb.getDocId(m_startKey);
int64_t d2 = g_posdb.getDocId(m_endKey);
if ( d1+1 == d2 ) singleDocIdQuery = d1;
}
// . try the LOCAL termlist cache
// . so when msg2 is evaluating a gbdocid:| query and it has to
// use msg0 to go across the network to get the same damn termlist
// over and over again for the same docid, this will help alot.
// . ideally it'd be nice if the seo pipe in xmldoc.cpp can try to
// send the same gbdocid:xxxx docids to the same hosts. maybe hash
// based on docid into the list of hosts and if that host is busy
// just chain until we find someone not busy.
if ( singleDocIdQuery &&
getListFromTermListCache ( coll,
m_startKey,
m_endKey,
maxCacheAge,
list ) )
// found!
return true;
*/
// but always local if only one host
if ( g_hostdb.getNumHosts() == 1 ) isLocal = true;
// . if the group is local then do it locally
// . Msg5::getList() returns false if blocked, true otherwise
// . Msg5::getList() sets g_errno on error
// . don't do this if m_hostId was specified
if ( isLocal ) {
logTrace( g_conf.m_logTraceMsg0, "isLocal" );
if ( msg5 ) {
m_msg5 = msg5;
m_deleteMsg5 = false;
}
else {
try { m_msg5 = new ( Msg5 ); }
catch ( ... ) {
g_errno = ENOMEM;
log("net: Local alloc for disk read failed "
"while tring to read data for %s. "
"Trying remote request.",
getDbnameFromId(m_rdbId));
goto skip;
}
mnew ( m_msg5 , sizeof(Msg5) , "Msg0::Msg5" );
m_deleteMsg5 = true;
}
QUICKPOLL(m_niceness);
if ( ! m_msg5->getList ( rdbId,
m_collnum ,
m_list ,
m_startKey ,
m_endKey ,
m_minRecSizes ,
includeTree , // include Tree?
addToCache , // addToCache?
maxCacheAge ,
startFileNum ,
numFiles ,
this ,
gotListWrapper2 ,
niceness ,
doErrorCorrection ,
NULL , // cacheKeyPtr
0 , // retryNum
-1 , // maxRetries
true , // compensateForMerge
syncPoint ,
m_isRealMerge ,
m_allowPageCache ) ) {
logTrace( g_conf.m_logTraceMsg0, "END, return false" );
return false;
}
// nuke it
reset();
logTrace( g_conf.m_logTraceMsg0, "END, return true" );
return true;
}
skip:
// debug msg
if ( g_conf.m_logDebugQuery )
log(LOG_DEBUG,"net: msg0: Sending request for data to "
"shard=%" PRIu32" "
"listPtr=%" PTRFMT" minRecSizes=%" PRId32" termId=%" PRIu64" "
//"startKey.n1=%" PRIx32",n0=%" PRIx64" (niceness=%" PRId32")",
"startKey.n1=%" PRIx64",n0=%" PRIx64" (niceness=%" PRId32")",
//g_hostdb.makeHostId ( m_groupId ) ,
m_shardNum,
(PTRTYPE)m_list,
m_minRecSizes, g_posdb.getTermId(m_startKey) ,
//m_startKey.n1,m_startKey.n0 , (int32_t)m_niceness);
KEY1(m_startKey,m_ks),KEY0(m_startKey),
(int32_t)m_niceness);
char *replyBuf = NULL;
int32_t replyBufMaxSize = 0;
bool freeReply = true;
// . make a request with the info above (note: not in network order)
// . IMPORTANT!!!!! if you change this change
// Multicast.cpp::sleepWrapper1 too!!!!!!!!!!!!
// no, not anymore, we commented out that request peeking code
char *p = m_request;
*(int64_t *) p = syncPoint ; p += 8;
//*(key_t *) p = m_startKey ; p += sizeof(key_t);
//*(key_t *) p = m_endKey ; p += sizeof(key_t);
*(int32_t *) p = m_minRecSizes ; p += 4;
*(int32_t *) p = startFileNum ; p += 4;
*(int32_t *) p = numFiles ; p += 4;
*(int32_t *) p = maxCacheAge ; p += 4;
if ( p - m_request != RDBIDOFFSET ) { char *xx=NULL;*xx=0; }
*p = m_rdbId ; p++;
*p = addToCache ; p++;
*p = doErrorCorrection; p++;
*p = includeTree ; p++;
*p = (char)niceness ; p++;
*p = (char)m_allowPageCache; p++;
KEYSET(p,m_startKey,m_ks); ; p+=m_ks;
KEYSET(p,m_endKey,m_ks); ; p+=m_ks;
// NULL terminated collection name
//strcpy ( p , coll ); p += gbstrlen ( coll ); *p++ = '\0';
*(collnum_t *)p = m_collnum; p += sizeof(collnum_t);
m_requestSize = p - m_request;
// ask an individual host for this list if hostId is NOT -1
if ( m_hostId != -1 ) {
// get Host
Host *h = g_hostdb.getHost ( m_hostId );
if ( ! h ) {
g_errno = EBADHOSTID;
log(LOG_LOGIC,"net: msg0: Bad hostId of %" PRId64".", m_hostId);
logTrace( g_conf.m_logTraceMsg0, "END, return true. Bad hostId" );
return true;
}
// if niceness is 0, use the higher priority udpServer
UdpServer *us ;
uint16_t port;
QUICKPOLL(m_niceness);
us = &g_udpServer ; port = h->m_port ;
// . returns false on error and sets g_errno, true otherwise
// . calls callback when reply is received (or error)
// . we return true if it returns false
if ( ! us->sendRequest ( m_request ,
m_requestSize ,
0x00 , // msgType
h->m_ip ,
port ,
m_hostId ,
NULL , // the slotPtr
this ,
gotSingleReplyWrapper ,
timeout ,
-1 , // backoff
-1 , // maxwait
replyBuf ,
replyBufMaxSize ,
m_niceness ) ) { // cback niceness
logTrace( g_conf.m_logTraceMsg0, "END, return true. Request sent" );
return true;
}
// return false cuz it blocked
logTrace( g_conf.m_logTraceMsg0, "END, return false. sendRequest blocked" );
return false;
}
// timing debug
if ( g_conf.m_logTimingNet )
m_startTime = gettimeofdayInMilliseconds();
else
m_startTime = 0;
// . get the top int32_t of the key
// . i guess this will work for 128 bit keys... hmmmmm
int32_t keyTop = hash32 ( (char *)startKey , m_ks );
// . otherwise, multicast to a host in group "groupId"
// . returns false and sets g_errno on error
// . calls callback on completion
// . select first host to send to in group based on upper 32 bits
// of termId (m_startKey.n1)
// . need to send out to all the indexdb split hosts
m_numRequests = 0;
m_numReplies = 0;
//for ( int32_t i = 0; i < m_numSplit; i++ ) {
QUICKPOLL(m_niceness);
//int32_t gr;
char *buf;
buf = replyBuf;
// get the multicast
Multicast *m = &m_mcast;
if ( ! m->send ( m_request ,
m_requestSize,
0x00 , // msgType 0x00
false , // does multicast own request?
m_shardNum ,
false , // send to whole group?
//m_startKey.n1, // key is passed on startKey
keyTop , // key is passed on startKey
this , // state data
NULL , // state data
gotMulticastReplyWrapper0 ,
timeout*1000 , // timeout
niceness ,
firstHostId ,
buf ,
replyBufMaxSize ,
freeReply , // free reply buf?
true , // do disk load balancing?
maxCacheAge ,
//(key_t *)cacheKey ,
// multicast uses it for determining the best
// host to send the request to when doing
// disk load balancing. if the host has our
// data cached, then it will probably get to
// handle the request. for now let's just assume
// this is a 96-bit key. TODO: fix...
0 , // *(key_t *)cacheKey ,
rdbId ,
minRecSizes ) )
{
log(LOG_ERROR, "net: Failed to send request for data from %s in shard "
"#%" PRIu32" over network: %s.",
getDbnameFromId(m_rdbId),m_shardNum, mstrerror(g_errno));
// but speed it up
m_errno = g_errno;
m->reset();
if ( m_numRequests > 0 ) {
logTrace( g_conf.m_logTraceMsg0, "END - returning false" );
return false;
}
logTrace( g_conf.m_logTraceMsg0, "END - returning true" );
return true;
}
m_numRequests++;
// we blocked
logTrace( g_conf.m_logTraceMsg0, "END - returning false, blocked" );
return false;
}
// . THIS Msg0 class must be alloc'd, i.e. not on the stack, etc.
// . if list is stored locally this tries to get it locally
// . otherwise tries to get the list from the network
// . returns false if blocked, true otherwise
// . sets g_errno on error
// . NOTE: i was having problems with queries being cached too long, you
// see the cache here is a NETWORK cache, so when the machines that owns
// the list updates it on disk it can't flush our cache... so use a small
// maxCacheAge of like , 30 seconds or so...
bool Msg0::getList ( int64_t hostId , // host to ask (-1 if none)
int32_t ip , // info on hostId
int16_t port ,
int32_t maxCacheAge , // max cached age in seconds
bool addToCache , // add net recv'd list to cache?
char rdbId , // specifies the rdb
//char *coll ,
collnum_t collnum ,
RdbList *list ,
//key_t startKey ,
//key_t endKey ,
char *startKey ,
char *endKey ,
int32_t minRecSizes , // use -1 for no max
void *state ,
void (* callback)(void *state ),//, RdbList *list ) ,
int32_t niceness ,
bool doErrorCorrection ,
bool includeTree ,
bool doMerge ,
int32_t firstHostId ,
int32_t startFileNum ,
int32_t numFiles ,
int32_t timeout ,
int64_t syncPoint ,
int32_t preferLocalReads ,
Msg5 *msg5 ,
Msg5 *msg5b ,
bool isRealMerge ,
//#ifdef SPLIT_INDEXDB
bool allowPageCache ,
bool forceLocalIndexdb ,
bool noSplit , // doIndexdbSplit ,
int32_t forceParitySplit ) {
//#else
// bool allowPageCache ) {
//#endif
// this is obsolete! mostly, but we need it for PageIndexdb.cpp to
// show a "termlist" for a given query term in its entirety so you
// don't have to check each machine in the network. if this is true it
// means to query each split and merge the results together into a
// single unified termlist. only applies to indexdb/datedb.
//if ( doIndexdbSplit ) { char *xx = NULL; *xx = 0; }
// note this because if caller is wrong it hurts performance major!!
//if ( doIndexdbSplit )
// logf(LOG_DEBUG,"net: doing msg0 with indexdb split true");
// warning
if ( collnum < 0 ) log(LOG_LOGIC,"net: NULL collection. msg0.");
//if ( doIndexdbSplit ) { char *xx=NULL;*xx=0; }
// reset the list they passed us
list->reset();
// get keySize of rdb
m_ks = getKeySizeFromRdbId ( rdbId );
// if startKey > endKey, don't read anything
//if ( startKey > endKey ) return true;
if ( KEYCMP(startKey,endKey,m_ks)>0 ) { char *xx=NULL;*xx=0; }//rettrue
// . reset hostid if it is dead
// . this is causing UOR queries to take forever when we have a dead
if ( hostId >= 0 && g_hostdb.isDead ( hostId ) ) hostId = -1;
// no longer accept negative minrecsize
if ( minRecSizes < 0 ) {
g_errno = EBADENGINEER;
log(LOG_LOGIC,
"net: msg0: Negative minRecSizes no longer supported.");
char *xx=NULL;*xx=0;
return true;
}
// debug msg
//if ( niceness != 0 ) log("HEY start");
// ensure startKey last bit clear, endKey last bit set
//if ( (startKey.n0 & 0x01) == 0x01 )
// log("Msg0::getList: warning startKey lastbit set");
//if ( (endKey.n0 & 0x01) == 0x00 )
// log("Msg0::getList: warning endKey lastbit clear");
// remember these
m_state = state;
m_callback = callback;
m_list = list;
m_hostId = hostId;
m_niceness = niceness;
//m_ip = ip;
//m_port = port;
m_addToCache = addToCache;
// . these define our request 100%
//m_startKey = startKey;
//m_endKey = endKey;
KEYSET(m_startKey,startKey,m_ks);
KEYSET(m_endKey,endKey,m_ks);
m_minRecSizes = minRecSizes;
m_rdbId = rdbId;
m_collnum = collnum;// = coll;
m_isRealMerge = isRealMerge;
m_allowPageCache = allowPageCache;
// . group to ask is based on the first key
// . we only do 1 group per call right now
// . groupMask must turn on higher bits first (count downwards kinda)
// . titledb and spiderdb use special masks to get groupId
// if diffbot.cpp is reading spiderdb from each shard we have to
// get groupid from hostid here lest we core in getGroupId() below.
// it does that for dumping spiderdb to the client browser. they
// can download the whole enchilada.
if ( hostId >= 0 && m_rdbId == RDB_SPIDERDB )
m_shardNum = 0;
// did they force it? core until i figure out what this is
else if ( forceParitySplit >= 0 )
//m_groupId = g_hostdb.getGroupId ( forceParitySplit );
m_shardNum = forceParitySplit;
else
//m_groupId = getGroupId ( m_rdbId , startKey , ! noSplit );
m_shardNum = getShardNum ( m_rdbId , startKey );
// if we are looking up a termlist in posdb that is split by termid and
// not the usual docid then we have to set this posdb key bit that tells
// us that ...
if ( noSplit && m_rdbId == RDB_POSDB )
m_shardNum = g_hostdb.getShardNumByTermId ( startKey );
// how is this used?
//if ( forceLocalIndexdb ) m_groupId = g_hostdb.m_groupId;
if ( forceLocalIndexdb ) m_shardNum = getMyShardNum();
// . store these parameters
// . get a handle to the rdb in case we can satisfy locally
// . returns NULL and sets g_errno on error
QUICKPOLL((m_niceness));
Rdb *rdb = getRdbFromId ( m_rdbId );
if ( ! rdb ) return true;
// we need the fixedDataSize
m_fixedDataSize = rdb->getFixedDataSize();
m_useHalfKeys = rdb->useHalfKeys();
// . debug msg
// . Msg2 does this when checking for a cached compound list.
// compound lists do not actually exist, they are merges of smaller
// UOR'd lists.
if ( maxCacheAge != 0 && ! addToCache && (numFiles > 0 || includeTree))
log(LOG_LOGIC,"net: msg0: "
"Weird. check but don't add... rdbid=%"INT32".",(int32_t)m_rdbId);
// set this here since we may not call msg5 if list not local
//m_list->setFixedDataSize ( m_fixedDataSize );
// . now that we do load balancing we don't want to do a disk lookup
// even if local if we are merging or dumping
// . UNLESS g_conf.m_preferLocalReads is true
if ( preferLocalReads == -1 )
preferLocalReads = g_conf.m_preferLocalReads;
// . always prefer local for full split clusterdb
// . and keep the tfndb/titledb lookups in the same stripe
// . so basically we can't do biased caches if fully split
//if ( g_conf.m_fullSplit ) preferLocalReads = true;
preferLocalReads = true;
// it it stored locally?
bool isLocal = ( m_hostId == -1 && //g_hostdb.m_groupId == m_groupId );
m_shardNum == getMyShardNum() );
// only do local lookups if this is true
if ( ! preferLocalReads ) isLocal = false;
/*
m_numSplit = 1;
if ( g_hostdb.m_indexSplits > 1 &&
( rdbId == RDB_POSDB || rdbId==RDB_DATEDB)&&
! forceLocalIndexdb && doIndexdbSplit ) {
isLocal = false;
//m_numSplit = INDEXDB_SPLIT;
m_numSplit = g_hostdb.m_indexSplits;
char *xx=NULL;*xx=0;
}
*/
/*
int64_t singleDocIdQuery = 0LL;
if ( rdbId == RDB_POSDB ) {
int64_t d1 = g_posdb.getDocId(m_startKey);
int64_t d2 = g_posdb.getDocId(m_endKey);
if ( d1+1 == d2 ) singleDocIdQuery = d1;
}
// . try the LOCAL termlist cache
// . so when msg2 is evaluating a gbdocid:| query and it has to
// use msg0 to go across the network to get the same damn termlist
// over and over again for the same docid, this will help alot.
// . ideally it'd be nice if the seo pipe in xmldoc.cpp can try to
// send the same gbdocid:xxxx docids to the same hosts. maybe hash
// based on docid into the list of hosts and if that host is busy
// just chain until we find someone not busy.
if ( singleDocIdQuery &&
getListFromTermListCache ( coll,
m_startKey,
m_endKey,
maxCacheAge,
list ) )
// found!
return true;
*/
// but always local if only one host
if ( g_hostdb.getNumHosts() == 1 ) isLocal = true;
// force a msg0 if doing a docid restrictive query like
// gbdocid:xxxx|<query> so we call cacheTermLists()
//if ( singleDocIdQuery ) isLocal = false;
// . if the group is local then do it locally
// . Msg5::getList() returns false if blocked, true otherwise
// . Msg5::getList() sets g_errno on error
// . don't do this if m_hostId was specified
if ( isLocal ) { // && !g_conf.m_interfaceMachine ) {
if ( msg5 ) {
m_msg5 = msg5;
m_deleteMsg5 = false;
}
else {
try { m_msg5 = new ( Msg5 ); }
catch ( ... ) {
g_errno = ENOMEM;
log("net: Local alloc for disk read failed "
"while tring to read data for %s. "
"Trying remote request.",
getDbnameFromId(m_rdbId));
goto skip;
}
mnew ( m_msg5 , sizeof(Msg5) , "Msg0" );
m_deleteMsg5 = true;
}
QUICKPOLL(m_niceness);
// same for msg5b
if ( msg5b ) {
m_msg5b = msg5b;
m_deleteMsg5b = false;
}
/*
else if ( m_rdbId == RDB_TITLEDB ) {
try { m_msg5b = new ( Msg5 ); }
catch ( ... ) {
g_errno = ENOMEM;
log("net: Local alloc for disk read failed "
"while tring to read data for %s. "
"Trying remote request. 2.",
getDbnameFromId(m_rdbId));
goto skip;
}
mnew ( m_msg5b , sizeof(Msg5) , "Msg0b" );
m_deleteMsg5b = true;
}
*/
QUICKPOLL(m_niceness);
if ( ! m_msg5->getList ( rdbId,
m_collnum ,
m_list ,
m_startKey ,
m_endKey ,
m_minRecSizes ,
includeTree , // include Tree?
addToCache , // addToCache?
maxCacheAge ,
startFileNum ,
numFiles ,
this ,
gotListWrapper2 ,
niceness ,
doErrorCorrection ,
NULL , // cacheKeyPtr
0 , // retryNum
-1 , // maxRetries
true , // compensateForMerge
syncPoint ,
NULL,//m_msg5b ,
m_isRealMerge ,
m_allowPageCache ) ) return false;
// nuke it
reset();
return true;
}
skip:
// debug msg
if ( g_conf.m_logDebugQuery )
log(LOG_DEBUG,"net: msg0: Sending request for data to "
"shard=%"UINT32" "
"listPtr=%"PTRFMT" minRecSizes=%"INT32" termId=%"UINT64" "
//"startKey.n1=%"XINT32",n0=%"XINT64" (niceness=%"INT32")",
"startKey.n1=%"XINT64",n0=%"XINT64" (niceness=%"INT32")",
//g_hostdb.makeHostId ( m_groupId ) ,
m_shardNum,
(PTRTYPE)m_list,
m_minRecSizes, g_posdb.getTermId(m_startKey) ,
//m_startKey.n1,m_startKey.n0 , (int32_t)m_niceness);
KEY1(m_startKey,m_ks),KEY0(m_startKey),
(int32_t)m_niceness);
char *replyBuf = NULL;
int32_t replyBufMaxSize = 0;
bool freeReply = true;
// adjust niceness for net transmission
bool realtime = false;
//if ( minRecSizes + 32 < TMPBUFSIZE ) realtime = true;
// if we're niceness 0 we need to pre-allocate for reply since it
// might be received within the asynchronous signal handler which
// cannot call mmalloc()
if ( realtime ) { // niceness <= 0 || netnice == 0 ) {
// . we should not get back more than minRecSizes bytes since
// we are now performing merges
// . it should not slow things down too much since the hashing
// is 10 times slower than merging anyhow...
// . CAUTION: if rdb is not fixed-datasize then this will
// not work for us! it can exceed m_minRecSizes.
replyBufMaxSize = m_minRecSizes ;
// . get a little extra to fix the error where we ask for 64
// but get 72
// . where is that coming from?
// . when getting titleRecs we often exceed the minRecSizes
// . ?Msg8? was having trouble. was int16_t 32 bytes sometimes.
replyBufMaxSize += 36;
// why add ten percent?
//replyBufMaxSize *= 110 ;
//replyBufMaxSize /= 100 ;
// make a buffer to hold the reply
//#ifdef SPLIT_INDEXDB
/*
if ( m_numSplit > 1 ) {
m_replyBufSize = replyBufMaxSize * m_numSplit;
replyBuf = (char *) mmalloc(m_replyBufSize, "Msg0");
m_replyBuf = replyBuf;
freeReply = false;
}
else
*/
//#endif
replyBuf = (char *) mmalloc(replyBufMaxSize , "Msg0");
// g_errno is set and we return true if it failed
if ( ! replyBuf ) {
log("net: Failed to pre-allocate %"INT32" bytes to hold "
"data read remotely from %s: %s.",
replyBufMaxSize,getDbnameFromId(m_rdbId),
mstrerror(g_errno));
return true;
}
}
// . make a request with the info above (note: not in network order)
// . IMPORTANT!!!!! if you change this change
// Multicast.cpp::sleepWrapper1 too!!!!!!!!!!!!
// no, not anymore, we commented out that request peeking code
char *p = m_request;
*(int64_t *) p = syncPoint ; p += 8;
//*(key_t *) p = m_startKey ; p += sizeof(key_t);
//*(key_t *) p = m_endKey ; p += sizeof(key_t);
*(int32_t *) p = m_minRecSizes ; p += 4;
*(int32_t *) p = startFileNum ; p += 4;
*(int32_t *) p = numFiles ; p += 4;
*(int32_t *) p = maxCacheAge ; p += 4;
if ( p - m_request != RDBIDOFFSET ) { char *xx=NULL;*xx=0; }
*p = m_rdbId ; p++;
*p = addToCache ; p++;
*p = doErrorCorrection; p++;
*p = includeTree ; p++;
*p = (char)niceness ; p++;
*p = (char)m_allowPageCache; p++;
KEYSET(p,m_startKey,m_ks); ; p+=m_ks;
KEYSET(p,m_endKey,m_ks); ; p+=m_ks;
// NULL terminated collection name
//strcpy ( p , coll ); p += gbstrlen ( coll ); *p++ = '\0';
*(collnum_t *)p = m_collnum; p += sizeof(collnum_t);
m_requestSize = p - m_request;
// ask an individual host for this list if hostId is NOT -1
if ( m_hostId != -1 ) {
// get Host
Host *h = g_hostdb.getHost ( m_hostId );
if ( ! h ) {
g_errno = EBADHOSTID;
log(LOG_LOGIC,"net: msg0: Bad hostId of %"INT64".",
m_hostId);
return true;
}
// if niceness is 0, use the higher priority udpServer
UdpServer *us ;
uint16_t port;
QUICKPOLL(m_niceness);
//if ( niceness <= 0 || netnice == 0 ) {
//if ( realtime ) {
// us = &g_udpServer2; port = h->m_port2; }
//else {
us = &g_udpServer ; port = h->m_port ;
// . returns false on error and sets g_errno, true otherwise
// . calls callback when reply is received (or error)
// . we return true if it returns false
if ( ! us->sendRequest ( m_request ,
m_requestSize ,
0x00 , // msgType
h->m_ip ,
port ,
m_hostId ,
NULL , // the slotPtr
this ,
gotSingleReplyWrapper ,
timeout ,
-1 , // backoff
-1 , // maxwait
replyBuf ,
replyBufMaxSize ,
m_niceness ) ) // cback niceness
return true;
// return false cuz it blocked
return false;
}
// timing debug
if ( g_conf.m_logTimingNet )
m_startTime = gettimeofdayInMilliseconds();
else
m_startTime = 0;
//if ( m_rdbId == RDB_INDEXDB ) log("Msg0:: getting remote indexlist. "
// "termId=%"UINT64", "
// "groupNum=%"UINT32"",
// g_indexdb.getTermId(m_startKey) ,
// g_hostdb.makeHostId ( m_groupId ) );
/*
// make the cache key so we can see what remote host cached it, if any
char cacheKey[MAX_KEY_BYTES];
//key_t cacheKey = makeCacheKey ( startKey ,
makeCacheKey ( startKey ,
endKey ,
includeTree ,
minRecSizes ,
startFileNum ,
numFiles ,
cacheKey ,
m_ks );
*/
// . get the top int32_t of the key
// . i guess this will work for 128 bit keys... hmmmmm
int32_t keyTop = hash32 ( (char *)startKey , m_ks );
/*
// allocate space
if ( m_numSplit > 1 ) {
int32_t need = m_numSplit * sizeof(Multicast) ;
char *buf = (char *)mmalloc ( need,"msg0mcast" );
if ( ! buf ) return true;
m_mcasts = (Multicast *)buf;
for ( int32_t i = 0; i < m_numSplit ; i++ )
m_mcasts[i].constructor();
}
*/
// . otherwise, multicast to a host in group "groupId"
// . returns false and sets g_errno on error
// . calls callback on completion
// . select first host to send to in group based on upper 32 bits
// of termId (m_startKey.n1)
//#ifdef SPLIT_INDEXDB
// . need to send out to all the indexdb split hosts
m_numRequests = 0;
m_numReplies = 0;
//for ( int32_t i = 0; i < m_numSplit; i++ ) {
QUICKPOLL(m_niceness);
//int32_t gr;
char *buf;
/*
if ( m_numSplit > 1 ) {
gr = g_indexdb.getSplitGroupId ( baseGroupId, i );
buf = &replyBuf[i*replyBufMaxSize];
}
else {
*/
//gr = m_groupId;
buf = replyBuf;
//}
// get the multicast
Multicast *m = &m_mcast;
//if ( m_numSplit > 1 ) m = &m_mcasts[i];
if ( ! m->send ( m_request ,
//#else
// if ( ! m_mcast.send ( m_request ,
//#endif
m_requestSize,
0x00 , // msgType 0x00
false , // does multicast own request?
m_shardNum ,
//#ifdef SPLIT_INDEXDB
// gr , // group + offset
//#else
// m_groupId , // group to send to (groupKey)
//#endif
false , // send to whole group?
//m_startKey.n1, // key is passed on startKey
keyTop , // key is passed on startKey
this , // state data
NULL , // state data
gotMulticastReplyWrapper0 ,
timeout , // timeout in seconds (was 30)
niceness ,
realtime ,
firstHostId ,
//#ifdef SPLIT_INDEXDB
// &replyBuf[i*replyBufMaxSize] ,
//#else
// replyBuf ,
//#endif
buf ,
replyBufMaxSize ,
freeReply , // free reply buf?
true , // do disk load balancing?
maxCacheAge ,
//(key_t *)cacheKey ,
// multicast uses it for determining the best
// host to send the request to when doing
// disk load balancing. if the host has our
// data cached, then it will probably get to
// handle the request. for now let's just assume
// this is a 96-bit key. TODO: fix...
0 , // *(key_t *)cacheKey ,
rdbId ,
minRecSizes ) ) {
log("net: Failed to send request for data from %s in shard "
"#%"UINT32" over network: %s.",
getDbnameFromId(m_rdbId),m_shardNum, mstrerror(g_errno));
// no, multicast will free this when it is destroyed
//if (replyBuf) mfree ( replyBuf , replyBufMaxSize , "Msg22" );
// but speed it up
//#ifdef SPLIT_INDEXDB
m_errno = g_errno;
m->reset();
if ( m_numRequests > 0 )
return false;
//#else
// m_mcast.reset();
//#endif
return true;
}
//#ifdef SPLIT_INDEXDB
m_numRequests++;
//#endif
// we blocked
return false;
}