// . return false if blocked, true otherwise
// . set g_errno on error
// . list should be truncated, possible have all negative keys removed,
// and de-duped thanks to RdbList::indexMerge_r() and RdbList::merge_r()
bool RdbMerge::dumpList ( ) {
// return true on g_errno
if ( g_errno ) return true;
// . it's suspended so we count this as blocking
// . resumeMerge() will call getNextList() again, not dumpList() so
// don't advance m_startKey
if ( m_isSuspended ) {
m_isReadyToSave = true;
return false;
}
// . set the list to only those records that should be in our group
// . filter the records that don't belong in this group via groupId
//filterList ( &m_list );
// . compute the new m_startKey to get the next list from disk
// . m_list was formed via RdbList::merge()
// . m_list may be empty because of negative/positive collisions
// but there may still be data left
//m_startKey = m_list.getLastKey() ;
//m_list.getLastKey(m_startKey) ;
// if we use getLastKey() for this the merge completes but then
// tries to merge two empty lists and cores in the merge function
// because of that. i guess it relies on endkey rollover only and
// not on reading less than minRecSizes to determine when to stop
// doing the merge.
m_list.getEndKey(m_startKey) ;
//m_startKey += (uint32_t)1;
KEYADD(m_startKey,m_ks);
/////
//
// dedup for spiderdb before we dump it. try to save disk space.
//
/////
if ( m_rdbId == RDB_SPIDERDB )
// removeNegRecs? = false
dedupSpiderdbList(&m_list, false);
// if the startKey rolled over we're done
//if ( m_startKey.n0 == 0LL && m_startKey.n1 == 0 ) m_doneMerging=true;
if ( KEYCMP(m_startKey,KEYMIN(),m_ks)==0 ) m_doneMerging = true;
// debug msg
log(LOG_DEBUG,"db: Dumping list.");
// debug msg
//fprintf(stderr,"list startKey.n1=%" PRIu32",n0=%" PRIu64", endKey.n1=%" PRIu32",n0=%" PRIu64","
// " size=%" PRId32"\n",
// m_list.getStartKey().n1,
// m_list.getStartKey().n0,
// m_list.getLastKey().n1,
// m_list.getLastKey().n0, m_list.getListSize() );
// . send the whole list to the dump
// . it returns false if blocked, true otherwise
// . it sets g_errno on error
// . it calls dumpListWrapper when done dumping
// . return true if m_dump had an error or it did not block
// . if it gets a EFILECLOSED error it will keep retrying forever
return m_dump.dumpList ( &m_list , m_niceness , false/*recall?*/ ) ;
}
// . 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;
}
// . 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();
}
// . returns a new, smaller endKey
// . shrinks endKey while still preserving the minRecSizes requirement
// . this is the most confusing subroutine in the project
// . this now OVERWRITES endKey with the new one
//key_t Msg3::setPageRanges ( RdbBase *base ,
void Msg3::setPageRanges ( RdbBase *base ,
int32_t *fileNums ,
int32_t numFileNums ,
const char *startKey ,
char *endKey ,
int32_t minRecSizes ) {
// sanity check
//if ( m_ks != 12 && m_ks != 16 ) { char *xx=NULL;*xx=0; }
// get the file maps from the rdb
RdbMap **maps = base->getMaps();
// . initialize the startpg/endpg for each file
// . we read from the first offset on m_startpg to offset on m_endpg
// . since we set them equal that means an empty range for each file
for ( int32_t i = 0 ; i < numFileNums ; i++ ) {
int32_t fn = fileNums[i];
if ( fn < 0 ) { char *xx = NULL; *xx = 0; }
m_startpg[i] = maps[fn]->getPage( startKey );
m_endpg [i] = m_startpg[i];
}
// just return if minRecSizes 0 (no reading needed)
//if ( minRecSizes <= 0 ) return endKey ;
if ( minRecSizes <= 0 ) return;
// calculate minKey minus one
//key_t lastMinKey ;
char lastMinKey[MAX_KEY_BYTES];
char lastMinKeyIsValid = 0;
// loop until we find the page ranges that barely satisfy "minRecSizes"
loop:
// find the map whose next page has the lowest key
int32_t minpg = -1;
//key_t minKey;
char minKey[MAX_KEY_BYTES];
for ( int32_t i = 0 ; i < numFileNums ; i++ ) {
int32_t fn = fileNums[i];
// this guy is out of race if his end key > "endKey" already
//if ( maps[fn]->getKey ( m_endpg[i] ) > endKey ) continue;
if(KEYCMP(maps[fn]->getKeyPtr(m_endpg[i]),endKey,m_ks)>0)
continue;
// get the next page after m_endpg[i]
int32_t nextpg = m_endpg[i] + 1;
// if endpg[i]+1 == m_numPages then we maxed out this range
if ( nextpg > maps[fn]->getNumPages() ) continue;
// . but this may have an offset of -1
// . which means the page has no key starting on it and
// it's occupied by a rec which starts on a previous page
while ( nextpg < maps[fn]->getNumPages() &&
maps[fn]->getOffset ( nextpg ) == -1 ) nextpg++;
// . continue if his next page doesn't have the minimum key
// . if nextpg == getNumPages() then it returns the LAST KEY
// contained in the corresponding RdbFile
//if ( minpg != -1 && maps[fn]->getKey ( nextpg ) > minKey )
if (minpg != -1 &&
KEYCMP(maps[fn]->getKeyPtr(nextpg),minKey,m_ks)>0)continue;
// . we got a winner, his next page has the current min key
// . if m_endpg[i]+1 == getNumPages() then getKey() returns the
// last key in the mapped file
// . minKey should never equal the key on m_endpg[i] UNLESS
// it's on page #m_numPages
//minKey = maps[fn]->getKey ( nextpg );
KEYSET(minKey,maps[fn]->getKeyPtr(nextpg),m_ks);
minpg = i;
// if minKey is same as the current key on this endpg, inc it
// so we cause some advancement, otherwise, we'll loop forever
//if ( minKey != maps[fn]->getKey ( m_endpg[i] ) ) continue;
if ( KEYCMP(minKey,maps[fn]->getKeyPtr(m_endpg[i]),m_ks)!=0)
continue;
//minKey += (uint32_t) 1;
KEYADD(minKey,m_ks);
}
// . we're done if we hit the end of all maps in the race
// . return the max end key
// key_t maxEndKey; maxEndKey.setMax(); return maxEndKey; }
// . no, just the endKey
//if ( minpg == -1 ) return endKey;
if ( minpg == -1 ) return;
// sanity check
if ( lastMinKeyIsValid && KEYCMP(minKey,lastMinKey,m_ks)<=0 ) {
g_errno = ECORRUPTDATA;
log("db: Got corrupted map in memory for %s. This is almost "
"always because of bad memory. Please replace your RAM.",
base->m_dbname);
// do not wait for any merge to complete... otherwise
// Rdb.cpp will not close until the merge is done
g_merge.m_isMerging = false;
g_merge2.m_isMerging = false;
// to complete
// shutdown with urgent=true so threads are disabled.
g_process.shutdown(true);
//g_numCorrupt++;
// sleep for now until we make sure this works
//sleep(2000);
return;
}
// don't let minKey exceed endKey, however
//if ( minKey > endKey ) {
if ( KEYCMP(minKey,endKey,m_ks)>0 ) {
//minKey = endKey ;
//minKey += (uint32_t) 1;
//lastMinKey = endKey;
KEYSET(minKey,endKey,m_ks);
KEYADD(minKey,m_ks);
KEYSET(lastMinKey,endKey,m_ks);
}
else {
//lastMinKey = minKey ;
//lastMinKey -= (uint32_t) 1;
KEYSET(lastMinKey,minKey,m_ks);
KEYSUB(lastMinKey,m_ks);
}
// it is now valid
lastMinKeyIsValid = 1;
// . advance m_endpg[i] so that next page < minKey
// . we want to read UP TO the first key on m_endpg[i]
for ( int32_t i = 0 ; i < m_numFileNums ; i++ ) {
int32_t fn = fileNums[i];
m_endpg[i] = maps[fn]->getEndPage ( m_endpg[i], lastMinKey );
}
// . if the minKey is BIGGER than the provided endKey we're done
// . we don't necessarily include records whose key is "minKey"
//if ( minKey > endKey ) return endKey;
if ( KEYCMP(minKey,endKey,m_ks)>0) return;
// . calculate recSizes per page within [startKey,minKey-1]
// . compute bytes of records in [startKey,minKey-1] for each map
// . this includes negative records so we may have annihilations
// when merging into "diskList" and get less than what we wanted
// but endKey should be shortened, so our caller will know to call
// again if he wants more
int32_t recSizes = 0;
for ( int32_t i = 0 ; i < m_numFileNums ; i++ ) {
int32_t fn = fileNums[i];
recSizes += maps[fn]->getMinRecSizes ( m_startpg[i] ,
m_endpg [i] ,
startKey ,
lastMinKey ,
false );
}
// if we hit it then return minKey -1 so we only read UP TO "minKey"
// not including "minKey"
//if ( recSizes >= minRecSizes )
if ( recSizes >= minRecSizes ) {
// . sanity check
// . this sanity check fails sometimes, but leave it
// out for now... causes the Illegal endkey msgs in
// RdbList::indexMerge_r()
//if ( KEYNEG(lastMinKey) ) { char *xx=NULL;*xx=0; }
KEYSET(endKey,lastMinKey,m_ks);
//return lastMinKey;
return;
}
// keep on truckin'
goto loop;
}
// . 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;
}
// . return false if blocked, true otherwise
// . sets g_errno on error
// . this one is also called by RdbMerge to dump lists
bool RdbDump::dumpList ( RdbList *list , int32_t niceness , bool recall ) {
// if we had a write error and are being recalled...
if ( recall ) { m_offset -= m_bytesToWrite; goto recallskip; }
// assume we don't hack the list
m_hacked = false;
m_hacked12 = false;
// save ptr to list... why?
m_list = list;
// nothing to do if list is empty
if ( m_list->isEmpty() ) return true;
// we're now in dump mode again
m_isDumping = true;
//#ifdef GBSANITYCHECK
// don't check list if we're dumping an unordered list from tree!
if ( g_conf.m_verifyWrites && m_orderedDump ) {
m_list->checkList_r ( false /*removedNegRecs?*/ );
// print list stats
// log("dump: sk=%s ",KEYSTR(m_list->m_startKey,m_ks));
// log("dump: ek=%s ",KEYSTR(m_list->m_endKey,m_ks));
}
//#endif
// before calling RdbMap::addList(), always reset list ptr
// since we no longer call this in RdbMap::addList() so we don't
// mess up the possible HACK below
m_list->resetListPtr();
// . SANITY CHECK
// . ensure first key is >= last key added to the map map
if ( m_offset > 0 && m_map ) {
//key_t k = m_list->getCurrentKey();
char k[MAX_KEY_BYTES];
m_list->getCurrentKey(k);
//key_t lastKey = m_map->getLastKey (); // m_lastKey
char lastKey[MAX_KEY_BYTES];
m_map->getLastKey(lastKey);
//char *lastKey = m_map->getLastKey();
//if ( k <= lastKey ) {
if ( KEYCMP(k,lastKey,m_ks)<=0 ) {
log(LOG_LOGIC,"db: Dumping list key out of order. "
//"lastKey.n1=%"XINT32" n0=%"XINT64" k.n1=%"XINT32" n0=%"XINT64"",
//lastKey.n1,lastKey.n0,k.n1,k.n0);
"lastKey=%s k=%s",
KEYSTR(lastKey,m_ks),
KEYSTR(k,m_ks));
g_errno = EBADENGINEER;
//return true;
char *xx = NULL; *xx = 0;
}
}
if ( g_conf.m_verifyWrites ) {
char rdbId = 0;
if ( m_rdb ) rdbId = m_rdb->m_rdbId;
m_list->checkList_r(false,false,rdbId);//RDB_POSDB);
m_list->resetListPtr();
}
// HACK! POSDB
if ( m_ks == 18 && m_orderedDump && m_offset > 0 ) {
char k[MAX_KEY_BYTES];
m_list->getCurrentKey(k);
// . same top 6 bytes as last key we added?
// . if so, we should only add 6 bytes from this key, not 12
// so on disk it is compressed consistently
if ( memcmp ( (k ) + (m_ks-12) ,
(m_prevLastKey ) + (m_ks-12) , 12 ) == 0 ) {
char tmp[MAX_KEY_BYTES];
char *p = m_list->getList();
// swap high 12 bytes with low 6 bytes for first key
gbmemcpy ( tmp , p , m_ks-12 );
gbmemcpy ( p , p + (m_ks-12) , 12 );
gbmemcpy ( p + 12, tmp , m_ks-12 );
// big hack here
m_list->m_list = p + 12;
m_list->m_listPtr = p + 12;
m_list->m_listPtrLo = p ;
m_list->m_listPtrHi = p + 6;
m_list->m_listSize -= 12 ;
// turn on both bits to indicate double compression
*(p+12) |= 0x06;
m_hacked12 = true;
}
}
// . HACK
// . if we're doing an ordered dump then hack the list's first 12 byte
// key to make it a 6 byte iff the last key we dumped last time
// shares the same top 6 bytes as the first key of this list
// . this way we maintain compression consistency on the disk
// so IndexTable.cpp can expect all 6 byte keys for the same termid
// and RdbList::checkList_r() can expect the half bits to always be
// on when they can be on
// . IMPORTANT: calling m_list->resetListPtr() will mess this HACK up!!
if ( m_useHalfKeys && m_orderedDump && m_offset > 0 && ! m_hacked12 ) {
//key_t k = m_list->getCurrentKey();
char k[MAX_KEY_BYTES];
m_list->getCurrentKey(k);
// . same top 6 bytes as last key we added?
// . if so, we should only add 6 bytes from this key, not 12
// so on disk it is compressed consistently
//if ( memcmp ( ((char *)&k ) + 6 ,
// ((char *)&m_prevLastKey ) + 6 , 6 ) == 0 ) {
if ( memcmp ( (k ) + (m_ks-6) ,
(m_prevLastKey ) + (m_ks-6) , 6 ) == 0 ) {
m_hacked = true;
//char tmp[6];
char tmp[MAX_KEY_BYTES];
char *p = m_list->getList();
//gbmemcpy ( tmp , p , 6 );
//gbmemcpy ( p , p + 6 , 6 );
//gbmemcpy ( p + 6 , tmp , 6 );
gbmemcpy ( tmp , p , m_ks-6 );
gbmemcpy ( p , p + (m_ks-6) , 6 );
gbmemcpy ( p + 6 , tmp , m_ks-6 );
// big hack here
m_list->m_list = p + 6;
m_list->m_listPtr = p + 6;
// make this work for POSDB, too
m_list->m_listPtrLo = p + 6 + 6;
m_list->m_listPtrHi = p ;
m_list->m_listSize -= 6 ;
// hack on the half bit, too
*(p+6) |= 0x02;
}
}
// update old last key
//m_prevLastKey = m_list->getLastKey();
m_list->getLastKey(m_prevLastKey);
// now write it to disk
m_buf = m_list->getList ();
m_bytesToWrite = m_list->getListSize();
//#ifdef GBSANITYCHECK
//if (m_list->getListSize()!=m_list->getListEnd() - m_list->getList()){
// log("RdbDump::dumpList: major problem here!");
// sleep(50000);
//}
//#endif
recallskip:
// make sure we have enough mem to add to map after a successful
// dump up here, otherwise, if we write it and fail to add to map
// the map is not in sync if we core thereafter
if ( m_addToMap && m_map && ! m_map->prealloc ( m_list ) ) {
log("db: Failed to prealloc list into map: %s.",
mstrerror(g_errno));
// g_errno should be set to something if that failed
if ( ! g_errno ) { char *xx = NULL; *xx = 0; }
return true;
}
// tab to the old offset
int64_t offset = m_offset;
// might as well update the offset now, even before write is done
m_offset += m_bytesToWrite ;
// write thread is out
m_writing = true;
//m_bytesWritten = 0;
// sanity check
//log("dump: writing %"INT32" bytes at offset %"INT64"",m_bytesToWrite,offset);
// . if we're called by RdbMerge directly use m_callback/m_state
// . otherwise, use doneWritingWrapper() which will call dumpTree()
// . BigFile::write() return 0 if blocked,-1 on error,>0 on completion
// . it also sets g_errno on error
bool isDone = m_file->write ( m_buf ,
m_bytesToWrite ,
offset ,
&m_fstate ,
this ,
doneWritingWrapper ,
niceness );
// debug msg
//log("RdbDump dumped %"INT32" bytes, done=%"INT32"\n",
// m_bytesToWrite,isDone);
// return false if it blocked
if ( ! isDone ) return false;
// done writing
m_writing = false;
// return true on error
if ( g_errno ) return true;
// . delete list from tree, incorporate list into cache, add to map
// . returns false if blocked, true otherwise, sets g_errno on error
// . will only block in calling updateTfndb()
return doneDumpingList ( true );
}
// . returns false if blocked, true otherwise
// . sets g_errno on error
// . dumps the RdbTree, m_tree, into m_file
// . also sets and writes the RdbMap for m_file
// . we methodically get RdbLists from the RdbTree
// . dumped recs are ordered by key if "orderedDump" was true in call to set()
// otherwise, lists are ordered by node #
// . we write each list of recs to the file until the whole tree has been done
// . we delete all records in list from the tree after we've written the list
// . if a cache was provided we incorporate the list into the cache before
// deleting it from the tree to keep the cache in sync. NO we do NOT!
// . called again by writeBuf() when it's done writing the whole list
bool RdbDump::dumpTree ( bool recall ) {
// set up some vars
//int32_t nextNode;
//key_t maxEndKey;
//maxEndKey.setMax();
char maxEndKey[MAX_KEY_BYTES];
KEYMAX(maxEndKey,m_ks);
// if dumping statsdb, we can only dump records 30 seconds old or
// more because Statsdb.cpp can "back modify" such records in the tree
// because it may have a query that took 10 seconds come in then it
// needs to add a partial stat to the last 10 stats for those 10 secs.
// we use Global time at this juncture
if ( m_rdb->m_rdbId == RDB_STATSDB ) {
int32_t nowSecs = getTimeGlobal();
StatKey *sk = (StatKey *)maxEndKey;
sk->m_zero = 0x01;
sk->m_labelHash = 0xffffffff;
// leave last 60 seconds in there just to be safe
sk->m_time1 = nowSecs - 60;
}
// this list will hold the list of nodes/recs from m_tree
m_list = &m_ourList;
// convert coll to collnum
//collnum_t collnum = g_collectiondb.getCollnum ( m_coll );
// a collnum of -1 is for collectionless rdbs
//if ( collnum < 0 ) {
// //if ( g_catdb->getRdb() == m_rdb )
// if ( ! m_rdb->m_isCollectionLess ) {
// char *xx=NULL;*xx=0; //return true;
// }
// g_errno = 0;
// collnum = 0;
//}
// getMemOccupiedForList2() can take some time, so breathe
int32_t niceness = 1;
loop:
// if the lastKey was the max end key last time then we're done
if ( m_rolledOver ) return true;
// this is set to -1 when we're done with our unordered dump
if ( m_nextNode == -1 ) return true;
// . NOTE: list's buffer space should be re-used!! (TODO)
// . "lastNode" is set to the last node # in the list
bool status = true;
//if ( ! m_orderedDump ) {
// status = ((RdbTree *)m_tree)->getListUnordered ( m_nextNode ,
// m_maxBufSize ,
// m_list ,
// &nextNode );
// // this is -1 when no more nodes are left
// m_nextNode = nextNode;
//}
// "lastKey" is set to the last key in the list
//else {
{
// can we remove neg recs?
// class RdbBase *base = m_rdb->getBase(m_collnum);
// bool removeNegRecs = false;
// if ( base->m_numFiles <= 0 ) removeNegRecs = true;
if ( recall ) goto skip;
// debug msg
//log("RdbDump:: getting list");
m_t1 = gettimeofdayInMilliseconds();
if(m_tree)
status = m_tree->getList ( m_collnum ,
m_nextKey ,
maxEndKey ,
m_maxBufSize , // max recSizes
m_list ,
&m_numPosRecs ,
&m_numNegRecs ,
m_useHalfKeys ,
niceness );
else if(m_buckets)
status = m_buckets->getList ( m_collnum,
m_nextKey ,
maxEndKey ,
m_maxBufSize , // max recSizes
m_list ,
&m_numPosRecs ,
&m_numNegRecs ,
m_useHalfKeys );
// don't dump out any neg recs if it is our first time dumping
// to a file for this rdb/coll. TODO: implement this later.
//if ( removeNegRecs )
// m_list.removeNegRecs();
// if(!m_list->checkList_r ( false , // removeNegRecs?
// false , // sleep on problem?
// m_rdb->m_rdbId )) {
// log("db: list to dump is not sane!");
// char *xx=NULL;*xx=0;
// }
skip:
int64_t t2;
//key_t lastKey;
char *lastKey;
// if error getting list (out of memory?)
if ( ! status ) goto hadError;
// debug msg
t2 = gettimeofdayInMilliseconds();
log(LOG_INFO,"db: Get list took %"INT64" ms. "
"%"INT32" positive. %"INT32" negative.",
t2 - m_t1 , m_numPosRecs , m_numNegRecs );
// keep a total count for reporting when done
m_totalPosDumped += m_numPosRecs;
m_totalNegDumped += m_numNegRecs;
// . check the list we got from the tree for problems
// . ensures keys are ordered from lowest to highest as well
//#ifdef GBSANITYCHECK
if ( g_conf.m_verifyWrites ) {
char *s = "none";
if ( m_rdb ) s = getDbnameFromId(m_rdb->m_rdbId);
log("dump: verifying list before dumping (rdb=%s)",s);
m_list->checkList_r ( false , // removeNegRecs?
false , // sleep on problem?
m_rdb->m_rdbId );
}
// if list is empty, we're done!
if ( status && m_list->isEmpty() ) {
// consider that a rollover?
if ( m_rdb->m_rdbId == RDB_STATSDB )
m_rolledOver = true;
return true;
}
// get the last key of the list
lastKey = m_list->getLastKey();
// advance m_nextKey
//m_nextKey = lastKey ;
//m_nextKey += (uint32_t)1;
//if ( m_nextKey < lastKey ) m_rolledOver = true;
KEYSET(m_nextKey,lastKey,m_ks);
KEYADD(m_nextKey,1,m_ks);
if (KEYCMP(m_nextKey,lastKey,m_ks)<0) m_rolledOver = true;
// debug msg
//log(0,"RdbDump:lastKey.n1=%"UINT32",n0=%"UINT64"",lastKey.n1,lastKey.n0);
//log(0,"RdbDump:next.n1=%"UINT32",n0=%"UINT64"",m_nextKey.n1,m_nextKey.n0);
}
// . return true on error, g_errno should have been set
// . this is probably out of memory error
if ( ! status ) {
hadError:
log("db: Had error getting data for dump: %s. Retrying.",
mstrerror(g_errno));
// debug msg
//log("RdbDump::getList: sleeping and retrying");
// retry for the remaining two types of errors
if (!g_loop.registerSleepCallback(1000,this,tryAgainWrapper2)){
log(
"db: Retry failed. Could not register callback.");
return true;
}
// wait for sleep
return false;
}
// if list is empty, we're done!
if ( m_list->isEmpty() ) return true;
// . set m_firstKeyInQueue and m_lastKeyInQueue
// . this doesn't work if you're doing an unordered dump, but we should
// not allow adds when closing
m_lastKeyInQueue = m_list->getLastKey();
//m_firstKeyInQueue = m_list->getCurrentKey();
m_list->getCurrentKey(m_firstKeyInQueue);
// . write this list to disk
// . returns false if blocked, true otherwise
// . sets g_errno on error
// . if this blocks it should call us (dumpTree() back)
if ( ! dumpList ( m_list , m_niceness , false ) ) return false;
// close up shop on a write/dumpList error
if ( g_errno ) return true;
// . if dumpList() did not block then keep on truckin'
// . otherwise, wait for callback of dumpTree()
goto loop;
}
int main()
{
atddtree_key min = 1;
atddtree_key max;// = 1000;
atddtree_key keys[10] = {2,3,5,8,16,23,26,35,48,50};
int i;
atddtree* t;
int nuser = 1024*1024-1; //1024*1024*64-1;
double mean_ia = 205;
gsl_rng *r;
const gsl_rng_type *T;
int n=5;
double u;
T=gsl_rng_ranlxs0; //设随机数生成器类型是 ranlxs0
//gen arrival
gsl_rng_default_seed = ((unsigned long)(time(NULL))); //设seed值为当前时间
r=gsl_rng_alloc(T); //生成实例
double* exp_sample_ir = MALLOC(nuser, double);
double abstemp = 0;
for(i=0;i<nuser;i++) {
exp_sample_ir[i] = gsl_ran_exponential(r, mean_ia);
//exp_sample_ir[i] = 2+(i%10000)*0.3;
#ifdef LOGISTIC
abstemp = gsl_ran_logistic(r, 1);
if(abstemp<0) {
abstemp=0-abstemp;
}
exp_sample_ir[i] = abstemp;
#endif
//exp_sample_ir[i] = 5*gsl_ran_beta(r, 5, 1);
//exp_sample_ir[i] = 5*gsl_ran_lognormal(r, 5, 0.25);
//printf("exp: %f\n", exp_sample_ir[i]);
}
double* arrival_real = MALLOC(nuser, double);
arrival_real[0] = 1.0;
for(i=1;i<nuser;i++) {
arrival_real[i] = arrival_real[i-1]+exp_sample_ir[i-1];
//printf("arrival_real: %f\n", arrival_real[i]);
}
atddtree_key* arrival = MALLOC(nuser, atddtree_key);
for(i=0;i<nuser;i++) {
arrival[i] = (atddtree_key)arrival_real[i];
//printf("arrival: %ld\n", arrival[i]);
}
max = 0;
for(i=0;i<nuser;i++) {
if(KEYCMP(arrival[i],max)>0) {
KEYCPY(max,arrival[i]);
}
}
printf("---max=%ld\n", max);
t = atddtree_create(&min, &max);
for(i=0;i<nuser;i++) {
atddtree_insert(t, arrival+i);
//printf("insert %ld, height=%d\n", arrival[i], t->h);
}
printf("height=%d\n", t->h);
}