static std::vector<keypair_t> parseMetaList(const char *metaList, int32_t metaListSize) {
std::vector<keypair_t> result;
const char *p = metaList;
const char *pend = metaList + metaListSize;
for (; p < pend;) {
// get rdbId
rdbid_t rdbId = (rdbid_t)(*p & 0x7f);
p++;
// key size
int32_t ks = getKeySizeFromRdbId(rdbId);
// get key
const char *key = p;
p += ks;
// . if key is negative, no data is present
// . the doledb key is negative for us here
int32_t ds = KEYNEG(key) ? 0 : getDataSizeFromRdbId(rdbId);
// if datasize variable, read it in
if (ds == -1) {
// get data size
ds = *(int32_t *)p;
// skip data size int32_t
p += 4;
}
// skip data if not zero
p += ds;
result.push_back(std::make_pair(rdbId,key));
}
return result;
}
// . 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;
}
// . reply to a request for an RdbList
// . MUST call g_udpServer::sendReply or sendErrorReply() so slot can
// be destroyed
void handleRequest0 ( UdpSlot *slot , int32_t netnice ) {
logTrace( g_conf.m_logTraceMsg0, "BEGIN. Got request for an RdbList" );
// if niceness is 0, use the higher priority udpServer
UdpServer *us = &g_udpServer;
//if ( netnice == 0 ) us = &g_udpServer2;
// get the request
char *request = slot->m_readBuf;
int32_t requestSize = slot->m_readBufSize;
// collection is now stored in the request, so i commented this out
//if ( requestSize != MSG0_REQ_SIZE ) {
// log("net: Received bad data request size of %" PRId32" bytes. "
// "Should be %" PRId32".", requestSize ,(int32_t)MSG0_REQ_SIZE);
// us->sendErrorReply ( slot , EBADREQUESTSIZE );
// return;
//}
// parse the request
char *p = request;
int64_t syncPoint = *(int64_t *)p ; p += 8;
//key_t startKey = *(key_t *)p ; p += sizeof(key_t);
//key_t endKey = *(key_t *)p ; p += sizeof(key_t);
int32_t minRecSizes = *(int32_t *)p ; p += 4;
int32_t startFileNum = *(int32_t *)p ; p += 4;
int32_t numFiles = *(int32_t *)p ; p += 4;
int32_t maxCacheAge = *(int32_t *)p ; p += 4;
char rdbId = *p++;
char addToCache = *p++;
char doErrorCorrection = *p++;
char includeTree = *p++;
// this was messing up our niceness conversion logic
int32_t niceness = slot->m_niceness;//(int32_t)(*p++);
// still need to skip it though!
p++;
bool allowPageCache = (bool)(*p++);
char ks = getKeySizeFromRdbId ( rdbId );
char *startKey = p; p+=ks;
char *endKey = p; p+=ks;
collnum_t collnum = *(collnum_t *)p; p += sizeof(collnum_t);
CollectionRec *xcr = g_collectiondb.getRec ( collnum );
if ( ! xcr ) g_errno = ENOCOLLREC;
if( g_conf.m_logTraceMsg0 ) {
logTrace( g_conf.m_logTraceMsg0, "rdbId....... %d", (int)rdbId );
logTrace( g_conf.m_logTraceMsg0, "key size.... %d", (int)ks );
logTrace( g_conf.m_logTraceMsg0, "startFileNum %" PRId32, startFileNum );
logTrace( g_conf.m_logTraceMsg0, "numFiles.... %" PRId32, numFiles );
}
// error set from XmlDoc::cacheTermLists()?
if ( g_errno ) {
logTrace( g_conf.m_logTraceMsg0, "END. Invalid collection" );
log(LOG_ERROR,"%s:%s:%d: call sendErrorReply.", __FILE__, __func__, __LINE__);
us->sendErrorReply ( slot , EBADRDBID );
return;
}
// . get the rdb we need to get the RdbList from
// . returns NULL and sets g_errno on error
//Msg0 msg0;
//Rdb *rdb = msg0.getRdb ( rdbId );
Rdb *rdb = getRdbFromId ( rdbId );
if ( ! rdb ) {
logTrace( g_conf.m_logTraceMsg0, "END. Invalid rdbId" );
log(LOG_ERROR,"%s:%s:%d: call sendErrorReply.", __FILE__, __func__, __LINE__);
us->sendErrorReply ( slot , EBADRDBID );
return;
}
// keep track of stats
rdb->readRequestGet ( requestSize );
// . do a local get
// . create a msg5 to get the list
State00 *st0 ;
try { st0 = new (State00); }
catch ( ... ) {
g_errno = ENOMEM;
log("Msg0: new(%" PRId32"): %s",
(int32_t)sizeof(State00),mstrerror(g_errno));
log(LOG_ERROR,"%s:%s:%d: call sendErrorReply.", __FILE__, __func__, __LINE__);
us->sendErrorReply ( slot , g_errno );
return;
}
mnew ( st0 , sizeof(State00) , "State00" );
// timing debug
if ( g_conf.m_logTimingNet )
st0->m_startTime = gettimeofdayInMilliseconds();
// save slot in state
st0->m_slot = slot;
// save udp server to send back reply on
st0->m_us = us;
// init this one
st0->m_niceness = niceness;
st0->m_rdbId = rdbId;
QUICKPOLL(niceness);
// debug msg
if ( maxCacheAge != 0 && ! addToCache ) {
log( LOG_LOGIC, "net: msg0: check but don't add... rdbid=%" PRId32".", ( int32_t ) rdbId );
}
// . if this request came over on the high priority udp server
// make sure the priority gets passed along
// . return if this blocks
// . we'll call sendReply later
if ( ! st0->m_msg5.getList ( rdbId ,
collnum ,
&st0->m_list ,
startKey ,
endKey ,
minRecSizes ,
includeTree , // include tree?
addToCache , // addToCache?
maxCacheAge ,
startFileNum ,
numFiles ,
st0 ,
gotListWrapper ,
niceness ,
doErrorCorrection ,
NULL , // cacheKeyPtr
0 , // retryNum
2 , // maxRetries
true , // compensateForMerge
syncPoint ,
false,
allowPageCache ) ) {
logTrace( g_conf.m_logTraceMsg0, "END. m_msg5.getList returned false" );
return;
}
// call wrapper ouselves
logTrace( g_conf.m_logTraceMsg0, "Calling gotListWrapper" );
gotListWrapper ( st0 , NULL , NULL );
logTrace( g_conf.m_logTraceMsg0, "END" );
}
// . return false if blocked, true otherwise
// . set g_errno on error
// . read list of keys in [startKey,endKey] range
// . read at least "minRecSizes" bytes of keys in that range
// . the "m_endKey" of resulting, merged list may have a smaller endKey
// than the argument, "endKey" due to limitation by "minRecSizes"
// . resulting list will contain ALL keys between ITS [m_startKey,m_endKey]
// . final merged list "should" try to have a size of at least "minRecSizes"
// but due to negative/postive rec elimination may be less
// . the endKey of the lists we read may be <= "endKey" provided
// . we try to shrink the endKey if minRecSizes is >= 0 in order to
// avoid excessive reading
// . by shrinking the endKey we cannot take into account the size of deleted
// records, so therefore we may fall short of "minRecSizes" in actuality,
// in fact, the returned list may even be empty with a shrunken endKey
// . we merge all lists read from disk into the provided "list"
// . caller should call Msg3.getList(int32_t i) and Msg3:getNumLists() to retrieve
// . this makes the query engine faster since we don't need to merge the docIds
// and can just send them across the network separately and they will be
// hashed into IndexTable's table w/o having to do time-wasting merging.
// . caller can specify array of filenums to read from so incremental syncing
// in Sync class can just read from titledb*.dat files that were formed
// since the last sync point.
bool Msg3::readList ( char rdbId ,
collnum_t collnum ,
const char *startKeyArg ,
const char *endKeyArg ,
int32_t minRecSizes , // max size of scan
int32_t startFileNum , // first file to scan
int32_t numFiles , // rel. to startFileNum
void *state , // for callback
void (* callback ) ( void *state ) ,
int32_t niceness ,
int32_t retryNum ,
int32_t maxRetries ,
bool compensateForMerge ,
bool justGetEndKey ,
bool allowPageCache ,
bool hitDisk ) {
// set this to true to validate
m_validateCache = false;//true;
// clear, this MUST be done so if we return true g_errno is correct
g_errno = 0;
// assume lists are not checked for corruption
m_listsChecked = false;
// warn
if ( minRecSizes < -1 ) {
log(LOG_LOGIC,"db: Msg3 got minRecSizes of %" PRId32", changing "
"to -1.",minRecSizes);
minRecSizes = -1;
}
// reset m_alloc and data in all lists in case we are a re-call
reset();
// warning
if ( collnum < 0 ) log(LOG_LOGIC,"net: NULL collection. msg3.");
// remember the callback
m_rdbId = rdbId;
m_collnum = collnum;
m_callback = callback;
m_state = state;
m_niceness = niceness;
m_numScansCompleted = 0;
m_retryNum = retryNum;
m_maxRetries = maxRetries;
m_compensateForMerge = compensateForMerge;
m_allowPageCache = allowPageCache;
m_hitDisk = hitDisk;
m_hadCorruption = false;
// get keySize of rdb
m_ks = getKeySizeFromRdbId ( m_rdbId );
// reset the group error
m_errno = 0;
// . reset all our lists
// . these are reset in call the RdbScan::setRead() below
//for ( int32_t i = 0 ; i < MAX_RDB_FILES ; i++ ) m_lists[i].reset();
// . ensure startKey last bit clear, endKey last bit set
// . no! this warning is now only in Msg5
// . if RdbMerge is merging some files, not involving the root
// file, then we can expect to get a lot of unmatched negative recs.
// . as a consequence, our endKeys may often be negative. This means
// it may not annihilate with the positive key, but we should only
// miss like this at the boundaries of the lists we fetch.
// . so in that case RdbList::merge will stop merging once the
// minRecSizes limit is reached even if it means ending on a negative
// rec key
//if ( (startKey.n0 & 0x01) == 0x01 )
if ( !KEYNEG(startKeyArg) )
log(LOG_REMIND,"net: msg3: StartKey lastbit set.");
if ( KEYNEG(endKeyArg) )
log(LOG_REMIND,"net: msg3: EndKey lastbit clear.");
// declare vars here becaues of 'goto skip' below
int32_t mergeFileNum = -1 ;
int32_t max ;
// get base, returns NULL and sets g_errno to ENOCOLLREC on error
RdbBase *base = getRdbBase( m_rdbId, m_collnum );
if ( ! base ) {
return true;
}
// store the file numbers in the array, these are the files we read
m_numFileNums = 0;
// save startFileNum here, just for recall
m_startFileNum = startFileNum;
m_numFiles = numFiles;
// . if we have a merge going on, we may have to change startFileNum
// . if some files get unlinked because merge completes then our
// reads will detect the error and loop back here
// . we launch are reads right after this without giving up the cpu
// and we use file descriptors, so any changes to Rdb::m_files[]
// should not hurt us
// . WARNING: just make sure you don't lose control of cpu until after
// you call RdbScan::set()
// . we use hasMergeFile() instead of isMerging() because he may not
// be merging cuz he got suspended or he restarted and
// hasn't called attemptMerge() yet, but he may still contain it
if ( g_conf.m_logDebugQuery )
log(LOG_DEBUG,
"net: msg3: "
"c=%" PRId32" hmf=%" PRId32" sfn=%" PRId32" msfn=%" PRId32" nf=%" PRId32" db=%s.",
(int32_t)compensateForMerge,(int32_t)base->hasMergeFile(),
(int32_t)startFileNum,(int32_t)base->m_mergeStartFileNum-1,
(int32_t)numFiles,base->m_dbname);
int32_t pre = -10;
if ( compensateForMerge && base->hasMergeFile() &&
startFileNum >= base->m_mergeStartFileNum - 1 &&
(startFileNum > 0 || numFiles != -1) ) {
// now also include the file being merged into, but only
// if we are reading from a file being merged...
if ( startFileNum < base->m_mergeStartFileNum +
base->m_numFilesToMerge - 1 )
//m_fileNums [ m_numFileNums++ ] =
// base->m_mergeStartFileNum - 1;
pre = base->m_mergeStartFileNum - 1;
// debug msg
if ( g_conf.m_logDebugQuery )
log(LOG_DEBUG,
"net: msg3: startFileNum from %" PRId32" to %" PRId32" (mfn=%" PRId32")",
startFileNum,startFileNum+1,mergeFileNum);
// if merge file was inserted before us, inc our file number
startFileNum++;
}
// adjust num files if we need to, as well
if ( compensateForMerge && base->hasMergeFile() &&
startFileNum < base->m_mergeStartFileNum - 1 &&
numFiles != -1 &&
startFileNum + numFiles - 1 >= base->m_mergeStartFileNum - 1 ) {
// debug msg
if ( g_conf.m_logDebugQuery )
log(LOG_DEBUG,"net: msg3: numFiles up one.");
// if merge file was inserted before us, inc our file number
numFiles++;
}
// . how many rdb files does this base have?
// . IMPORTANT: this can change since files are unstable because they
// might have all got merged into one!
// . so do this check to make sure we're safe... especially if
// there was an error before and we called readList() on ourselves
max = base->getNumFiles();
// -1 means we should scan ALL the files in the base
if ( numFiles == -1 ) numFiles = max;
// limit it by startFileNum, however
if ( numFiles > max - startFileNum ) numFiles = max - startFileNum;
// set g_errno and return true if it is < 0
if ( numFiles < 0 ) {
log(LOG_LOGIC,
"net: msg3: readList: numFiles = %" PRId32" < 0 (max=%" PRId32")(sf=%" PRId32")",
numFiles , max , startFileNum );
g_errno = EBADENGINEER;
// force core dump
char *xx=NULL;*xx=0;
return true;
}
// . allocate buffer space
// . m_scans, m_startpg, m_endpg, m_hintKeys, m_hintOffsets,
// m_fileNums, m_lists
int32_t chunk = sizeof(RdbScan) + // m_scans
4 + // m_startpg
4 + // m_endpg
//sizeof(key_t) + // m_hintKeys
m_ks + // m_hintKeys
4 + // m_hintOffsets
4 + // m_fileNums
sizeof(RdbList) ; // m_lists
int32_t nn = numFiles;
if ( pre != -10 ) nn++;
m_numChunks = nn;
int32_t need = nn * (chunk);
m_alloc = m_buf;
if ( need > (int32_t)MSG3_BUF_SIZE ) {
m_allocSize = need;
m_alloc = (char *)mcalloc ( need , "Msg3" );
if ( ! m_alloc ) {
log("disk: Could not allocate %" PRId32" bytes read "
"structures to read %s.",need,base->m_dbname);
return true;
}
}
char *p = m_alloc;
m_scans = (RdbScan *)p; p += nn * sizeof(RdbScan);
m_startpg = (int32_t *)p; p += nn * 4;
m_endpg = (int32_t *)p; p += nn * 4;
//m_hintKeys = (key_t *)p; p += nn * sizeof(key_t);
m_hintKeys = (char *)p; p += nn * m_ks;
m_hintOffsets = (int32_t *)p; p += nn * 4;
m_fileNums = (int32_t *)p; p += nn * 4;
m_lists = (RdbList *)p; p += nn * sizeof(RdbList);
// sanity check
if ( p - m_alloc != need ) {
log(LOG_LOGIC,"disk: Bad malloc in Msg3.cpp.");
char *xx = NULL; *xx = 0;
}
// call constructors
for ( int32_t i = 0 ; i < nn ; i++ ) m_lists[i].constructor();
// make fix from up top
if ( pre != -10 ) m_fileNums [ m_numFileNums++ ] = pre;
// store them all
for ( int32_t i = startFileNum ; i < startFileNum + numFiles ; i++ )
m_fileNums [ m_numFileNums++ ] = i;
// . remove file nums that are being unlinked after a merge now
// . keep it here (below skip: label) so sync point reads can use it
int32_t n = 0;
for ( int32_t i = 0 ; i < m_numFileNums ; i++ ) {
// skip those that are being unlinked after the merge
if ( base->m_isUnlinking &&
m_fileNums[i] >= base->m_mergeStartFileNum &&
m_fileNums[i] < base->m_mergeStartFileNum +
base->m_numFilesToMerge )
continue;
// otherwise, keep it
m_fileNums[n++] = m_fileNums[i];
}
m_numFileNums = n;
// . if root file is being merged, he's file #0, & root file is file #1
// . this is a hack so caller gets what he wants
//if ( startFileNum == 0 && base->getFileId(0) == 0 && numFiles == 1 )
// numFiles = 2;
// remember the file range we should scan
m_numScansStarted = 0;
m_numScansCompleted = 0;
//m_startKey = startKey;
//m_endKey = endKey;
//m_constrainKey = endKey; // set in case justGetEndKey is true
KEYSET(m_startKey,startKeyArg,m_ks);
KEYSET(m_endKey,endKeyArg,m_ks);
KEYSET(m_constrainKey,endKeyArg,m_ks);//set incase justGetEndKey istrue
m_minRecSizes = minRecSizes;
m_compensateForMerge = compensateForMerge;
// bail if 0 files to scan -- no! need to set startKey/endKey
if ( numFiles == 0 ) return true;
// don't read anything if endKey < startKey
//if ( m_startKey > m_endKey ) return true;
if ( KEYCMP(m_startKey,m_endKey,m_ks)>0 ) return true;
// keep the original in tact in case g_errno == ETRYAGAIN
//m_endKeyOrig = endKey;
KEYSET(m_endKeyOrig,endKeyArg,m_ks);
m_minRecSizesOrig = minRecSizes;
// start reading at this key
m_fileStartKey = startKeyArg;
// start the timer, keep it fast for clusterdb though
if ( g_conf.m_logTimingDb ) m_startTime = gettimeofdayInMilliseconds();
// translate base to an id, for the sake of m_msg0
//char baseId = m_msg0->getRdbId ( base );
// map ptrs
RdbMap **maps = base->getMaps();
// . we now boost m_minRecSizes to account for negative recs
// . but not if only reading one list, cuz it won't get merged and
// it will be too big to send back
if ( m_numFileNums > 1 ) compensateForNegativeRecs ( base );
// . often endKey is too big for an efficient read of minRecSizes bytes
// because we end up reading too much from all the files
// . this will set m_startpg[i], m_endpg[i] for each RdbScan/RdbFile
// to ensure we read "minRecSizes" worth of records, not much more
// . returns the new endKey for all ranges
// . now this just overwrites m_endKey
//m_endKey = setPageRanges ( base ,
setPageRanges ( base ,
m_fileNums ,
m_numFileNums ,
m_fileStartKey , // start reading @ key
m_endKey , // stop reading @ key
m_minRecSizes );
// . NEVER let m_endKey be a negative key, because it will
// always be unmatched, since delbit is cleared
// . adjusting it here ensures our generated hints are valid
// . we will use this key to call constrain() with
//m_constrainKey = m_endKey;
//if ( ( m_constrainKey.n0 & 0x01) == 0x00 )
// m_constrainKey -= (uint32_t)1;
KEYSET(m_constrainKey,m_endKey,m_ks);
if ( KEYNEG(m_constrainKey) )
KEYSUB(m_constrainKey,m_ks);
// Msg5 likes to get the endkey for getting the list from the tree
if ( justGetEndKey ) return true;
// sanity check
if ( m_numFileNums > nn ) {
log(LOG_LOGIC,"disk: Failed sanity check in Msg3.");
char *xx = NULL; *xx = 0;
}
// debug msg
//log("msg3 getting list (msg5=%" PRIu32")",m_state);
// . MDW removed this -- go ahead an end on a delete key
// . RdbMerge might not pick it up this round, but oh well
// . so we can have both positive and negative co-existing in same file
// make sure the last bit is set so we don't end on a delete key
//m_endKey.n0 |= 0x01LL;
// . now start reading/scanning the files
// . our m_scans array starts at 0
for ( int32_t i = 0 ; i < m_numFileNums ; i++ ) {
// get the page range
//int32_t p1 = m_startpg [ i ];
//int32_t p2 = m_endpg [ i ];
//#ifdef GBSANITYCHECK
int32_t fn = m_fileNums[i];
// this can happen somehow!
if ( fn < 0 ) {
log(LOG_LOGIC,"net: msg3: fn=%" PRId32". Bad engineer.",fn);
continue;
}
// sanity check
if ( i > 0 && m_fileNums[i-1] >= fn ) {
log(LOG_LOGIC,
"net: msg3: files must be read in order "
"from oldest to newest so RdbList::indexMerge_r "
"works properly. Otherwise, corruption will "
"result. ");
char *xx = NULL; *xx = 0;
return true;
}
// . sanity check?
// . no, we must get again since we turn on endKey's last bit
int32_t p1 , p2;
maps[fn]->getPageRange ( m_fileStartKey ,
m_endKey ,
&p1 ,
&p2 ,
NULL );
//if ( p1 != p1c || p2 != p2c ) {
// fprintf(stderr,"Msg3::bad page range\n");
// sleep(50000);
//}
// sanity check, each endpg's key should be > endKey
//if ( p2 < maps[fn]->getNumPages() &&
// maps[fn]->getKey ( p2 ) <= m_endKey ) {
// fprintf(stderr,"Msg3::bad page range 2\n");
// sleep(50000);
//}
//#endif
//int32_t p1 , p2;
//maps[fn]->getPageRange (startKey,endKey,minRecSizes,&p1,&p2);
// now get some read info
int64_t offset = maps[fn]->getAbsoluteOffset ( p1 );
int32_t bytesToRead = maps[fn]->getRecSizes ( p1, p2, false);
// max out the endkey for this list
// debug msg
//#ifdef _DEBUG_
//if ( minRecSizes == 2000000 )
//log("Msg3:: reading %" PRId32" bytes from file #%" PRId32,bytesToRead,i);
//#endif
// inc our m_numScans
m_numScansStarted++;
// . keep stats on our disk accesses
// . count disk seeks (assuming no fragmentation)
// . count disk bytes read
if ( bytesToRead > 0 ) {
base->m_rdb->didSeek ( );
base->m_rdb->didRead ( bytesToRead );
}
// . the startKey may be different for each RdbScan class
// . RdbLists must have all keys within their [startKey,endKey]
// . therefore set startKey individually from first page in map
// . this endKey must be >= m_endKey
// . this startKey must be < m_startKey
//key_t startKey = maps[fn]->getKey ( p1 );
//key_t endKey = maps[fn]->getKey ( p2 );
char startKey2 [ MAX_KEY_BYTES ];
char endKey2 [ MAX_KEY_BYTES ];
maps[fn]->getKey ( p1 , startKey2 );
maps[fn]->getKey ( p2 , endKey2 );
//char *startKey = maps[fn]->getKeyPtr ( p1 );
//char *endKey = maps[fn]->getKeyPtr ( p2 );
// store in here
m_startpg [ i ] = p1;
m_endpg [ i ] = p2;
// . we read UP TO that endKey, so reduce by 1
// . but iff p2 is NOT the last page in the map/file
// . maps[fn]->getKey(lastPage) will return the LAST KEY
// and maps[fn]->getOffset(lastPage) the length of the file
//if ( maps[fn]->getNumPages()!=p2) endKey -=(uint32_t)1;
if ( maps[fn]->getNumPages() != p2 ) KEYSUB(endKey2,m_ks);
// otherwise, if we're reading all pages, then force the
// endKey to virtual inifinite
//else endKey.setMax();
else KEYMAX(endKey2,m_ks);
// . set up the hints
// . these are only used if we are only reading from 1 file
// . these are used to call constrain() so we can constrain
// the end of the list w/o looping through all the recs
// in the list
int32_t h2 = p2 ;
// decrease by one page if we're on the last page
if ( h2 > p1 && maps[fn]->getNumPages() == h2 ) h2--;
// . decrease hint page until key is <= endKey on that page
// AND offset is NOT -1 because the old way would give
// us hints passed the endkey
// . also decrease so we can constrain on minRecSizes in
// case we're the only list being read
// . use >= m_minRecSizes instead of >, otherwise we may
// never be able to set "size" in RdbList::constrain()
// because "p" could equal "maxPtr" right away
while ( h2 > p1 &&
//( maps[fn]->getKey (h2) > m_constrainKey ||
(KEYCMP(maps[fn]->getKeyPtr(h2),m_constrainKey,m_ks)>0||
maps[fn]->getOffset(h2) == -1 ||
maps[fn]->getAbsoluteOffset(h2) - offset >=
m_minRecSizes ) )
h2--;
// now set the hint
m_hintOffsets [ i ] = maps[fn]->getAbsoluteOffset ( h2 ) -
maps[fn]->getAbsoluteOffset ( p1 ) ;
//m_hintKeys [ i ] = maps[fn]->getKey ( h2 );
KEYSET(&m_hintKeys[i*m_ks],maps[fn]->getKeyPtr(h2),m_ks);
// reset g_errno before calling setRead()
g_errno = 0;
// . this fix is now in RdbList::checklist_r()
// . we can now have dup keys, so, we may read in
// a rec with key "lastMinKey" even though we don't read
// in the first key on the end page, so don't subtract 1...
//if ( endKey != m_endKeyOrig )
// endKey += (uint32_t) 1;
// timing debug
if ( g_conf.m_logTimingDb )
log(LOG_TIMING,
"net: msg: reading %" PRId32" bytes from %s file #%" PRId32" "
"(niceness=%" PRId32")",
bytesToRead,base->m_dbname,i,m_niceness);
// log huge reads, those hurt us
if ( bytesToRead > 150000000 ) {
logf(LOG_INFO,"disk: Reading %" PRId32" bytes at offset %" PRId64" "
"from %s.",
bytesToRead,offset,base->m_dbname);
}
// if any keys in the map are the same report corruption
char tmpKey [16];
char lastTmpKey[16];
int32_t ccount = 0;
if ( bytesToRead > 10000000 &&
bytesToRead / 2 > m_minRecSizes &&
base->m_fixedDataSize >= 0 ) {
for ( int32_t pn = p1 ; pn <= p2 ; pn++ ) {
maps[fn]->getKey ( pn , tmpKey );
if ( KEYCMP(tmpKey,lastTmpKey,m_ks) == 0 )
ccount++;
gbmemcpy(lastTmpKey,tmpKey,m_ks);
}
}
if ( ccount > 10 ) {
logf(LOG_INFO,"disk: Reading %" PRId32" bytes from %s file #"
"%" PRId32" when min "
"required is %" PRId32". Map is corrupt and has %" PRId32" "
"identical consecutive page keys because the "
"map was \"repaired\" because out of order keys "
"in the index.",
(int32_t)bytesToRead,
base->m_dbname,fn,
(int32_t)m_minRecSizes,
(int32_t)ccount);
m_numScansCompleted++;
m_errno = ECORRUPTDATA;
m_hadCorruption = true;
//m_maxRetries = 0;
break;
}
////////
//
// try to get from PAGE CACHE
//
////////
BigFile *ff = base->getFile(m_fileNums[i]);
RdbCache *rpc = getDiskPageCache ( m_rdbId );
if ( ! m_allowPageCache ) rpc = NULL;
// . vfd is unique 64 bit file id
// . if file is opened vfd is -1, only set in call to open()
int64_t vfd = ff->getVfd();
key192_t ck = makeCacheKey ( vfd , offset, bytesToRead);
char *rec; int32_t recSize;
bool inCache = false;
if ( rpc && vfd != -1 && ! m_validateCache )
inCache = rpc->getRecord ( (collnum_t)0 , // collnum
(char *)&ck ,
&rec ,
&recSize ,
true , // copy?
-1 , // maxAge, none
true ); // inccounts?
m_scans[i].m_inPageCache = false;
if ( inCache ) {
m_scans[i].m_inPageCache = true;
m_numScansCompleted++;
// now we have to store this value, 6 or 12 so
// we can modify the hint appropriately
m_scans[i].m_shifted = *rec;
m_lists[i].set ( rec +1,
recSize-1 ,
rec , // alloc
recSize , // allocSize
startKey2 ,
endKey2 ,
base->m_fixedDataSize ,
true , // owndata
base->useHalfKeys() ,
getKeySizeFromRdbId ( m_rdbId ) );
continue;
}
// . do the scan/read of file #i
// . this returns false if blocked, true otherwise
// . this will set g_errno on error
bool done = m_scans[i].setRead (base->getFile(m_fileNums[i]),
base->m_fixedDataSize ,
offset ,
bytesToRead ,
startKey2 ,
endKey2 ,
m_ks ,
&m_lists[i] ,
this ,
doneScanningWrapper ,
base->useHalfKeys() ,
m_rdbId,
m_niceness ,
m_allowPageCache ,
m_hitDisk ) ;
// . damn, usually the above will indirectly launch a thread
// to do the reading, but it sets g_errno to EINTR,
// "interrupted system call"!
// . i guess the thread does the read w/o blocking and then
// queues the signal on g_loop's queue before it exits
// . try ignoring, and keep going
if ( g_errno == EINTR ) {
log("net: Interrupted system call while reading file. "
"Ignoring.");
g_errno = 0;
}
// debug msg
//fprintf(stderr,"Msg3:: reading %" PRId32" bytes from file #%" PRId32","
// "done=%" PRId32",offset=%" PRId64",g_errno=%s,"
// "startKey=n1=%" PRIu32",n0=%" PRIu64", "
// "endKey=n1=%" PRIu32",n0=%" PRIu64"\n",
// bytesToRead,i,(int32_t)done,offset,mstrerror(g_errno),
// m_startKey,m_endKey);
//if ( bytesToRead == 0 )
// fprintf(stderr,"shit\n");
// if it did not block then it completed, so count it
if ( done ) m_numScansCompleted++;
// break on an error, and remember g_errno in case we block
if ( g_errno && g_errno != ENOTHREADSLOTS ) {
int32_t tt = LOG_WARN;
if ( g_errno == EFILECLOSED ) tt = LOG_INFO;
log(tt,"disk: Reading %s had error: %s.",
base->m_dbname, mstrerror(g_errno));
m_errno = g_errno;
break;
}
}
// debug test
//if ( rand() % 100 <= 10 ) m_errno = EIO;
// if we blocked, return false
if ( m_numScansCompleted < m_numScansStarted ) return false;
// . if all scans completed without blocking then wrap it up & ret true
// . doneScanning may now block if it finds data corruption and must
// get the list remotely
return doneScanning();
}
// . 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;
}
// . reply to a request for an RdbList
// . MUST call g_udpServer::sendReply or sendErrorReply() so slot can
// be destroyed
void handleRequest0 ( UdpSlot *slot , int32_t netnice ) {
// if niceness is 0, use the higher priority udpServer
UdpServer *us = &g_udpServer;
//if ( netnice == 0 ) us = &g_udpServer2;
// get the request
char *request = slot->m_readBuf;
int32_t requestSize = slot->m_readBufSize;
// collection is now stored in the request, so i commented this out
//if ( requestSize != MSG0_REQ_SIZE ) {
// log("net: Received bad data request size of %"INT32" bytes. "
// "Should be %"INT32".", requestSize ,(int32_t)MSG0_REQ_SIZE);
// us->sendErrorReply ( slot , EBADREQUESTSIZE );
// return;
//}
// parse the request
char *p = request;
int64_t syncPoint = *(int64_t *)p ; p += 8;
//key_t startKey = *(key_t *)p ; p += sizeof(key_t);
//key_t endKey = *(key_t *)p ; p += sizeof(key_t);
int32_t minRecSizes = *(int32_t *)p ; p += 4;
int32_t startFileNum = *(int32_t *)p ; p += 4;
int32_t numFiles = *(int32_t *)p ; p += 4;
int32_t maxCacheAge = *(int32_t *)p ; p += 4;
char rdbId = *p++;
char addToCache = *p++;
char doErrorCorrection = *p++;
char includeTree = *p++;
// this was messing up our niceness conversion logic
int32_t niceness = slot->m_niceness;//(int32_t)(*p++);
// still need to skip it though!
p++;
bool allowPageCache = (bool)(*p++);
char ks = getKeySizeFromRdbId ( rdbId );
char *startKey = p; p+=ks;
char *endKey = p; p+=ks;
// then null terminated collection
//char *coll = p;
collnum_t collnum = *(collnum_t *)p; p += sizeof(collnum_t);
CollectionRec *xcr = g_collectiondb.getRec ( collnum );
if ( ! xcr ) g_errno = ENOCOLLREC;
// error set from XmlDoc::cacheTermLists()?
if ( g_errno ) {
us->sendErrorReply ( slot , EBADRDBID ); return;}
// is this being called from callWaitingHandlers()
//bool isRecall = (netnice == 99);
// . get the rdb we need to get the RdbList from
// . returns NULL and sets g_errno on error
//Msg0 msg0;
//Rdb *rdb = msg0.getRdb ( rdbId );
Rdb *rdb = getRdbFromId ( rdbId );
if ( ! rdb ) {
us->sendErrorReply ( slot , EBADRDBID ); return;}
// keep track of stats
rdb->readRequestGet ( requestSize );
/*
// keep track of stats
if ( ! isRecall ) rdb->readRequestGet ( requestSize );
int64_t singleDocId2 = 0LL;
if ( rdbId == RDB_POSDB && maxCacheAge ) {
int64_t d1 = g_posdb.getDocId(startKey);
int64_t d2 = g_posdb.getDocId(endKey);
if ( d1+1 == d2 ) singleDocId2 = d1;
}
// have we parsed this docid and cached its termlists?
bool shouldBeCached2 = false;
if ( singleDocId2 &&
isDocIdInTermListCache ( singleDocId2 , coll ) )
shouldBeCached2 = true;
// if in the termlist cache, send it back right away
char *trec;
int32_t trecSize;
if ( singleDocId2 &&
getRecFromTermListCache(coll,
startKey,
endKey,
maxCacheAge,
&trec,
&trecSize) ) {
// if in cache send it back!
us->sendReply_ass(trec,trecSize,trec,trecSize,slot);
return;
}
// if should be cached but was not found then it's probably a
// synonym form not in the doc content. make an empty list then.
if ( shouldBeCached2 ) {
// send back an empty termlist
us->sendReply_ass(NULL,0,NULL,0,slot);
return;
}
// MUST be in termlist cache! if not in there it is a probably
// a synonym term termlist of a word in the doc.
if ( isRecall ) {
// send back an empty termlist
us->sendReply_ass(NULL,0,NULL,0,slot);
return;
}
// init waiting table?
static bool s_waitInit = false;
if ( ! s_waitInit ) {
// do not repeat
s_waitInit = true;
// niceness = 0
if ( ! g_waitingTable.set(8,4,2048,NULL,0,true,0,"m5wtbl")){
log("msg5: failed to init waiting table");
// error kills us!
us->sendErrorReply ( slot , EBADRDBID );
return;
}
}
// wait in waiting table?
if ( singleDocId2 && g_waitingTable.isInTable ( &singleDocId2 ) ) {
g_waitingTable.addKey ( &singleDocId2 , &slot );
return;
}
// if it's for a special gbdocid: query then cache ALL termlists
// for this docid into g_termListCache right now
if ( singleDocId2 ) {
// have all further incoming requests for this docid
// wait in the waiting table
g_waitingTable.addKey ( &singleDocId2 , &slot );
// load the title rec and store its posdb termlists in cache
XmlDoc *xd;
try { xd = new ( XmlDoc ); }
catch ( ... ) {
g_errno = ENOMEM;
us->sendErrorReply ( slot , g_errno );
return;
}
mnew ( xd, sizeof(XmlDoc),"msg0xd");
// always use niceness 1 now even though we use niceness 0
// to make the cache hits fast
//niceness = 1;
// . load the old title rec first and just recycle all
// . typically there might be a few hundred related docids
// each with 50,000 matching queries on average to evaluate
// with the gbdocid:xxxx| restriction?
if ( ! xd->set3 ( singleDocId2 , coll , niceness ) ) {
us->sendErrorReply ( slot , g_errno ); return;}
// init the new xmldoc
xd->m_callback1 = callWaitingHandlers;
xd->m_state = xd;
// . if this blocks then return
// . should call loadOldTitleRec() and get JUST the posdb recs
// by setting m_useTitledb, etc. to false. then it should
// make posdb termlists with the compression using
// RdbList::addRecord() and add those lists to
// g_termListCache
if ( ! xd->cacheTermLists ( ) ) return;
// otherwise, it completed right away!
callWaitingHandlers ( xd );
return;
}
*/
/*
// init special sectiondb cache?
if ( rdbId == RDB_SECTIONDB && ! s_initCache ) {
// try to init cache
if ( ! s_sectiondbCache.init ( 20000000 , // 20MB max mem
-1 , // fixed data size
false , // support lists?
20000 , // 20k max recs
false , // use half keys?
"secdbche", // dbname
false, // load from disk?
sizeof(key128_t), //cachekeysize
0 , // data key size
20000 )) // numPtrs max
log("msg0: failed to init sectiondb cache: %s",
mstrerror(g_errno));
else
s_initCache = true;
}
// check the sectiondb cache
if ( rdbId == RDB_SECTIONDB ) {
//int64_t sh48 = g_datedb.getTermId((key128_t *)startKey);
// use the start key now!!!
char *data;
int32_t dataSize;
if (s_sectiondbCache.getRecord ( coll,
startKey,//&sh48,
&data,
&dataSize,
true, // docopy?
600, // maxage (10 mins)
true, // inc counts?
NULL, // cachedtime
true // promoteRec?
)){
// debug
//log("msg0: got sectiondblist in cache datasize=%"INT32"",
// dataSize);
// send that back
g_udpServer.sendReply_ass ( data ,
dataSize ,
data ,
dataSize ,
slot ,
60 ,
NULL ,
doneSending_ass ,
-1 ,
-1 ,
true );
return;
}
}
*/
// . do a local get
// . create a msg5 to get the list
State00 *st0 ;
try { st0 = new (State00); }
catch ( ... ) {
g_errno = ENOMEM;
log("Msg0: new(%"INT32"): %s",
(int32_t)sizeof(State00),mstrerror(g_errno));
us->sendErrorReply ( slot , g_errno );
return;
}
mnew ( st0 , sizeof(State00) , "State00" );
// timing debug
if ( g_conf.m_logTimingNet )
st0->m_startTime = gettimeofdayInMilliseconds();
// save slot in state
st0->m_slot = slot;
// save udp server to send back reply on
st0->m_us = us;
// init this one
st0->m_niceness = niceness;
st0->m_rdbId = rdbId;
QUICKPOLL(niceness);
// debug msg
if ( maxCacheAge != 0 && ! addToCache )
log(LOG_LOGIC,"net: msg0: check but don't add... rdbid=%"INT32".",
(int32_t)rdbId);
// . if this request came over on the high priority udp server
// make sure the priority gets passed along
// . return if this blocks
// . we'll call sendReply later
if ( ! st0->m_msg5.getList ( rdbId ,
collnum ,
&st0->m_list ,
startKey ,
endKey ,
minRecSizes ,
includeTree , // include tree?
addToCache , // addToCache?
maxCacheAge ,
startFileNum ,
numFiles ,
st0 ,
gotListWrapper ,
niceness ,
doErrorCorrection ,
NULL , // cacheKeyPtr
0 , // retryNum
2 , // maxRetries
true , // compensateForMerge
syncPoint ,
NULL,//&st0->m_msg5b ,
false,
allowPageCache ) )
return;
// call wrapper ouselves
gotListWrapper ( st0 , NULL , NULL );
}
// . return false if blocked, true otherwise
// . sets g_errno on error
bool Msg1::sendData ( unsigned long shardNum, char *listData , long listSize) {
// debug msg
//log("sendData: mcast=%lu listSize=%li",
// (long)&m_mcast,(long)listSize);
// bail if this is an interface machine, don't write to the main
if ( g_conf.m_interfaceMachine ) return true;
// return true if no data
if ( listSize == 0 ) return true;
// how many hosts in this group
//long numHosts = g_hostdb.getNumHostsPerShard();
// . NOTE: for now i'm removing this until I handle ETRYAGAIN errors
// properly... by waiting and retrying...
// . if this is local data just for us just do an addList to OUR rdb
/*
if ( groupId == g_hostdb.m_groupId && numHosts == 1 ) {
// this sets g_errno on error
Msg0 msg0;
Rdb *rdb = msg0.getRdb ( (char) m_rdbId );
if ( ! rdb ) return true;
// make a list from this data
RdbList list;
list.set (listData,listSize,listSize,rdb->getFixedDataSize(),
false) ; // ownData?
// this returns false and sets g_errno on error
rdb->addList ( &list );
// . if we got a ETRYAGAIN cuz the buffer we add to was full
// then we should sleep and try again!
// . return false cuz this blocks for a period of time
// before trying again
if ( g_errno == ETRYAGAIN ) {
// try adding again in 1 second
registerSleepCallback ( 1000, slot, tryAgainWrapper1 );
// return now
return false;
}
// . always return true cuz we did not block
// . g_errno may be set
return true;
}
*/
// if the data is being added to our group, don't send ourselves
// a msg1, if we can add it right now
bool sendToSelf = true;
if ( shardNum == getMyShardNum() &&
! g_conf.m_interfaceMachine ) {
// get the rdb to which it belongs, use Msg0::getRdb()
Rdb *rdb = getRdbFromId ( (char) m_rdbId );
if ( ! rdb ) goto skip;
// key size
long ks = getKeySizeFromRdbId ( m_rdbId );
// reset g_errno
g_errno = 0;
// . make a list from this data
// . skip over the first 4 bytes which is the rdbId
// . TODO: embed the rdbId in the msgtype or something...
RdbList list;
// set the list
list.set ( listData ,
listSize ,
listData ,
listSize ,
rdb->getFixedDataSize() ,
false , // ownData?
rdb->useHalfKeys() ,
ks );
// note that
//log("msg1: local addlist niceness=%li",m_niceness);
// this returns false and sets g_errno on error
rdb->addList ( m_coll , &list , m_niceness );
// if titledb, add tfndb recs to map the title recs
//if ( ! g_errno && rdb == g_titledb.getRdb() && m_injecting )
// // this returns false and sets g_errno on error
// updateTfndb ( m_coll , &list , true , m_niceness);
// if no error, no need to use a Msg1 UdpSlot for ourselves
if ( ! g_errno ) sendToSelf = false;
else {
log("rdb: msg1 had error: %s",mstrerror(g_errno));
// this is messing up generate catdb's huge rdblist add
// why did we put it in there??? from msg9b.cpp
//return true;
}
QUICKPOLL(m_niceness);
// if we're the only one in the group, bail, we're done
if ( ! sendToSelf &&
g_hostdb.getNumHostsPerShard() == 1 ) return true;
}
skip:
// . make an add record request to multicast to a bunch of machines
// . this will alloc new space, returns NULL on failure
//char *request = makeRequest ( listData, listSize, groupId ,
//m_rdbId , &requestLen );
long collLen = gbstrlen ( m_coll );
// . returns NULL and sets g_errno on error
// . calculate total size of the record
// . 1 byte for rdbId, 1 byte for flags,
// then collection NULL terminated, then list
long requestLen = 1 + 1 + collLen + 1 + listSize ;
// make the request
char *request = (char *) mmalloc ( requestLen ,"Msg1" );
if ( ! request ) return true;
char *p = request;
// store the rdbId at top of request
*p++ = m_rdbId;
// then the flags
*p = 0;
if ( m_injecting ) *p |= 0x80;
p++;
// then collection name
memcpy ( p , m_coll , collLen );
p += collLen;
*p++ = '\0';
// sanity check
if ( collLen <= 0 ) {
log(LOG_LOGIC,"net: No collection specified for list add.");
//char *xx = NULL; *xx = 0;
g_errno = ENOCOLLREC;
return true;
}
//if ( m_deleteRecs ) request[1] |= 0x80;
//if ( m_overwriteRecs ) request[1] |= 0x40;
// store the list after coll
memcpy ( p , listData , listSize );
QUICKPOLL(m_niceness);
// debug msg
//if ( ! m_waitForReply ) // (m_rdbId == RDB_SPIDERDB ||
//m_rdbId == RDB_TFNDB) )
// // if we don't get here we lose it!!!!!!!!!!!!!!!!!!!!!
// log("using mcast=%lu rdbId=%li listData=%lu listSize=%lu "
// "gid=%lu",
// (long)&m_mcast,(long)m_rdbId,(long)listData,(long)listSize,
// groupId);
// for small packets
//long niceness = 2;
//if ( requestLen < TMPBUFSIZE - 32 ) niceness = 0;
//log("msg1: sending mcast niceness=%li",m_niceness);
// . multicast to all hosts in group "groupId"
// . multicast::send() returns false and sets g_errno on error
// . we return false if we block, true otherwise
// . will loop indefinitely if a host in this group is down
key_t k; k.setMin();
if ( m_mcast.send ( request , // sets mcast->m_msg to this
requestLen , // sets mcast->m_msgLen to this
0x01 , // msgType for add rdb record
true , // does multicast own msg?
shardNum , // group to send to (groupKey)
true , // send to whole group?
0 , // key is useless for us
this , // state data
NULL , // state data
gotReplyWrapper1 ,
60 , // timeout in secs
m_niceness , // niceness
false , // realtime
-1 , // first host to try
NULL , // replyBuf = NULL ,
0 , // replyBufMaxSize = 0 ,
true , // freeReplyBuf = true ,
false , // doDiskLoadBalancing = false ,
-1 , // no max cache age limit
//(key_t)0 , // cache key
k , // cache key
RDB_NONE , // bogus rdbId
-1 , // unknown minRecSizes read size
sendToSelf ))
return false;
QUICKPOLL(m_niceness);
// g_errno should be set
log("net: Had error when sending request to add data to %s in shard "
"#%lu: %s.", getDbnameFromId(m_rdbId),shardNum,mstrerror(g_errno));
return true;
}
