Index: Makefile
==================================================================
--- Makefile
+++ Makefile
@@ -24,10 +24,11 @@
 	fossil timeline -n 350 -t ci -F "%h,%a,%b,%t,\"%c\"" > recent-commits.csv
 
 
 SHELL=/bin/bash
 PREFIX=$(PWD)
+# CSCOPTS=-lfa2 -specialize -inline-global
 CSCOPTS=
 INSTALL=install
 SRCFILES = common.scm items.scm launch.scm ods.scm runconfig.scm	\
            server.scm configf.scm db.scm keys.scm		\
            process.scm runs.scm tasks.scm tests.scm genexample.scm	\
@@ -339,10 +340,13 @@
 
 $(PREFIX)/bin/mt-new-to-old.sh : utils/mt-new-to-old.sh
 	$(INSTALL) $< $@
 	chmod a+x $@
 
+$(PREFIX)/bin/convert-db.sh : utils/convert-db.sh
+	$(INSTALL) $< $@
+	chmod a+x $@
 
 deploytarg/nbfake : utils/nbfake
 	$(INSTALL) $< $@
 	chmod a+x $@
 
@@ -385,10 +389,11 @@
 
 install : $(PREFIX)/bin/.$(ARCHSTR) $(PREFIX)/bin/.$(ARCHSTR)/mtest $(PREFIX)/bin/megatest \
           $(PREFIX)/bin/.$(ARCHSTR)/dboard $(PREFIX)/bin/dashboard $(HELPERS) $(PREFIX)/bin/nbfake \
 	  $(PREFIX)/bin/nbfind $(PREFIX)/bin/mtrunner $(PREFIX)/bin/viewscreen $(PREFIX)/bin/mt_xterm \
           $(PREFIX)/bin/mt-old-to-new.sh $(PREFIX)/bin/mt-new-to-old.sh \
+          $(PREFIX)/bin/convert-db.sh $(PREFIX)/bin/convert-db.sh \
 	  $(PREFIX)/share/docs/megatest_manual.html $(PREFIX)/bin/remrun \
 	  $(PREFIX)/share/docs/megatest_manual.html $(PREFIX)/bin/remrun $(PREFIX)/bin/mtutil \
 	  $(PREFIX)/share/db/mt-pg.sql \
           $(PREFIX)/share/js/jquery-3.1.0.slim.min.js \
           $(PREFIX)/bin/.$(ARCHSTR)/lib/libpangox-1.0.so \

Index: api.scm
==================================================================
--- api.scm
+++ api.scm
@@ -155,12 +155,18 @@
 
 (define *db-write-mutexes* (make-hash-table))
 (define *server-signature* #f)
 
 (define *api-threads* '())
-(define (api:register-thread th-in)
-  (set! *api-threads* (cons (cons th-in (current-seconds)) *api-threads*)))
+(define (api:register-thread th-in command)
+  (set! *api-threads* (cons (list th-in (current-seconds) command) *api-threads*)))
+
+(define (api:get-thread-command th-in)
+  (let ((thread-data (assoc th-in *api-threads*)))
+    (if thread-data
+        (third thread-data) ; Assuming the command is the third element in the list
+        #f))) ; Return #f if the thread is not found
 
 (define (api:unregister-thread th-in)
   (set! *api-threads* (filter (lambda (thdat)
 				(not (eq? th-in (car thdat))))
 			      *api-threads*)))
@@ -170,10 +176,19 @@
 				(not (member (thread-state (car thdat)) '(terminated dead))))
 			      *api-threads*)))
 
 (define (api:get-count-threads-alive)
   (length *api-threads*))
+
+(define (api:get-threads)
+  (map (lambda (thdat)
+         (let ((thread (first thdat))
+               (timestamp (second thdat))
+               (command (third thdat)))
+           (format "\nThread: ~a, age: ~a, Command: ~a" thread (- (current-seconds) timestamp) command)))
+       *api-threads*))
+
 
 (define *api:last-stats-print* 0)
 (define *api-print-db-stats-mutex* (make-mutex))
 (define (api:print-db-stats)
   (debug:print-info 0 *default-log-port* "Started periodic db stats printer")
@@ -196,11 +211,11 @@
 (define (api:tcp-dispatch-request-make-handler dbstruct) ;; cmd run-id params)
   (assert *toppath* "FATAL: api:tcp-dispatch-request-make-handler called but *toppath* not set.")
   (if (not *server-signature*)
       (set! *server-signature* (tt:mk-signature *toppath*)))
   (lambda (indat)
-    (api:register-thread (current-thread))
+    (api:register-thread (current-thread) (car indat))
     (let* ((result 
 	    (let* ((numthreads (api:get-count-threads-alive))
 		   (delay-wait (if (> numthreads 10)
 				   (- numthreads 10)
 				   0))
@@ -228,17 +243,24 @@
 					  (assert ok "FATAL: database file and run-id not aligned.")))))
 			     (ttdat   *server-info*)
 			     (server-state (tt-state ttdat))
 			     (maxthreads   20) ;; make this a parameter?
 			     (status  (cond
-				       ((and (> numthreads maxthreads)
-					     (> (random 100) 70)) ;; allow a 30% probability to go through so we can figure out what is going wrong in main.db server.
-					'busy)
-				       ;; ((> numthreads 5) 'loaded) ;; this gets transmitted to the client which calls tt:backoff-incr to slow stuff down.
+                                       ((> numthreads maxthreads)
+					(let* ((testsuite (common:get-testsuite-name))
+					       (mtexe     (common:find-local-megatest))
+					       (proc      (lambda ()
+							    ;; we are overloaded, try to start another server
+							    (debug:print 0 *default-log-port* "Too many threads running, starting another server")
+							    (tt:server-process-run *toppath* testsuite mtexe run-id))))
+					  (set! *server-start-requests* (cons proc *server-start-requests*)))
+					;; 'busy
+					'loaded ;; not ideal since the client will not backoff
+					)
 				       (else 'ok)))
 			     (errmsg  (case status
-					((busy)   (conc "Server overloaded, "numthreads" threads in flight"))
+					((busy)   (conc "Server overloaded, "numthreads" threads in flight, current cmd: " cmd "\n current threads: " (api:get-threads)))
 					((loaded) (conc "Server loaded, "numthreads" threads in flight"))
 					(else     #f)))
 			     (result  (case status
 					((busy)
 					 (if (eq? cmd 'ping)
@@ -251,11 +273,11 @@
 					 ;; 				(thread-sleep! 0.5))
 					 (normal-proc cmd run-id params))
 					(else
 					 (normal-proc cmd run-id params))))
 			     (meta   (case cmd
-				       ((ping) `((sstate . ,server-state)))
+				       ((ping) `((sstate . ,server-state)(sload . ,numthreads)))
 				       (else   `((wait . ,delay-wait)))))
 			     (payload (list status errmsg result meta)))
 			;; (cmd run-id params meta)
 			(db:add-stats cmd run-id params (- (current-milliseconds) start-t))
 			payload))
@@ -264,12 +286,10 @@
       ;; (set! *api-process-request-count* (- *api-process-request-count* 1))
       ;; (serialize payload)
      
       (api:unregister-thread (current-thread))
       result)))
-
-
 
 (define *api-halt-writes* #f)
 
 (define (api:dispatch-request dbstruct cmd run-id params)
   (if (not *no-sync-db*)

Index: common.scm
==================================================================
--- common.scm
+++ common.scm
@@ -38,13 +38,15 @@
 
 (import commonmod
 	debugprint
 	rmtmod
 	(prefix mtargs args:))
-
+	
+(define (remove-server-files directory-path)
+  (let ((files (glob (string-append directory-path "/server*"))))
+    (for-each delete-file* files)))
 (include "common_records.scm")
-
 (define (remove-files filespec)
   (let ((files (glob filespec)))
     (for-each delete-file* files)))
 
 (define (stop-the-train)
@@ -53,10 +55,11 @@
 				  (if (and *toppath*
 					   (file-exists? (conc *toppath*"/stop-the-train")))
 				      (let* ((msg (conc "ERROR: found file "*toppath*"/stop-the-train, exiting immediately")))
 					;; yes, print to current-output-port AND *default-log-port*, annoying but necessary I think
 					(print msg)
+                                        (remove-server-files (conc *toppath* "/logs"))
 					(debug:print 0 *default-log-port* msg)
                                         (remove-files (conc *toppath* "/logs/server*"))
                                         (remove-files (conc *toppath* "/.servinfo/*"))
                                         (remove-files (conc *toppath* "/.mtdb/*lock"))
 					(exit 1)))
@@ -397,21 +400,19 @@
 (define (common:version-changed?)
   (not (equal? (common:get-last-run-version)
                (common:version-signature))))
 
 
-;; From 1.70 to 1.80, db's are compatible.
-
+;; From 1.70 to 1.81, db's are compatible.
+;;
+;; BUG: This logic is almost certainly not quite correct.
+;;
 (define (common:api-changed?)
-  (let* (
-    (megatest-major-version (substring (->string megatest-version) 0 4))
-    (run-major-version (substring (conc (common:get-last-run-version)) 0 4))
-   )
-   (and (not (equal? megatest-major-version "1.80"))
-     (not (equal? megatest-major-version megatest-run-version)))
-  )
-)
+  (let* ((megatest-major-version (substring (->string megatest-version) 0 4))
+	 (run-major-version (substring (conc (common:get-last-run-version)) 0 4)))
+    (and (not (member megatest-major-version '("1.81" "1.80")))
+	 (not (equal? megatest-major-version run-major-version)))))
 
 ;;======================================================================
 ;; Move me elsewhere ...
 ;; RADT => Why do we meed the version check here, this is called only if version misma
 ;;

Index: dashboard-tests.scm
==================================================================
--- dashboard-tests.scm
+++ dashboard-tests.scm
@@ -95,11 +95,12 @@
 			      "Current state: "
 			      "Current status: "
 			      "Test comment: "
 			      "Test id: "
 			      "Test date: "))
-		   (list (iup:label "" #:expand "VERTICAL"))))
+		   (list (iup:label "" #:expand "VERTICAL"
+				    ))))
     (apply iup:vbox  ; #:expand "YES"
 	   (list 
 	    (store-label "testname"
 			 (iup:label (db:test-get-testname  testdat) #:expand "HORIZONTAL")
 			 (lambda (testdat)(db:test-get-testname testdat)))
@@ -163,11 +164,12 @@
 			(list "Author: "
 			      "Owner: "
 			      "Reviewed: "
 			      "Tags: "
 			      "Description: "))
-		   (list (iup:label "" #:expand "VERTICAL"))))
+		   (list (iup:label "" #:expand "VERTICAL"
+				    ))))
     (apply iup:vbox  ; #:expand "YES"
 	   (list 
 	    (store-meta "author"
 			 (iup:label (db:testmeta-get-author testmeta) #:expand "HORIZONTAL")
 			 (lambda (testmeta)(db:testmeta-get-author testmeta)))
@@ -179,11 +181,12 @@
 			 (lambda (testmeta)(db:testmeta-get-reviewed testmeta)))
 	    (store-meta "tags" 
 			 (iup:label (db:testmeta-get-tags testmeta) #:expand "HORIZONTAL")
 			 (lambda (testmeta)(db:testmeta-get-tags testmeta)))
 	    (store-meta "description" 
-			 (iup:label (test-meta-panel-get-description testmeta) #:size "x50"); #:expand "HORIZONTAL")
+			(iup:label (test-meta-panel-get-description testmeta) ;; #:wordwrap "YES" ;; #:size "x50"
+				   ) ;; #:expand "HORIZONTAL")
 			 (lambda (testmeta)
 			   (test-meta-panel-get-description testmeta)))
 	    )))))
 
 
@@ -207,16 +210,21 @@
 		     (list (iup:label "runname ")
 			   (iup:label "run-id")
 			   (iup:label "run-date"))))
       (apply iup:vbox
 	     (append (map (lambda (keyval)
-			    (iup:label (cadr keyval) #:expand "HORIZONTAL"))
+			    (iup:vbox
+			     (iup:label (cadr keyval) #:expand "HORIZONTAL")
+			     ;; (iup:label "" #:expand "BOTH")
+			     )
+			    )
 			  keydat)
 		     (list (iup:label runname)
 			   (iup:label (conc run-id))
 			   (iup:label (seconds->year-work-week/day-time event_time))
-			   (iup:label "" #:expand "VERTICAL"))))))))
+			   (iup:label "" ;;#:expand "VERTICAL"
+				      ))))))))
   
 ;;======================================================================
 ;; Host info panel
 ;;======================================================================
 (define (host-info-panel testdat store-label)
@@ -232,11 +240,12 @@
 			      "CPU Load: "
 			      "Run duration: "
 			      "Logfile: "
 			      "Top process id: "
 			      "Uname -a: "))
-		   (iup:label "" #:expand "VERTICAL")))
+		   (iup:label "" ;; #:expand "VERTICAL"
+			      )))
     (apply iup:vbox ; #:expand "YES"
 	   (list
 	    ;; NOTE: Yes, the host can change!
 	    (store-label "HostName"
 			 (iup:label ;; (sdb:qry 'getstr 
@@ -268,18 +277,20 @@
 ;;
 (define (submegatest-panel dbstruct keydat testdat runname testconfig)
   (let* ((test-run-dir      (db:test-get-rundir testdat))
 	 (subarea           (subrun:get-runarea test-run-dir))
 	 (area-exists       (and subarea (common:file-exists? subarea silent: #t))))
-    (if subarea
-	(iup:frame 
-	 #:title "Megatest Run Info" ; #:expand "YES"
+    (iup:frame 
+     #:title "Megatest Run Info" ;; #:expand "HORIZONTAL"
+     (if subarea
 	 (iup:button
 	  "Launch Dashboard"
 	  #:action (lambda (obj)
-                     (subrun:launch-dashboard test-run-dir))))
-	(iup:vbox))))
+                     (subrun:launch-dashboard test-run-dir)))
+	 (iup:vbox
+	  (iup:label "Not a subrun..." #:expand "HORIZONTAL")
+	  )))))
 
 ;; use a global for setting the buttons colors
 ;;                           state status teststeps
 (define *state-status* (vector #f #f #f))
 (define (update-state-status-buttons testdat)
@@ -671,10 +682,19 @@
 			     (conc "megatest -target " keystring " -runname " runname 
 				   " -run -testpatt " (conc testname "/" (if (equal? item-path "")
 									"%" 
 									item-path))
 				   " -clean-cache"
+				   ))))
+	       (rerun-clean  (lambda (x)
+			    (iup:attribute-set! 
+			     command-text-box "VALUE"
+			     (conc "megatest -target " keystring " -runname " runname 
+				   " -rerun-clean -testpatt " (conc testname "/" (if (equal? item-path "")
+									"%" 
+									item-path))
+				   " -clean-cache"
 				   ))))
 	       (remove-test (lambda (x)
 			      (iup:attribute-set!
 			       command-text-box "VALUE"
 			       (conc "megatest -remove-runs -target " keystring " -runname " runname
@@ -719,31 +739,45 @@
 	   (else
 	    ;;  (test-set-status! db run-id test-name state status itemdat)
 	    (set! self ; 
 		  (iup:dialog #:close_cb (lambda (a)(exit)) ; #:expand "YES"
 			      #:title testfullname
-			      (iup:vbox ; #:expand "YES"
-			       ;; The run and test info
-			       (iup:hbox  ; #:expand "YES"
-				(run-info-panel dbstruct keydat testdat runname)
-				(test-info-panel testdat store-label widgets)
-				(test-meta-panel testmeta store-meta))
+			      (iup:vbox
 			       (iup:hbox
-				(host-info-panel testdat store-label)
-				(submegatest-panel dbstruct keydat testdat runname testconfig))
+				(iup:vbox ; #:expand "YES"
+				 ;; The run and test info
+				 (iup:hbox  ; #:expand "YES"
+				  (run-info-panel dbstruct keydat testdat runname)
+				  (test-info-panel testdat store-label widgets))
+				 (host-info-panel testdat store-label))
+				(iup:vbox
+				 (test-meta-panel testmeta store-meta)
+				 (submegatest-panel dbstruct keydat testdat runname testconfig)))
 			       ;; The controls
-			       (iup:frame #:title "Actions" 
+			       (iup:hbox ;; frame #:title "Actions" 
 					  (iup:vbox
-					   (iup:hbox 
-					    (iup:button "View Log"      #:action viewlog      #:size "80x")
-					    (iup:button "Start Xterm"   #:action xterm        #:size "80x")
-					    (iup:button "Run Test"      #:action run-test     #:size "80x")
-					    (iup:button "Clean Test"    #:action remove-test  #:size "80x")
-					    (iup:button "CleanRunExecute!"    #:action clean-run-execute #:size "80x")
-					    (iup:button "Kill All Jobs" #:action kill-jobs    #:size "80x")
-					    (iup:button "Archive Test"  #:action archive-test #:size "80x")
-					    (iup:button "Close"         #:action (lambda (x)(exit)) #:size "80x"))
+					   (iup:hbox
+					    (iup:frame
+					     #:title "Immediate"
+					     (iup:hbox
+					      (iup:button "Start Xterm"   #:action xterm        #:size "80x")
+					      (iup:button "CleanRunExecute!"    #:action clean-run-execute #:size "80x")
+					      (iup:button "View Log"      #:action viewlog      #:size "80x")))
+					    (iup:frame
+					     #:title "Command line"
+					     (iup:hbox
+					      (iup:button "Run Test"      #:action run-test     #:size "80x")
+					      (iup:button "Rerun-clean"   #:action rerun-clean  #:size "80x")
+					      (iup:button "Clean Test"    #:action remove-test  #:size "80x")
+					      (iup:button "Kill All Jobs" #:action kill-jobs    #:size "80x")))
+					    (iup:label "" #:expand "HORIZONTAL")
+					    (iup:frame
+					     #:title "Other"
+					     (iup:hbox
+					      ;; (iup:button "Archive Test"  #:action archive-test #:size "80x")
+					      (iup:button "Close"         #:action (lambda (x)(exit)) #:size "80x")
+					      )))
 					   (apply 
 					    iup:hbox
 					    (list command-text-box command-launch-button))))
 			       (set-fields-panel dbstruct run-id test-id testdat)
 			       (let ((tabs 

Index: docs/Makefile
==================================================================
--- docs/Makefile
+++ docs/Makefile
@@ -33,5 +33,8 @@
 megatest.pdf : megatest.lyx
 	lyx -e pdf2 megatest.lyx
 
 pkts.pdf : pkts.dot
 	dot -Tpdf pkts.dot -o pkts.pdf
+
+stepwise.pdf : stepwise-rpc-via-direct-and-tcp-or-http.dot
+	dot stepwise-rpc-via-direct-and-tcp-or-http.dot -Tpdf -o stepwise.pdf

ADDED   docs/stepwise-rpc-via-direct-and-tcp-or-http.dot
Index: docs/stepwise-rpc-via-direct-and-tcp-or-http.dot
==================================================================
--- /dev/null
+++ docs/stepwise-rpc-via-direct-and-tcp-or-http.dot
@@ -0,0 +1,98 @@
+//  Copyright 2006-2017, Matthew Welland.
+// 
+// This file is part of Megatest.
+// 
+//     Megatest is free software: you can redistribute it and/or modify
+//     it under the terms of the GNU General Public License as published by
+//     the Free Software Foundation, either version 3 of the License, or
+//     (at your option) any later version.
+// 
+//     Megatest is distributed in the hope that it will be useful,
+//     but WITHOUT ANY WARRANTY; without even the implied warranty of
+//     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+//     GNU General Public License for more details.
+// 
+//     You should have received a copy of the GNU General Public License
+//     along with Megatest.  If not, see <http://www.gnu.org/licenses/>.
+//
+
+digraph megatest_state_status {
+  ranksep=0.05;
+  // rankdir=LR
+    
+  node [shape=box,style=filled];
+  
+  "START" -> "have_server";
+  "DONE" [label="ALL DONE"];
+  "send_request" -> "receive_response";
+  "receive_response" -> "DONE";
+  "have_server" -> "start_http_server" [label="no"];
+  
+  subgraph cluster_start_server {
+    label="Start Server";
+    "start_http_server" -> "create_servinfo_file";
+    "start_http_server" -> "enable_direct_mode";
+    "create_servinfo_file" -> "delay_60_sec";
+    "delay_60_sec" -> "set_up_tmp_cache";
+    "set_up_tmp_cache" -> "switch_mode_to_tmp";
+  }
+  
+  subgraph cluster_direct_access {
+    label="Direct Access";
+   
+    "direct_access" -> "touch_access_file";
+    "touch_access_file" -> "touch_host_pid_file";
+    "touch_host_pid_file" -> "count_host_pid_files";
+    "count_host_pid_files" -> "call_query_proc" [label="count < 5"];
+    "wait_for_low_count_host_pid_files" [label="sleep 1"];
+    "count_host_pid_files" -> "wait_for_low_count_host_pid_files"[label="count >= 5"];
+    "wait_for_low_count_host_pid_files" -> "count_host_pid_files";
+    "call_query_proc" -> "DONE";
+  }
+
+  "have_server" -> "ping_server" [label="yes"];
+  "have_server" -> "direct_access" [label="no" ];
+  "ping_server" -> "send_request" [label="server alive"];
+  "ping_server" -> "remove_server_file" [label="server not alive"];
+  "remove_server_file" -> "START";
+
+
+}
+// subgraph cluster_notstarted {
+//   label="Not started";
+
+//   "NOT_STARTED FAILS" [
+//   label = "{ NOT_STARTED/FAILS |{ NO_ITEMS |<here> FAIL_PREREQ |<here> FAIL_TIMEOUT }}";
+//   shape= "record";
+//   ]
+// 
+// "NOT_STARTED n/a" -> "LAUNCHED n/a" [label=" launch"];
+// "NOT_STARTED WAIT" -> "LAUNCHED n/a"
+// 
+//   "NOT_STARTED n/a";
+//   "NOT_STARTED WAIT" [
+//   label = "{NOT_STARTED WAIT|{ NO_SLOTS | <here> WAIT_PREREQ}}";
+//   shape = "record";
+// ]
+// 
+// // struct3 [shape=record,label="hello\nworld |{ b |{c|<here> d|e}| f}| g | h"];
+// 
+//   "NOT_STARTED n/a" -> "NOT_STARTED FAILS";
+//   "NOT_STARTED n/a" -> "NOT_STARTED WAIT";
+// 
+//   "RUNNING" [
+//      shape="record";
+//      label="{RUNNING|{n/a|<here> PASS |<here> FAIL}}";
+//   ]
+// 
+//   "COMPLETED" [
+//       shape="record";
+//       label = "{COMPLETED|{PASS | SKIP | WAIVED | FAIL | CHECK| ABORT}}";
+//   ]
+// 
+// 
+// "RUNNING" -> "COMPLETED";
+// "RUNNING" -> "INCOMPLETE" [label="test dead for > 24hrs"];
+// 
+// 
+// "LAUNCHED n/a" -> "REMOTEHOSTSTART n/a" -> "RUNNING";

Index: launch.scm
==================================================================
--- launch.scm
+++ launch.scm
@@ -380,10 +380,55 @@
 		   ;; (change-file-mode name (bitwise-ior perm/irwxg perm/irwxu))
 		   )))
 	      (else
 	       (debug:print-info 0 "Invalid logpro definiton found in [logpro] section of testconfig. \"" logprodat "\""))))
      logpros)))
+
+;; read testconfig and create .logpro and script files
+;;    - use #f for tconfigreg to re-read the testconfigs from disk
+;;
+(define (launch:extract-scripts-logpro  test-dir test-name item-path tconfigreg-in)
+  (let* ((tconfigreg      (or tconfigreg-in
+			      (tests:get-all)))
+	 (tconfig-fname   (conc test-dir "/.testconfig"))
+	 (tconfig-tmpfile (conc tconfig-fname ".tmp"))
+	 (tconfig         (tests:get-testconfig test-name item-path tconfigreg #t force-create: #t)) ;; 'return-procs)))
+	 (scripts         (configf:get-section tconfig "scripts"))
+	 (logpros         (configf:get-section tconfig "logpro")))
+    ;; create .testconfig file
+    (configf:write-alist tconfig tconfig-tmpfile)
+    (file-move tconfig-tmpfile tconfig-fname #t)
+    (delete-file* ".final-status")
+    
+    ;; extract scripts from testconfig and write them to files in test run dir
+    (for-each
+     (lambda (scriptdat)
+       (match scriptdat
+	      ((name content)
+	       (debug:print-info 2 *default-log-port* "Creating script "(current-directory)"/"name)
+	       (with-output-to-file name
+		 (lambda ()
+		   (print content)))
+	       (change-file-mode name (bitwise-ior perm/irwxg perm/irwxu)))
+	      (else
+	       (debug:print-info 0 "Invalid script definiton found in [scripts] section of testconfig. \"" scriptdat "\""))))
+     scripts)
+
+    ;; extract logpro from testconfig and write them to files in test run dir
+    (for-each
+     (lambda (logprodat)
+       (match logprodat
+	      ((name content)
+	       (debug:print-info 2 *default-log-port* "Creating logpro file "(current-directory)"/"name)
+	       (with-output-to-file name
+		 (lambda ()
+		   (print content)
+		   ;; (change-file-mode name (bitwise-ior perm/irwxg perm/irwxu))
+		   )))
+	      (else
+	       (debug:print-info 0 "Invalid logpro definiton found in [logpro] section of testconfig. \"" logprodat "\""))))
+     logpros)))
 
 (define (launch:execute encoded-cmd)
   (let* ((cmdinfo    (common:read-encoded-string encoded-cmd))
 	 (tconfigreg #f))
     (setenv "MT_CMDINFO" encoded-cmd)
@@ -733,16 +778,19 @@
 		 (th1          (make-thread monitorjob "monitor job"))
 		 (th2          (make-thread runit "run job"))
                  (tconfig         (tests:get-testconfig test-name item-path tconfigreg #t))
                  (propagate-exit-code (configf:lookup *configdat* "setup" "propagate-exit-code"))
                  (propagate-status-list '("FAIL" "KILLED" "ABORT" "DEAD" "CHECK" "SKIP" "WAIVED"))
-                 (test-status "not set")
+                 (test-status     "not set")
+		 (test-state      "not set")
 		 (precmd          (configf:lookup tconfig "setup" "precmd"))
 		 (postcmd         (configf:lookup tconfig "setup" "postcmd")))
 	    ;; first, if set, run the precmd
 	    (if precmd ;; (file-exists? precmd)(file-execute-access? precmd))
-		(system precmd)) ;; up to test author to put nbfake if desired.
+		(begin
+		  ;; (save-environment-as-files "precmd-envt")
+		  (system precmd))) ;; up to test author to put nbfake if desired.
 	    (set! job-thread th2)
 	    (thread-start! th1)
 	    (thread-start! th2)
 	    (thread-join! th2)
 	    (debug:print-info 0 *default-log-port* "Megatest execute of test " test-name ", item path " item-path " complete. Notifying the db ...")
@@ -801,25 +849,30 @@
 	    (mutex-unlock! m)
             (launch:end-of-run-check run-id )
 	    (debug:print 2 *default-log-port* "Output from running " fullrunscript ", pid " (launch:einf-pid exit-info) " in work area " 
 			 work-area ":\n====\n exit code " (launch:einf-exit-code exit-info) "\n" "====\n")
 
-
-            (set! test-status (db:test-get-status (rmt:get-testinfo-state-status run-id test-id)))
+	    (let* ((testrec  (rmt:get-testinfo-state-status run-id test-id)))
+              (set! test-status (db:test-get-status testrec))
+	      (set! test-state  (db:test-get-state  testrec)))
 
             ;; If the propagate-exit-code option has been set in the megatest config, and the test status matches the list, set the exit code to 1.
 
-	    (if postcmd
-		(system postcmd))
 
             (if (and propagate-exit-code (string=? propagate-exit-code "yes") (member test-status propagate-status-list))
                (begin
                 (debug:print 1 *default-log-port* "Setting exit status to 1 because of test status of " test-status) 
                 (set! *globalexitstatus* 1)
                )
             )
 
+	    (if postcmd
+		(begin
+		  (setenv "MT_TEST_STATE" test-state)
+		  (setenv "MT_TEST_STATUS" test-status)
+		  ;; (save-environment-as-files "postcmd-envt")
+		  (system postcmd)))
 	    (if (not (launch:einf-exit-status exit-info))
 		(exit 4))))
         )))
 
 ;; Spec for End of test

Index: megatest-version.scm
==================================================================
--- megatest-version.scm
+++ megatest-version.scm
@@ -18,6 +18,6 @@
 ;; Always use two or four digit decimal
 ;; 1.01, 1.02...1.10,1.11,1.1101 ... 1.99,2.00..
 
 ;; (declare (unit megatest-version))
 
-(define megatest-version 1.8031)
+(define megatest-version 1.8101)

Index: megatest.scm
==================================================================
--- megatest.scm
+++ megatest.scm
@@ -113,11 +113,13 @@
 ;;
 (let ((debugcontrolf (conc (get-environment-variable "HOME") "/.megatestrc")))
   (if (common:file-exists? debugcontrolf)
     (begin
       ;; for some reason, debug:print does not work here. Had to use print.
-      (print (conc "WARNING: loading " debugcontrolf))
+      (with-output-to-port (current-error-port)
+	(lambda ()
+	  (print (conc "WARNING: loading " debugcontrolf))))
       (load debugcontrolf)
     )
   )
 )
 
@@ -263,11 +265,10 @@
   -debug N|N,M,O...       : enable debug 0-N or N and M and O ...
   -debug-noprop N|M,M,O...: enable debug but do not propagate to subprocesses via MT_DEBUG
   -config fname           : override the megatest.config file with fname
   -append-config fname    : append fname to the megatest.config file
   -import-sexpr fname     : import a sexpr file (use -list-runs % -dumpmode sexpr to create)
-  -remove-dbs all         : remove Megatest DBs before importing sexpr. (Use only with -import-sexpr)
   -regen-testfiles        : regenerate scripts and logpro files from testconfig, run in test context
   
 Utilities
   -env2file fname         : write the environment to fname.csh and fname.sh
   -envcap a               : save current variables labeled as context 'a' in file envdat.db
@@ -2130,10 +2131,25 @@
 		  (paths    (tests:test-get-paths-matching keys target (args:get-arg "-test-files"))))
 	     (for-each (lambda (path)
 			 (print path))
 		       paths))))))
 
+;;======================================================================
+;; Utils for test areas
+;;======================================================================
+
+(if (args:get-arg "-regen-testfiles")
+    (if (getenv "MT_TEST_RUN_DIR")
+	(begin
+	  (launch:setup)
+	  (change-directory (getenv "MT_TEST_RUN_DIR"))
+	  (let* ((testname (getenv "MT_TEST_NAME"))
+		 (itempath (getenv "MT_ITEMPATH")))
+	    (launch:extract-scripts-logpro (getenv "MT_TEST_RUN_DIR") testname itempath #f))
+	  (set! *didsomething* #t))
+	(debug:print 0 *default-log-port* "ERROR: Must run -regen-testfiles in a test environment (i.e. test xterm from dashboard)")))
+		 	  
 ;;======================================================================
 ;; Utils for test areas
 ;;======================================================================
 
 (if (args:get-arg "-regen-testfiles")

Index: portlogger.scm
==================================================================
--- portlogger.scm
+++ portlogger.scm
@@ -15,10 +15,11 @@
 ;; 
 ;;     You should have received a copy of the GNU General Public License
 ;;     along with Megatest.  If not, see <http://www.gnu.org/licenses/>.
 ;;
 
+(require-extension (srfi 18) extras tcp s11n)
 
 (declare (unit portlogger))
 (declare (uses debugprint))
 (declare (uses dbmod))
 

Index: rmt.scm
==================================================================
--- rmt.scm
+++ rmt.scm
@@ -790,5 +790,8 @@
   (let* ((cfg-deadtime             (configf:lookup-number *configdat* "setup" "deadtime"))
 	 (test-stats-update-period (configf:lookup-number *configdat* "setup" "test-stats-update-period")))
    (rmt:find-and-mark-incomplete-engine run-id ovr-deadtime cfg-deadtime test-stats-update-period)
    ;;call end of eud of run detection for posthook
    (launch:end-of-run-check run-id)))
+
+;; orphaned from cherrypick merge
+;;         (debug:print 0 *default-log-port* "Inserting " (length tests-data) " tests in run " runname)

Index: rmtmod.scm
==================================================================
--- rmtmod.scm
+++ rmtmod.scm
@@ -18,12 +18,17 @@
 
 ;;======================================================================
 
 (declare (unit rmtmod))
 (declare (uses debugprint))
+;; (declare (uses debugprint.import))
 (declare (uses commonmod))
+;; (declare (uses commonmod.import))
 (declare (uses dbfile))    ;; needed for records
+(declare (uses dbmod))
+;; (declare (uses tcp-transportmod))
+;; (declare (uses tcp-transportmod.import))
 
 ;; (declare (uses apimod))
 ;; (declare (uses apimod.import))
 ;; (declare (uses ulex))
 
@@ -33,10 +38,14 @@
 	*
 	
 (import scheme chicken data-structures extras matchable srfi-69)
 (import (prefix sqlite3 sqlite3:) posix typed-records srfi-18)
 (import commonmod dbfile debugprint) ;; (prefix commonmod cmod:))
+(import dbmod
+	;; tcp-transportmod
+	)
+
 ;; (import apimod)
 ;; (import (prefix ulex ulex:))
 
 (include "db_records.scm")
 
@@ -305,7 +314,25 @@
 	     ;; call end of eud of run detection for posthook - from merge, is it needed?
 	     ;; (launch:end-of-run-check run-id)
 	     all-ids)
 	    )))))
 
+;;======================================================================
+;; Misc
+;;======================================================================
+
+;; (define (rmtmod:wait-on-server-load run-id ttdat)
+;;   (let* ((dbfname                 (dbmod:run-id->dbfname run-id))
+;; 	 (get-lowest-thread-load
+;; 	  (lambda ()
+;; 	    (let* ((sdats (tt:get-server-info-sorted ttdat dbfname)))
+;; 	      (car (map tt:get-server-threads sdats))))))
+;;     (if ttdat
+;; 	(let loop ()
+;; 	  (if (> (get-lowest-thread-load) 5) ;; load is pretty high
+;; 	      (begin
+;; 		(debug:print 0 *default-log-port* "Servers appear overloaded, waiting...")
+;; 		(thread-sleep! 1)
+;; 		(loop))))
+;; 	(debug:print 0 *default-log-port* "Can't wait on server load, *ttdat* not set"))))
 
 )

Index: runs.scm
==================================================================
--- runs.scm
+++ runs.scm
@@ -29,10 +29,11 @@
 (declare (uses mt))
 (declare (uses archive))
 (declare (uses mtargs))
 (declare (uses rmtmod))
 (declare (uses dbfile))
+(declare (uses tcp-transportmod))
 
 (use (prefix sqlite3 sqlite3:) srfi-1 posix regex regex-case srfi-69 (srfi 18) 
      posix-extras directory-utils pathname-expand typed-records format  sxml-serializer
      sxml-modifications matchable)
 
@@ -48,10 +49,11 @@
 
 (import commonmod
 	debugprint
 	rmtmod
 	dbfile
+	tcp-transportmod
 	(prefix mtargs args:))
 
 ;; use this struct to facilitate refactoring
 ;;
 
@@ -811,31 +813,21 @@
 	(debug:print-info 1 *default-log-port* "Adding \"" (string-intersperse required-tests " ") "\" to the run queue"))
     ;; NOTE: these are all parent tests, items are not expanded yet.
     (debug:print-info 4 *default-log-port* "test-records=" (hash-table->alist test-records))
     (let ((reglen (configf:lookup *configdat* "setup" "runqueue")))
       (if (> (length (hash-table-keys test-records)) 0)
-	  (let* ((keep-going        #t)
-		 (run-queue-retries 5)
-		 (run-ids (rmt:get-all-run-ids)))
-	    #;(for-each (lambda (run-id)
-			(if keep-going
-			    (handle-exceptions
-				exn
-			      (debug:print 0 *default-log-port* "error in calling find-and-mark-incomplete for run-id " run-id ", exn=" exn)
-			      (rmt:find-and-mark-incomplete run-id #f)))) ;; ovr-deadtime))) ;; could be root of https://hsdes.intel.com/appstore/article/#/220546828/main -- Title: Megatest jobs show DEAD even though they are still running (1.64/27)
-		      run-ids)
+	  (let* ()
 	    (runs:run-tests-queue run-id runname test-records keyvals flags test-patts required-tests
 				  (any->number reglen) all-tests-registry)
-	    (set! keep-going #f)
 	    (if (> run-count 0) ;; handle reruns
 		(begin
 		  (if (not (hash-table-ref/default flags "-preclean" #f))
 		      (hash-table-set! flags "-preclean" #t))
 		  (if (not (hash-table-ref/default flags "-rerun" #f))
 		      (hash-table-set! flags "-rerun" "ABORT,STUCK/DEAD,n/a,ZERO_ITEMS"))
-      (runs:run-tests target runname test-patts user flags run-count: (- run-count 1)))
-                (launch:end-of-run-check run-id)))
+                  (runs:run-tests target runname test-patts user flags run-count: (- run-count 1)))
+                 (launch:end-of-run-check run-id)))
 	  (debug:print-info 0 *default-log-port* "No tests to run")))
     (debug:print-info 4 *default-log-port* "All done by here")
     ;; TODO: try putting post hook call here
       
     ;  (debug:print-info 2 *default-log-port* " run-count " run-count)
@@ -1201,13 +1193,17 @@
 	       (common:wait-for-normalized-load maxload "Waiting for load to drop before starting more tests" #f))
 	   
 	   ;; jobtools maxhomehostload is intended to prevent overloading on the homehost which can cause database corruption issues
 	   (if maxhomehostload
 	       (common:wait-for-homehost-load maxhomehostload
-					      (conc "Waiting for homehost load to drop below normalized value of " maxhomehostload))))))
+					      (conc "Waiting for homehost load to drop below normalized value of " maxhomehostload)))
+
+	   ;; lastly lets check the servers are not overloaded by looking at threads
+	   (tt:wait-on-server-load run-id *ttdat*)
+	   
+	   )))
     
- 
     
     (if (and (not (null? prereqs-not-met))
 	     (runs:lownoise (conc "waiting on tests " prereqs-not-met hed) 60))
 	(debug:print-info 2 *default-log-port* "waiting on tests; " (string-intersperse (runs:mixed-list-testname-and-testrec->list-of-strings prereqs-not-met) ", ")))
 
@@ -1381,12 +1377,16 @@
 			      #f ;; yes, really
 			      (list (car tal)(cdr tal) reg reruns))
 			  (begin
 			    (if (runs:lownoise (conc "FAILED prerequisites or other issue" hed) 60)
 				(debug:print 0 *default-log-port* "WARNING: test " hed " has FAILED prerequisites or other issue. Internal state >" nth-try "< will be overridden and we'll retry."))
-			    ;; was: (mt:test-set-state-status-by-testname run-id test-name item-path "NOT_STARTED" "KEEP_TRYING" #f)
-                            (mt:test-set-state-status-by-testname-unless-completed run-id test-name item-path "COMPLETED" "PREQ_FAIL" #f)
+			    (let* ((test-id      (rmt:get-test-id run-id testname item-path))
+				   (test-info    (rmt:get-testinfo-state-status run-id test-id)) ;; we need *current* info
+				   (status       (db:test-status test-info)))
+			      (if (equal? status "KEEP_TRYING")
+				  (mt:test-set-state-status-by-testname-unless-completed run-id test-name item-path "COMPLETED" "PREQ_FAIL" #f)
+				  (mt:test-set-state-status-by-testname run-id test-name item-path "NOT_STARTED" "KEEP_TRYING" #f)))
 			    (hash-table-set! test-registry hed 'removed) ;; was 0
                             (if (not (and (null? reg) (null? tal)))
                                 (runs:loop-values tal reg reglen regfull reruns)
                                 #f))))
 		     (else

Index: server.scm
==================================================================
--- server.scm
+++ server.scm
@@ -393,11 +393,11 @@
 		      (set! new-server-key (with-input-from-file start-flag (lambda () (read-line))))
 		      (equal? server-key new-server-key)))
 	       #t
                ;; If either of the above conditions is not true, print a "Gating server start" message, wait <idle-time> + 1, then call this function recursively. 
 	       (begin
-		 (debug:print-info 0 *default-log-port* "Gating server start, last start: "
+		 (debug:print-info 2 *default-log-port* "Gating server start, last start: "
 				   (seconds->time-string fmodtime) ", time since last start: " delta ", required idletime: " idletime ", gating reason:" (if old-enough "another job started a server" "too soon to start another server"))
 		 
 		 (thread-sleep! ( + 1 idletime))
 		 (server:wait-for-server-start-last-flag areapath)))))))
 

Index: tcp-transportmod.scm
==================================================================
--- tcp-transportmod.scm
+++ tcp-transportmod.scm
@@ -159,14 +159,40 @@
 (define (tt:get-conn ttdat dbfname)
   (hash-table-ref/default (tt-conns ttdat) dbfname #f))
 
 ;; do all the busy work of finding and setting up conn for
 ;; connecting to a server
+;; This function, `tt:client-connect-to-server`, is designed to manage connections between a client and a server within a testing framework.
+;; The function takes four arguments:
+;; 1. `ttdat`: a data structure that holds information about the testing environment or connections.
+;; 2. `dbfname`: The name of the database file that the client wants to connect to.
+;; 3. `run-id`: An identifier for the current run of the test suite.
+;; 4. `testsuite`: The test suite that is being run.
+;;
+;; Here's a step-by-step explanation of what the function does:
+;;
+;; 1. It first asserts that the `run-id` is valid for the given `dbfname` using the `tt:valid-run-id` function. If the `run-id` is not valid, it raises a fatal error.
+;; 2. It prints debug information indicating that the function `tt:client-connect-to-server` has been called with the given `dbfname`.
+;; 3. It attempts to retrieve an existing connection to the server from a hash table (`tt-conns`) using the `dbfname` as the key. If a connection already exists, it prints debug information and returns the existing connection.
+;; 4. If no existing connection is found, it retrieves the current server information from the servinfo file, using the `tt:get-current-server-info` function.
+;; 5. It uses pattern matching to destructure the server information into variables (`host`, `port`, `start-time`, `server-id`, `pid`, `dbfname2`, `servinffile`). It then asserts that the `dbfname` from the server info matches the one provided to the function.
+;; 6. It constructs a connection object (`conn`) with the server information.
+;; 7. It attempts to ping the server using `tt:timed-ping` to verify that the server is running and can be communicated with.
+;; 8. Depending on the result of the ping:
+;;    - If the server is running (`running`), it prints debug information, saves the connection in the hash table, and returns the connection.
+;;    - If the server is starting (`starting`), it sleeps for 2 seconds and then recursively calls itself to retry the connection.
+;;    - If the server is neither running nor starting, it checks if it's been more than 10 seconds since the last server start attempt. If so, it attempts to start the server using `server-start-proc` and then sleeps for 1 second before retrying the connection.
+;; 9. If no server information is found (`else` case), it checks if it's been more than 3 seconds since the last server start attempt. If so, it starts a new server using `server-start-proc`, updates the last server start time, and sleeps for 4 seconds.
+;; 10. It then sleeps for 1 second and prints debug information before recursively calling itself to retry the connection.
+;;
+;; The function uses recursion to keep trying to connect to the server, with various sleep intervals to prevent overwhelming the system with connection attempts or server starts.
+;; It also uses a hash table to cache connections and avoid reconnecting to a server if a connection already exists.
+;; The function is designed to handle different server states and ensure that a server is running and available before returning a valid connection to the caller.
 ;; 
 (define (tt:client-connect-to-server ttdat dbfname run-id testsuite server-start-proc)
   (assert (tt:valid-run-id run-id dbfname) "FATAL: invalid run-id "run-id)
-  (debug:print-info 2 *default-log-port* "tt:client-connect-to-server " dbfname " " run-id)
+  (debug:print-info 2 *default-log-port* "tt:client-connect-to-server " dbfname)
   (let* ((conn              (tt:get-conn ttdat dbfname))
 	 (server-start-proc (or server-start-proc
 				(lambda ()
 				  (assert (equal? dbfname "main.db") ;; only main.db is started here
 					  "FATAL: called server-start-proc for db other than main.db")
@@ -175,20 +201,31 @@
 				   testsuite ;; (dbfile:testsuite-name)
 				   (common:find-local-megatest)
 				   run-id)))))
     (if conn
 	(begin 
-          (debug:print-info 2 *default-log-port* "already connected to a server")
+          (debug:print-info 2 *default-log-port* "already connected to a server for " dbfname)
            conn) ;; we are already connected to the server
 
 	;; no conn
+
+	;; find server with lowest number of threads running (i.e. lowest load)
+	;;
         (let* ((sdats (tt:get-server-info-sorted ttdat dbfname))
 	       (sdat  (if (null? sdats)
 			  #f
-			  (car sdats))))
-	   (debug:print-info 2 *default-log-port* "found sdat " sdat)
-           (match sdat
+			  ;; choose server with lowest threads count
+			  (car (sort sdats
+				     (lambda (a b)
+				       (let* ((load-a (tt:get-server-threads a))
+					      (load-b (tt:get-server-threads b)))
+					 (< load-a load-b))))))))
+				     
+	  ;; (let ((indx (max (random (- (length sdats) 1)) 0)))
+	  ;;    (list-ref sdats indx)))))
+	  ;; (debug:print-info 1 *default-log-port* "found sdat " sdat" from sdats: "sdats)
+          (match sdat
 	    ((host port start-time server-id pid dbfname2 servinffile)
 	     (assert (equal? dbfname dbfname2) "FATAL: read server info from wrong file.")
              (debug:print-info 2 *default-log-port* "no conn - in match servinffile:" servinffile)
 	     (let* ((host-port (conc host":"port))
 		    (conn (make-tt-conn
@@ -248,11 +285,57 @@
 ;; returns ( result . ping_time )
 (define (tt:timed-ping host port server-id)
   (let* ((start-time (current-milliseconds))
 	 (result     (tt:ping host port server-id)))
     (cons result (- (current-milliseconds) start-time))))
-    
+
+;; host:port => ( meta . when-updated)
+(define *server-load* (make-hash-table))
+
+(define (tt:save-server-meta host port meta)
+  (hash-table-set! *server-load* (conc host":"port) (cons meta (current-seconds))))
+
+(define (tt:get-server-threads dat)
+  (let* ((host (car  dat))
+	 (port (cadr dat))
+	 (dat  (tt:get-server-meta host port #t)))
+    ;; (debug:print 0 *default-log-port* "host: "host" port: "port" dat: "dat)
+    (if (list? dat)
+	(or (alist-ref 'sload dat) 99998)
+	99999))) ;; absurd number means don't use this one
+
+;; lazy get, does not auto-refresh meta, this might be a problem
+;;
+(define (tt:get-server-meta host port #!optional (do-ping #f))
+  (let* ((get-meta (lambda ()
+		     (let* ((dat  (hash-table-ref/default *server-load* (conc host":"port) #f)))
+		       (if dat (car dat) #f))))
+	 (meta     (get-meta)))
+    (if (and (not meta)
+	     do-ping)
+	(begin
+	  (tt:timed-ping host port #f)
+	  (get-meta))
+	meta)))
+
+(define (tt:wait-on-server-load run-id ttdat)
+  (if ttdat ;; if no server yet just pass on through
+      (let* ((dbfname                 (dbmod:run-id->dbfname run-id))
+	     (get-lowest-thread-load
+	      (lambda ()
+		(let* ((sdats (tt:get-server-info-sorted ttdat dbfname)))
+		  (car (map tt:get-server-threads sdats))))))
+	(if ttdat
+	    (let loop ((count 0))
+	      (let* ((lowestload (get-lowest-thread-load)))
+		(if (> lowestload 5) ;; load is pretty high
+		    (begin
+		      (debug:print 0 *default-log-port* "Servers appear overloaded with "lowestload" threads, waiting...")
+		      (thread-sleep! 1)
+		      (if (< count 10)
+			  (loop (+ count 1)))))))
+	    (debug:print 0 *default-log-port* "Can't wait on server load, *ttdat* not set")))))
 
 (define (tt:ping host port server-id #!optional (tries-left 5))
   (let*  ((res      (tt:send-receive-direct host port `(ping #f #f #f) ping-mode: #t)) ;; please send me your server-id
 	  (try-again (lambda ()
 		       (if (> tries-left 0)
@@ -263,25 +346,27 @@
     ;;
     ;; need two threads, one a 5 second timer
     ;;
     (match res
       ((status errmsg result meta)
+       (tt:save-server-meta host port meta)
        (if (equal? result server-id)
 	   (let* ((server-state (alist-ref 'sstate meta)))
 	     ;; (debug:print 0 *default-log-port* "Ping to "host":"port" successful.")
 	     (or server-state 'unk)) ;; then we are good
 	   (begin
-	     (debug:print 0 *default-log-port* "WARNING: server-id does not match, expected: "server-id", got: "result)
+	     (if server-id
+		 (debug:print 0 *default-log-port* "WARNING: server-id does not match, expected: "server-id", got: "result))
 	     #f)))
       (else
        ;; (debug:print 0 *default-log-port* "res not in form (status errmsg result meta), got: "res)
        (try-again)))))
 
 ;; client side handler
 ;;
 ;;(tt:handler #<tt> get-keys #f () 2 #f "/home/matt/data/megatest/ext-tests" #f "main.db" "ext-tests" "/home/matt/data/megatest/bin/.22.04/../megatest")
-;;
+;;g
 (define (tt:handler ttdat cmd run-id params attemptnum readonly-mode dbfname testsuite mtexe server-start-proc)
   ;; connect-to-server will start a server if needed.
   (let* ((areapath (tt-areapath ttdat))
 	 (conn     (tt:client-connect-to-server ttdat dbfname run-id testsuite server-start-proc))) ;; looks up conn keyed by dbfname
     (if conn
@@ -300,15 +385,19 @@
 	     (case status
 	       ((busy) ;; result will be how long the server wants you to delay
 		(let* ((raw-dly  (if (number? result) result 0.1))
 		       (dly      (+ raw-dly (/ attemptnum 10)))) ;; (* raw-dly (/ attemptnum 2))))
 		  (debug:print 0 *default-log-port* "WARNING: server for "dbfname" is busy, cmd is "cmd", will try again in "dly" seconds. This is attempt "(- attemptnum 1))
+                  (debug:print 0 *default-log-port* errmsg)
 		  (thread-sleep! dly)
 		  (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe server-start-proc)))
 	       ((loaded)
 		(debug:print 0 *default-log-port* "WARNING: server for "dbfname" is loaded, slowing queries.")
 		(tt:backoff-incr (tt-conn-host conn)(tt-conn-port conn))
+
+		;; this would be a good place to force reconnection and connect to a different server
+		
 		result) ;; (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))
 	       (else
 		result)))
 	    (else ;; did not receive properly formated result
 	     (if (not res) ;; tt:send-receive telling us that communication failed
@@ -355,11 +444,13 @@
 	(begin
 	  (thread-sleep! 1) ;; no conn yet set up, give it a rest and try again
 	  (tt:handler ttdat cmd run-id params attemptnum readonly-mode dbfname testsuite mtexe server-start-proc)))))
 
 ;; gets server info and appends path to server file
-;; sorts by age, oldest first
+;; sorts by age, --oldest-- now newest first
+;;
+;; move the ping here?
 ;;
 ;; returns list of (host port startseconds server-id servinfofile)
 ;;
 (define (tt:get-server-info-sorted ttdat dbfname)
   (let* ((areapath (tt-areapath ttdat))
@@ -368,11 +459,11 @@
 	 (sorted   (sort sdats (lambda (a b)
 				 (let* ((starta (list-ref a 2))
 					(startb (list-ref b 2)))
 				   (if (eq? starta startb)
 				       (string>? (list-ref a 3)(list-ref b 3)) ;; if servers started at same time look at server-id
-				       (< starta startb))))))
+				       (> starta startb))))))
 	 (count    0))
     (for-each
      (lambda (rec)
        (if (or (> (length sorted) 1)
 	       (common:low-noise-print 120 "server info sorted"))
@@ -491,10 +582,12 @@
 ;; server
 ;;======================================================================
 
 (define (tt:sync-dbs ttdat)
   #f)
+
+(define *server-start-requests* '())
 
 ;; start the listener and start responding to requests
 ;;
 ;; NOTE: organise by dbfname, not run-id so we don't need
 ;;       to pull in more modules
@@ -546,32 +639,36 @@
 				     (keep-srv  (and good-ping same-host)))
 				(if keep-srv	
 				    (loop (cdr servrs)
 					  host
 					  (cons servdat result))
-				    (begin
-                                      ;; (debug:print-info 0 *default-log-port* "good-ping: " good-ping " same-host: " same-host "keep-srv: " keep-srv)
-				      (handle-exceptions
-				       exn
-				       (debug:print-info 0 *default-log-port* "Error removing server info file: "servinfofile", "
-							 (condition->list exn))
-				       (delete-file* servinfofile))
-				      (loop (cdr servrs) prime-host result)))))
+				    (let* ((modtime (file-modification-time servinfofile)))
+				      ;; if the .servinfo hasn't been touched in five min
+				      ;; we can be pretty sure the server is truly dead
+				      (if (> (- (current-seconds) modtime) 360)
+					  (handle-exceptions
+					   exn
+					   (debug:print-info 0 *default-log-port*
+							     "Error removing server info file: "servinfofile", "
+							     (condition->list exn))
+					   (delete-file* servinfofile))
+					  (loop (cdr servrs) prime-host result))))))
 			     (else
-			      ;; can't delete it as we don't have a filename. NOTE: Should really never get here.
+			      ;; can't delete it as we don't have a filename. NOTE: Should never get here.
 			      (debug:print-info 0 *default-log-port* "ERROR: bad servinfo record \""servdat"\"")
 			      (loop (cdr servrs) prime-host result)) ;; drop 
 			     )))))
 	       (home-host (if (null? good-srvrs)
 			      #f
 			      (caar good-srvrs))))
 	  ;; by here we have a trustworthy list of servers and we have removed the .servinfo file for any unresponsive servers
 	  ;; and the list is in good-srvrs
+	  ;;
 	  (cond
 	   ((not home-host) ;; no servers yet, go ahead and start
 	    (debug:print-info 0 *default-log-port* "No servers yet, starting on "(get-host-name)))
-	   ((> (length good-srvrs) 2) ;; don't need more, just exit
+	   ((> (length good-srvrs) 3) ;; don't need more, just exit
 	    (debug:print-info 0 *default-log-port* "Have "(length good-srvrs)", no need for more, exiting.")
 	    (exit))
 	   ((not (equal? home-host (get-host-name))) ;; there is a home-host and we are not on it
 	    (debug:print-info 0 *default-log-port* "Prime main server is on host "home-host", but we are on host "(get-host-name)", exiting.")
 	    (exit))
@@ -595,10 +692,12 @@
 		    (begin
 		      (debug:print 0 *default-log-port* "ERROR: (tt-port ttdat) no port set! Exiting.")
 		      (exit)))))
 	  
 	  ;; create a servinfo file start keep-running
+	  ;; On WSL there seems to be a race condition where the .servinfo file
+	  ;; is not created fast enough
           (debug:print 0 *default-log-port* "Creating servinfo file for " dbfname)
 	  (tt:create-server-registration-file ttdat dbfname)
 	  (procinf-status-set! *procinf* "running")
 	  (tt-state-set! ttdat 'running)
 	  (dbfile:with-no-sync-db
@@ -621,10 +720,12 @@
 	     (dbfile:insert-or-update-process nsdb *procinf*)))
 	  (debug:print 0 *default-log-port* "Exiting now.")
 	  (exit))))))
 
 (define (tt:keep-running ttdat dbfname dbstruct)
+
+  (thread-sleep! 1)
   
   ;; at this point the server is running and responding to calls, we just monitor
   ;; for db calls and exit if there are none.
 
   ;; if I am not in the first 3 servers, exit
@@ -637,18 +738,21 @@
 	     (my-index  (list-index (lambda (x)
 				      (equal? (list-ref x 6)
 					      (tt-servinf-file ttdat)))
 				    servers))
 	     (ok         (cond
+			  ((not my-index)
+			   (debug:print 0 *default-log-port* "WARNING: Apparently I don't exist.")
+			   #f) ;; keep trying or give up?
 			  ((not *server-run*)
 			   (debug:print 0 *default-log-port* "WARNING: received a stop server from client by remote request.")
 			   #f)
 			  ((null? servers)
 			   (debug:print 0 *default-log-port* "WARNING: no servinfo files found, this cannot be.")
 			   #f) ;; not ok
-			  ((> my-index 2)
-			   (debug:print 0 *default-log-port* "WARNING: there are more than two servers ahead of me, I'm not needed, exiting.")
+			  ((> my-index 3)
+			   (debug:print 0 *default-log-port* "WARNING: there are more than three servers ahead of me, I'm not needed, exiting.")
 			   #f) ;; not ok to not be in first three
 			  ((eq? (tt-state ttdat) 'running) #t) ;; we are good to keep going
 			  ((> (- (current-seconds) start-time) 30)
 			   (debug:print 0 *default-log-port* "WARNING: over 30 seconds and not yet in runnning mode. Exiting.")
 			   #f)
@@ -669,11 +773,19 @@
 		(set! (file-modification-time sinfo-file) (current-seconds))
 		((dbr:dbstruct-sync-proc dbstruct) last-update)
 		(dbr:dbstruct-last-update-set! dbstruct curr-secs))))
 	
 	(if (< (- (current-seconds) (tt-last-access ttdat)) (tt-server-timeout-param))
-	    (begin
+	    ;; process any requests to start a new server due to load on this one
+	    (let* ((requests *server-start-requests*))
+	      (set! *server-start-requests* '())
+	      (if (> (length requests) 0)
+		  (debug:print-info 0 *default-log-port* "Processing "(length requests)" server start requests"))
+	      (for-each (lambda (proc)
+			  (proc)
+			  (thread-sleep! 1))
+			requests)
 	      (thread-sleep! 5)
 	      (loop)))))
     (tt:shutdown-server ttdat)
     (debug:print 0 *default-log-port* "INFO: Server timed out, exiting from tt:keep-running.")))
 
@@ -701,16 +813,26 @@
     (assert (and host port) "FATAL: tt:create-server-registration-file called with no conn, dbfname="dbfname)
     (tt-servinf-file-set! ttdat servinf)
     (with-output-to-file servinf
       (lambda ()
 	(print "SERVER STARTED: "host":"port" AT "(current-seconds)" server-id: "serv-id" pid: "(current-process-id)" dbfname: "dbfname)))
-      serv-id))
+    (let loop ((count 0))
+      (if (not (file-exists? servinf))
+	  (begin
+	    (debug:print 0 *default-log-port* "WARNING: file "servinf" was created but it doesn't show up on disk! We'll try again.")
+	    (thread-sleep! 1)
+	    (if (> count 10)
+		(debug:print 0 *default-log-port* "WARNING: file "servinf" was not created.")
+		(loop (+ count 1))))))
+    serv-id))
 
 ;; find valid server
 ;; get servers listed, last part of name must match :<dbfname>
 ;; if more than one, wait one second and look again
-;; future: ping oldest, if alive remove other :<dbfname> files
+;; 
+;; NOTE: this only gets the servinfo data, no network activity here
+;;       i.e. no ping etc.
 ;;
 (define (tt:find-server areapath dbfname)
   (let* ((servdir  (tt:get-servinfo-dir areapath))
 	 (sfiles   (glob (conc servdir"/*:"dbfname)))
 	 (goodfiles '()))
@@ -775,10 +897,12 @@
 
 (define (tt:too-recent-server-start dbfname)
   (let* ((last-run-time (hash-table-ref/default *last-server-start* dbfname #f)))
     (and last-run-time
 	 (< (- (current-seconds) last-run-time) 5))))
+
+(define *last-server-start-request-time* 0)
     
 ;; Given an area path,  start a server process    ### NOTE ### > file 2>&1 
 ;; if the target-host is set 
 ;; try running on that host
 ;;   incidental: rotate logs in logs/ dir.
@@ -787,26 +911,28 @@
   (assert areapath  "FATAL: tt:server-process-run called without areapath defined.")
   (assert testsuite "FATAL: tt:server-process-run called without testsuite defined.")
   (assert mtexe     "FATAL: tt:server-process-run called without mtexe defined.")
   ;; mtest -server - -m testsuite:ext-tests -db 6.db
   (let* ((dbfname  (dbmod:run-id->dbfname run-id)))
-    (if (tt:too-recent-server-start dbfname)
+    (if (or (< (- (current-seconds) *last-server-start-request-time*) 5) ;; attempted start less than 5 sec ago
+	    (tt:too-recent-server-start dbfname))
 	#f
 	(let* ((load     (get-normalized-cpu-load))
 	       (srvrs    (tt:find-server areapath dbfname))
 	       (trying   (length srvrs))
 	       (nrun     (number-of-processes-running (conc "mtest.*server.*"testsuite".*"dbfname))))
+	  (set! *last-server-start-request-time* (current-seconds))
 	  (cond
-	   ((> load 2.0)
-	    (debug:print 0 *default-log-port* "Normalized load "load" on " (get-host-name) " is over the limit of 2.0. Not starting a server. Please reduce the load on "(get-host-name)" by killing some processes")
-	    (thread-sleep! 1)
+	   ((> load 4.0)
+	    (debug:print 0 *default-log-port* "Normalized load " load " over 4, (load: " (commonmod:get-cpu-load) " cores: " (get-current-host-cores) " exiting...")
+	    (thread-sleep! 1) ;; I'm not convinced that a delay here is helpful. -mrw-
 	    #f)
 	   ((> nrun 100)
 	    (debug:print 0 *default-log-port* nrun" servers running on " (get-host-name) ", not starting another.")
 	    (thread-sleep! 1)
 	    #f)
-	   ((> trying 2)
+	   ((> trying 3)
 	    (debug:print 0 *default-log-port* trying" servers registered in .servinfo dir. not starting another.")
 	    (thread-sleep! 1)
 	    #f)
 	   (else
 	    (if (not (file-exists? (conc areapath"/logs")))

ADDED   utils/convert-db.sh
Index: utils/convert-db.sh
==================================================================
--- /dev/null
+++ utils/convert-db.sh
@@ -0,0 +1,25 @@
+#!/bin/bash
+
+if [ -z "megatest.config" ]; then
+    echo "The file 'megatest.config' does not exist. This must be run in a megatest area."
+    exit 1
+fi
+if [ -d ".mtdb" ]; then
+    echo "The .mtdb directory already exists. Will not do the conversion"
+    exit 1
+fi
+if [ -d ".megatest" ]; then
+    echo "Found a .megatest directory. Will convert from megatest 1.70 to 1.71/1.80 format"
+    /p/foundry/env/pkgs/megatest/1.70/16/bin/megatest -list-runs % -dumpmode sexpr > data.sexpr
+else 
+    if [ -f "megatest.db" ]; then
+        echo "Found megatest.db. Will convert from megatest 1.65 to 1.71/1.80 format"
+        /p/foundry/env/pkgs/megatest/1.65/92/bin/megatest -list-runs % -dumpmode sexpr > data.sexpr
+    else
+        echo "Did not find .megatest or megatest.db. Cannot do the conversion"
+        exit 1
+    fi
+fi
+which megatest
+megatest -import-sexpr data.sexpr
+