# See the file LICENSE for redistribution information. # # Copyright (c) 1999, 2012 Oracle and/or its affiliates. All rights reserved. # # $Id$ # proc build_all_subdb { dbname methods psize dups {nentries 100} {dbargs ""}} { set nsubdbs [llength $dups] set mlen [llength $methods] set savearg $dbargs for {set i 0} {$i < $nsubdbs} { incr i } { set m [lindex $methods [expr $i % $mlen]] set dbargs $savearg subdb_build $dbname $nentries [lindex $dups $i] \ $i $m $psize sub$i.db $dbargs } } proc subdb_build { name nkeys ndups dup_interval method psize subdb dbargs} { source ./include.tcl set dbargs [convert_args $method $dbargs] set omethod [convert_method $method] puts "Method: $method" set txnenv 0 set eindex [lsearch -exact $dbargs "-env"] if { $eindex != -1 } { incr eindex set env [lindex $dbargs $eindex] set txnenv [is_txnenv $env] } # Create the database and open the dictionary set oflags "-create -mode 0644 $omethod \ -pagesize $psize $dbargs {$name} $subdb" set db [eval {berkdb_open} $oflags] error_check_good dbopen [is_valid_db $db] TRUE set did [open $dict] set count 0 if { $ndups >= 0 } { puts "\tBuilding $method {$name} $subdb. \ $nkeys keys with $ndups duplicates at interval of $dup_interval" } if { $ndups < 0 } { puts "\tBuilding $method {$name} $subdb. \ $nkeys unique keys of pagesize $psize" # # If ndups is < 0, we want unique keys in each subdb, # so skip ahead in the dict by nkeys * iteration # for { set count 0 } \ { $count < [expr $nkeys * $dup_interval] } { incr count} { set ret [gets $did str] if { $ret == -1 } { break } } } set txn "" for { set count 0 } { [gets $did str] != -1 && $count < $nkeys } { incr count} { for { set i 0 } { $i < $ndups } { incr i } { set data [format "%04d" [expr $i * $dup_interval]] if { $txnenv == 1 } { set t [$env txn] error_check_good txn [is_valid_txn $t $env] TRUE set txn "-txn $t" } set ret [eval {$db put} $txn {$str \ [chop_data $method $data]}] error_check_good put $ret 0 if { $txnenv == 1 } { error_check_good txn [$t commit] 0 } } if { $txnenv == 1 } { set t [$env txn] error_check_good txn [is_valid_txn $t $env] TRUE set txn "-txn $t" } if { $ndups == 0 } { set ret [eval {$db put} $txn {$str \ [chop_data $method NODUP]}] error_check_good put $ret 0 } elseif { $ndups < 0 } { if { [is_record_based $method] == 1 } { global kvals set num [expr $nkeys * $dup_interval] set num [expr $num + $count + 1] set ret [eval {$db put} $txn {$num \ [chop_data $method $str]}] set kvals($num) [pad_data $method $str] error_check_good put $ret 0 } else { set ret [eval {$db put} $txn \ {$str [chop_data $method $str]}] error_check_good put $ret 0 } } if { $txnenv == 1 } { error_check_good txn [$t commit] 0 } } close $did error_check_good close:$name [$db close] 0 } proc do_join_subdb { db primary subdbs key oargs } { source ./include.tcl puts "\tJoining: $subdbs on $key" # Open all the databases set p [eval {berkdb_open -unknown} $oargs { $db } $primary] error_check_good "primary open" [is_valid_db $p] TRUE set dblist "" set curslist "" foreach i $subdbs { set jdb [eval {berkdb_open -unknown} $oargs { $db } sub$i.db] error_check_good "sub$i.db open" [is_valid_db $jdb] TRUE lappend jlist [list $jdb $key] lappend dblist $jdb } set join_res [eval {$p get_join} $jlist] set ndups [llength $join_res] # Calculate how many dups we expect. # We go through the list of indices. If we find a 0, then we # expect 0 dups. For everything else, we look at pairs of numbers, # if the are relatively prime, multiply them and figure out how # many times that goes into 50. If they aren't relatively prime, # take the number of times the larger goes into 50. set expected 50 set last 1 foreach n $subdbs { if { $n == 0 } { set expected 0 break } if { $last == $n } { continue } if { [expr $last % $n] == 0 || [expr $n % $last] == 0 } { if { $n > $last } { set last $n set expected [expr 50 / $last] } } else { set last [expr $n * $last / [gcd $n $last]] set expected [expr 50 / $last] } } error_check_good number_of_dups:$subdbs $ndups $expected # # If we get here, we have the number expected, now loop # through each and see if it is what we expected. # for { set i 0 } { $i < $ndups } { incr i } { set pair [lindex $join_res $i] set k [lindex $pair 0] foreach j $subdbs { error_check_bad valid_dup:$j:$subdbs $j 0 set kval [string trimleft $k 0] if { [string length $kval] == 0 } { set kval 0 } error_check_good \ valid_dup:$j:$subdbs [expr $kval % $j] 0 } } error_check_good close_primary [$p close] 0 foreach i $dblist { error_check_good close_index:$i [$i close] 0 } } proc n_to_subname { n } { if { $n == 0 } { return null.db; } else { return sub$n.db; } }