Megatest

Artifact [a172975f82]
Login

Artifact a172975f82f14d89542bf8be35bd9c45e1abeb70:


;;======================================================================
;; Copyright 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/>.

;;======================================================================

;;======================================================================
;; Cpumod:
;;
;;   Put things here don't fit anywhere else
;;======================================================================

(declare (unit launchmod))
(declare (uses debugprint))
(declare (uses mtargs))
(declare (uses commonmod))
(declare (uses configfmod))
(declare (uses dbfile))
(declare (uses dbmod))
(declare (uses rmtmod))
(declare (uses servermod))
(declare (uses processmod))
(declare (uses pgdb))
(declare (uses mtmod))
(declare (uses megatestmod))
(declare (uses tasksmod))
(declare (uses subrunmod))
(declare (uses testsmod))
(declare (uses runsmod))
(declare (uses fsmod))

(use srfi-69)

(module launchmod
	*

(import scheme)
(cond-expand
 (chicken-4
  
  (import chicken
	  ports
	  data-structures
	  extras
	  files
	  matchable
	  pathname-expand
	  posix
	  posix-extras
	  regex
	  regex-case
	  sparse-vectors
	  
	  )
  (use srfi-69))
 (chicken-5
  (import (prefix sqlite3 sqlite3:)
	  chicken.base
	  chicken.condition
	  chicken.file
	  chicken.file.posix
	  chicken.io
	  chicken.pathname
	  chicken.port
	  chicken.process
	  chicken.process-context
	  chicken.process-context.posix
	  chicken.sort
	  chicken.string
	  chicken.time
	  chicken.time.posix
	  
	  matchable
	  md5
	  message-digest
	  pathname-expand
	  regex
	  regex-case
	  system-information

  )))

;; imports common to ck4 and ck5
(import srfi-1
	srfi-13
	srfi-18
	srfi-69
	typed-records
	(prefix base64 base64:)
	(prefix sqlite3 sqlite3:)
	md5
	message-digest
	z3
	csv
	directory-utils
	call-with-environment-variables
	
	debugprint
	commonmod
	configfmod
	(prefix mtargs args:)
	dbmod
	dbfile
	rmtmod
	servermod
	processmod
	pgdb
	mtmod
	megatestmod
	tasksmod
	subrunmod
	testsmod
	runsmod
	fsmod
	)

(include "common_records.scm")
(include "key_records.scm")
(include "db_records.scm")
(include "megatest-fossil-hash.scm")

;;======================================================================
;; ezsteps
;;======================================================================

;; ezsteps were going to be coded as
;; stepname[,predstep1,predstep2 ...] [{VAR1=first,second,third}] command to execute
;;   BUT
;; now are
;; stepname {VAR=first,second,third ...} command ...
;; where the {VAR=first,second,third ...} is optional.

;; given an exit code and whether or not logpro was used calculate OK/BAD
;; return #t if we are ok, #f otherwise
(define (steprun-good? logpro exitcode stepparms)
  (or (eq? exitcode 0)
      (and logpro  (member exitcode '( 2 4 6)))
      (let* ((params (alist-ref 'params stepparms)) ;; get the params section
	     (keep-going (if params
			     (alist-ref "keep-going" params equal?)
			     #f)))
	(debug:print 0 *default-log-port* "keep-going=" keep-going)
	(and keep-going (equal? (car keep-going) "yes")))))

;; if handed a string, process it, else look for MT_CMDINFO
(define (launch:get-cmdinfo-assoc-list #!key (encoded-cmd #f))
  (let ((enccmd (if encoded-cmd encoded-cmd (getenv "MT_CMDINFO"))))
    (if enccmd
	(common:read-encoded-string enccmd)
	'())))

;;                       0           1              2              3
(defstruct launch:einf (pid #t)(exit-status #t)(exit-code #t)(rollup-status 0))

;; return (conc status ": " comment) from the final section so that
;;   the comment can be set in the step record in launch.scm
;;
(define (launch:load-logpro-dat run-id test-id stepname)
  (let ((cname (conc stepname ".dat")))
    (if (common:file-exists? cname)
	(let* ((dat  (read-config cname #f #f))
	       (csvr (db:logpro-dat->csv dat stepname))
	       (csvt (let-values (((fmt-cell fmt-record fmt-csv) (make-format ",")))
		       (fmt-csv (map list->csv-record csvr))))
	       (status (configf:lookup dat "final" "exit-status"))
	       (msg     (configf:lookup dat "final" "message")))
          (if csvt  ;; this if blocked stack dump caused by .dat file from logpro being 0-byte.  fixed by upgrading logpro
              (rmt:csv->test-data run-id test-id csvt)
	      (debug:print 0 *default-log-port* "ERROR: no csvdat exists for run-id: " run-id " test-id: " test-id " stepname: " stepname ", check that logpro version is 1.15 or newer"))
	  ;;  (debug:print-info 13 *default-log-port* "Error: run-id/test-id/stepname="run-id"/"test-id"/"stepname" => bad csvr="csvr)
	  ;;  )
	  (cond
	   ((equal? status "PASS") "PASS") ;; skip the message part if status is pass
	   (status (conc (configf:lookup dat "final" "exit-status") ": " (if msg msg "no message")))
	   (else #f)))
	#f)))

(define (launch:manage-steps run-id test-id item-path fullrunscript ezsteps subrun test-name tconfigreg exit-info m)
  ;; (let-values
  ;;  (((pid exit-status exit-code)
  ;;    (run-n-wait fullrunscript)))
  ;; (tests:test-set-status! test-id "RUNNING" "n/a" #f #f)
  ;; Since we should have a clean slate at this time there is no need to do 
  ;; any of the other stuff that tests:test-set-status! does. Let's just 
  ;; force RUNNING/n/a

  ;; (thread-sleep! 0.3)
  ;; (tests:test-force-state-status! run-id test-id "RUNNING" "n/a")
  (rmt:set-state-status-and-roll-up-items run-id test-name item-path "RUNNING" #f #f) 
  ;; (thread-sleep! 0.3) ;; NFS slowness has caused grief here

  ;; if there is a runscript do it first
  (if fullrunscript
      (let ((pid (process-run fullrunscript)))
	(rmt:test-set-top-process-pid run-id test-id pid)
	(let loop ((i 0))
	  (let-values
	   (((pid-val exit-status exit-code) (process-wait pid #t)))
	   (mutex-lock! m)
	   (launch:einf-pid-set!           exit-info  pid)         ;; (vector-set! exit-info 0 pid)
	   (launch:einf-exit-status-set!   exit-info  exit-status) ;; (vector-set! exit-info 1 exit-status)
	   (launch:einf-exit-code-set!     exit-info  exit-code)   ;; (vector-set! exit-info 2 exit-code)
	   (launch:einf-rollup-status-set! exit-info  exit-code)   ;; (vector-set! exit-info 3 exit-code)  ;; rollup status
	   (mutex-unlock! m)
	   (if (eq? pid-val 0)
	       (begin
		 (thread-sleep! 2)
		 (loop (+ i 1)))
	       )))))
  ;; then, if runscript ran ok (or did not get called)
  ;; do all the ezsteps (if any)
  (if (or ezsteps subrun)
      (let* ((test-run-dir (tests:get-test-path-from-environment))
             (testconfig ;; (read-config (conc work-area "/testconfig") #f #t environ-patt: "pre-launch-env-vars")) ;; FIXME??? is allow-system ok here?
	      ;; NOTE: it is tempting to turn off force-create of testconfig but dynamic
	      ;;       ezstep names need a full re-eval here.
	      (tests:get-testconfig test-name item-path tconfigreg #t force-create: #t)) ;; 'return-procs)))
	     (ezstepslst (if (hash-table? testconfig)
			     (hash-table-ref/default testconfig "ezsteps" '())
			     #f)))
	(if testconfig
	    (hash-table-set! *testconfigs* test-name testconfig) ;; cached for lazy reads later ...
	    (begin
	      (launch:setup)
	      (debug:print 0 *default-log-port* "WARNING: no testconfig found for " test-name " in search path:\n  "
			   (string-intersperse (tests:get-tests-search-path *configdat*) "\n  "))))
	;; after all that, still no testconfig? Time to abort
	(if (not testconfig)
	    (begin
	      (debug:print-error 0 *default-log-port* "Failed to resolve megatest.config, runconfigs.config and testconfig issues. Giving up now")
	      (exit 1)))

	;; create a proc for the subrun if requested, save that proc in the ezsteps table as the last entry
	;; 1. get section [runarun]
	;; 2. unset MT_* vars
	;; 3. fix target
	;; 4. fix runname
	;; 5. fix testpatt or calculate it from contour
	;; 6. launch the run
	;; 7. roll up the run result and or roll up the logpro processed result
	(when (configf:lookup testconfig "subrun" "runwait") ;; we use runwait as the flag that a subrun is requested
            (subrun:initialize-toprun-test testconfig test-run-dir)
	    (let* ((mt-cmd (subrun:launch-cmd test-run-dir (configf:lookup testconfig "subrun" "runwait"))))
              (debug:print-info 0 *default-log-port* "Subrun command is \"" mt-cmd "\"")
              (set! ezsteps #t) ;; set the needed flag
	      (set! ezstepslst
                    (append (or ezstepslst '())
                            (list (list "subrun" (conc "{subrun=true} " mt-cmd)))))))

	;; process the ezsteps
	(if ezsteps
	    (let* ((all-steps-dat (make-hash-table))) ;; keep all the info around as stepname ==> alist; where  'params is the params list (add other stuff as needed)
	      (if (not (common:file-exists? ".ezsteps"))(create-directory ".ezsteps"))
	      ;; if ezsteps was defined then we are sure to have at least one step but check anyway
	      (if (not (> (length ezstepslst) 0))
		  (debug:print-error 0 *default-log-port* "ezsteps defined but ezstepslst is zero length")
                  (let ((all-step-names (map car ezstepslst))
                        (status-file (file-open "ezsteps.status" (+ open/append open/wronly open/creat)))
                        )
		   (setenv "MT_STEP_NAMES" (string-intersperse all-step-names " "))
		   (let loop ((ezstep (car ezstepslst))
			     (tal    (cdr ezstepslst))
			     (prevstep #f))
                    (debug:print-info 0 *default-log-port* "Processing ezstep \"" (string-intersperse ezstep " ") "\"")
		    ;; check exit-info (vector-ref exit-info 1)
		    (if (launch:einf-exit-status exit-info) ;; (vector-ref exit-info 1)
			(let* ((logpro-used (ezsteps:runstep ezstep run-id test-id exit-info m tal testconfig all-steps-dat))
			       (stepname    (car ezstep))
			       (stepparms   (hash-table-ref all-steps-dat stepname)))
			  (setenv "MT_STEP_NAME" stepname)
			  (pp (hash-table->alist all-steps-dat))
			  ;; if logpro-used read in the stepname.dat file
			  (if (and logpro-used (common:file-exists? (conc stepname ".dat")))
			      (launch:load-logpro-dat run-id test-id stepname))

                          (file-write status-file (conc stepname " " (launch:einf-exit-code exit-info) "\n"))

			  (if (steprun-good? logpro-used (launch:einf-exit-code exit-info) stepparms)
			      (if (not (null? tal))
				  (loop (car tal) (cdr tal) stepname))
			      (debug:print 0 *default-log-port* "WARNING: step " (car ezstep) " failed. Stopping")))
			(debug:print 0 *default-log-port* "WARNING: a prior step failed, stopping at " ezstep)))
                        (file-close status-file)
                        )

                        ))))))

(define (launch:monitor-job run-id test-id item-path fullrunscript ezsteps test-name tconfigreg exit-info m work-area runtlim misc-flags)
  (let* ((update-period (string->number (or (configf:lookup *configdat* "setup" "test-stats-update-period") "60")))
         (start-seconds (current-seconds))
	 (calc-minutes  (lambda ()
			  (inexact->exact 
			   (round 
			    (- 
			     (current-seconds) 
			     start-seconds)))))
	 (kill-tries 0))
    ;; (tests:set-full-meta-info #f test-id run-id (calc-minutes) work-area)
    ;; (tests:set-full-meta-info test-id run-id (calc-minutes) work-area)
    (tests:set-full-meta-info #f test-id run-id (calc-minutes) work-area 10)

    (let loop ((minutes   (calc-minutes))
	       (cpu-load  (alist-ref 'adj-core-load (common:get-normalized-cpu-load #f)))
	       (disk-free (get-df (current-directory)))
               (last-sync (current-seconds)))
      (let* ((over-time     (> (current-seconds) (+ last-sync update-period)))
             (new-cpu-load  (let* ((load  (alist-ref 'adj-core-load (common:get-normalized-cpu-load #f)))
                                   (delta (abs (- load cpu-load))))
                              (if (> delta 0.1) ;; don't bother updating with small changes
                                  load
                                  #f)))
             (new-disk-free (let* ((df    (if over-time ;; only get df every 30 seconds
                                              (get-df (current-directory))
                                              disk-free))
                                   (delta (abs (- df disk-free))))
                              (if (and (> df 0)
                                       (> (/ delta df) 0.1)) ;; (> delta 200) ;; ignore changes under 200 Meg
                                  df
                                  #f)))
             (do-sync       (or new-cpu-load new-disk-free over-time))

             (test-info   (rmt:get-test-state-status-by-id run-id test-id))
             (state       (car test-info));; (db:test-get-state test-info))
             (status      (cdr test-info));; (db:test-get-status test-info))
	     (killreq     (equal? state "KILLREQ"))
             (kill-reason  "no kill reason specified")
             (kill-job?    #f))
        ;; (common:telemetry-log "zombie" (conc "launch:monitor-job - decision time encountered at "(current-seconds)" with last-sync="last-sync" do-sync="do-sync" over-time="over-time" update-period="update-period))
        (cond
         (killreq
          (set! kill-reason "KILLING TEST since received kill request (KILLREQ)")
          (set! kill-job? #t))
         ((and runtlim (> (- (current-seconds) start-seconds) runtlim))
          (set! kill-reason (conc "KILLING TEST DUE TO TIME LIMIT EXCEEDED! Runtime=" (- (current-seconds) start-seconds) " seconds, limit=" runtlim))
          (set! kill-job? #t))
         ((equal? status "DEAD")
          (tests:update-central-meta-info run-id test-id new-cpu-load new-disk-free (calc-minutes) #f #f)
          (rmt:set-state-status-and-roll-up-items run-id test-id 'foo "RUNNING" "n/a" "was marked dead; really still running.")
          ;;(set! kill-reason "KILLING TEST because it was marked as DEAD by launch:handle-zombie-tests (might indicate really overloaded server or else overzealous setup.deadtime)") ;; MARK RUNNING
          (set! kill-job? #f)))

        (debug:print 4 *default-log-port* "cpu: " new-cpu-load " disk: " new-disk-free " last-sync: " last-sync " do-sync: " do-sync)
        (if (common:low-noise-print 600 "run zombie") ;; every five minutes is plenty
	    (launch:handle-zombie-tests run-id))
        (when do-sync
          (tests:update-central-meta-info run-id test-id new-cpu-load new-disk-free (calc-minutes) #f #f))
        
	(if kill-job? 
	    (begin
              (debug:print-info 0 *default-log-port* "proceeding to kill test: "kill-reason)
	      (mutex-lock! m)
	      ;; NOTE: The pid can change as different steps are run. Do we need handshaking between this
	      ;;       section and the runit section? Or add a loop that tries three times with a 1/4 second
	      ;;       between tries?
	      (let* ((pid1 (launch:einf-pid exit-info)) ;; (vector-ref exit-info 0))
		     (pid2 (rmt:test-get-top-process-pid run-id test-id))
		     (pids (delete-duplicates (filter number? (list pid1 pid2)))))
		(if (not (null? pids))
		    (begin
		      (for-each
		       (lambda (pid)
			 (handle-exceptions
			  exn
			  (begin
			    (debug:print-info 0 *default-log-port* "Unable to kill process with pid " pid ", possibly already killed.")
			    (debug:print 0 *default-log-port* " message: " ((condition-property-accessor 'exn 'message) exn) ", exn=" exn))
			  (debug:print 0 *default-log-port* "WARNING: Request received to kill job " pid) ;;  " (attempt # " kill-tries ")")
			  (debug:print-info 0 *default-log-port* "Signal mask=" (signal-mask))
			  ;; (if (process:alive? pid)
			  ;;     (begin
			  (map (lambda (pid-num)
				 (process-signal pid-num signal/term))
			       (process:get-sub-pids pid))
			  (thread-sleep! 5)
			  ;; (if (process:process-alive? pid)
			  (map (lambda (pid-num)
				 (handle-exceptions
				     exn
				   (begin
				     (debug:print 0 *default-log-port* " .... had trouble sending kill to " pid-num ", exn=" exn)
				     #f)
				   (process-signal pid-num signal/kill)))
			       (process:get-sub-pids pid))))
		       ;;    (debug:print-info 0 *default-log-port* "not killing process " pid " as it is not alive"))))
		       pids)
                      ;; BB: question to Matt -- does the tests:test-state-status! encompass rollup to toplevel?  If not, should it?
		      (tests:test-set-status! run-id test-id "KILLED"  "KILLED" (conc (args:get-arg "-m")" "kill-reason) #f)) ;; BB ADDED kill-reason -- confirm OK with Matt
		    (begin
		      (debug:print-error 0 *default-log-port* "Nothing to kill, pid1=" pid1 ", pid2=" pid2)
		      (tests:test-set-status! run-id test-id "KILLED"  "FAILED TO KILL" (conc (args:get-arg "-m")" "kill-reason) #f) ;; BB ADDED kill-reason -- confirm OK with Matt
		      )))
	      (mutex-unlock! m)
	      ;; no point in sticking around. Exit now. But run end of run before exiting?
              (launch:end-of-run-check run-id)
	      (exit)))
	(if (hash-table-ref/default misc-flags 'keep-going #f)  ;; keep originals for cpu-load and disk-free unless they change more than the allowed delta
	    (begin
	      (thread-sleep! 6) ;; was 3
	      (loop (calc-minutes)
		    (or new-cpu-load cpu-load)
		    (or new-disk-free disk-free)
		    (if do-sync (current-seconds) last-sync))))))
    (tests:update-central-meta-info run-id test-id (commonmod:get-cpu-load) (get-df (current-directory))(calc-minutes) #f #f))) ;; NOTE: Checking twice for keep-going is intentional


;; 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".logpro")
	       (with-output-to-file (conc name".logpro")
		 (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)
    ;;(bb-check-path msg: "launch:execute incoming")
    (if (list? cmdinfo) ;; ((testpath /tmp/mrwellan/jazzmind/src/example_run/tests/sqlitespeed)
	;; (test-name sqlitespeed) (runscript runscript.rb) (db-host localhost) (run-id 1))
	(let* ((testpath  (assoc/default 'testpath  cmdinfo))  ;; testpath is the test spec area
	       (top-path  (assoc/default 'toppath   cmdinfo))
	       (work-area (assoc/default 'work-area cmdinfo))  ;; work-area is the test run area
	       (test-name (assoc/default 'test-name cmdinfo))
	       (runscript (assoc/default 'runscript cmdinfo))
	       (ezsteps   (assoc/default 'ezsteps   cmdinfo))
	       (subrun    (assoc/default 'subrun    cmdinfo))
	       ;; (runremote (assoc/default 'runremote cmdinfo))
	       ;; (transport (assoc/default 'transport cmdinfo))  ;; not used
	       ;; (serverinf (assoc/default 'serverinf cmdinfo))
	       ;; (port      (assoc/default 'port      cmdinfo))
	       (serverurl (assoc/default 'serverurl cmdinfo))
	       (homehost  (assoc/default 'homehost  cmdinfo))
	       (run-id    (assoc/default 'run-id    cmdinfo))
	       (test-id   (assoc/default 'test-id   cmdinfo))
	       (target    (assoc/default 'target    cmdinfo))
	       (areaname  (assoc/default 'areaname  cmdinfo))
	       (itemdat   (assoc/default 'itemdat   cmdinfo))
	       (env-ovrd  (assoc/default 'env-ovrd  cmdinfo))
	       (set-vars  (assoc/default 'set-vars  cmdinfo)) ;; pre-overrides from -setvar
	       (runname   (assoc/default 'runname   cmdinfo))
	       (megatest  (assoc/default 'megatest  cmdinfo))
	       (runtlim   (assoc/default 'runtlim   cmdinfo))
	       (contour   (assoc/default 'contour   cmdinfo))
	       (item-path (item-list->path itemdat))
	       (mt-bindir-path (assoc/default 'mt-bindir-path cmdinfo))
	       (keys      #f)
	       (keyvals   #f)
	       (fullrunscript (if (not runscript)
                                  #f
                                  (if (substring-index "/" runscript)
                                      runscript ;; use unadultered if contains slashes
                                      (let ((fulln (conc work-area "/" runscript)))
	                                  (if (and (common:file-exists? fulln)
                                                   (file-execute-access? fulln))
                                              fulln
                                              runscript))))) ;; assume it is on the path
               (check-work-area           (lambda ()
                                            ;; NFS might not have propagated the directory meta data to the run host - give it time if needed
                                            (let loop ((count 0))
                                              (if (or (common:directory-exists? work-area)
                                                      (> count 10))
                                                  (change-directory work-area)
                                                  (begin
                                                    (debug:print 0 *default-log-port* "INFO: Not starting job yet - directory " work-area " not found")
                                                    (thread-sleep! 10)
                                                    (loop (+ count 1)))))

                                            (if (not (string=?  (common:real-path work-area)(common:real-path (current-directory))))
                                                (begin
                                                  (debug:print 0 *default-log-port*
                                                               "INFO: we are expecting to be in directory " work-area "\n"
                                                               "     but we are actually in the directory " (current-directory) "\n"
                                                               "     doing another change dir.")
                                                  (change-directory work-area)))
                                            
                                            ;; spot check that the files in testpath are available. Too often NFS delays cause problems here.
                                            (let ((files      (glob (conc testpath "/*")))
                                                  (bad-files '()))
                                              (for-each
                                               (lambda (fullname)
                                                 (let* ((fname (pathname-strip-directory fullname))
                                                        (targn (conc work-area "/" fname)))
                                                   (if (not (file-exists? targn))
                                                       (set! bad-files (cons fname bad-files)))))
                                               files)
                                              (if (not (null? bad-files))
                                                  (begin
                                                    (debug:print 0 *default-log-port* "INFO: test data from " testpath " not copied properly or filesystem problems causing data to not be found. Re-running the copy command.")
                                                    (debug:print 0 *default-log-port* "INFO: missing files from " work-area ": " (string-intersperse bad-files ", "))
                                                    (launch:test-copy testpath work-area))))
                                            ;; one more time, change to the work-area directory
                                            (change-directory work-area)))
	       ) ;; let*

	  (if contour (setenv "MT_CONTOUR" contour))
	  
	  ;; immediated set some key variables from CMDINFO data, yes, these will be set again below ...
	  ;;
	  (setenv "MT_TESTSUITENAME" areaname)
	  (setenv "MT_RUN_AREA_HOME" top-path)
	  (set! *toppath* top-path)
          (change-directory *toppath*) ;; temporarily switch to the run area home
	  (setenv "MT_TEST_RUN_DIR"  work-area)

	  (launch:setup) ;; should be properly in the run area home now

	  (if contour (setenv "MT_CONTOUR" contour))
	  
	  ;; immediated set some key variables from CMDINFO data, yes, these will be set again below ...
	  ;;
	  (setenv "MT_TESTSUITENAME" areaname)
	  (setenv "MT_RUN_AREA_HOME" top-path)
	  (set! *toppath* top-path)
          (change-directory *toppath*) ;; temporarily switch to the run area home
	  (setenv "MT_TEST_RUN_DIR"  work-area)

	  (launch:setup) ;; should be properly in the run area home now
          
	  (set! tconfigreg (tests:get-all)) ;; mapping of testname => test source path
	  (let ((sighand (lambda (signum)
			   ;; (signal-mask! signum) ;; to mask or not? seems to cause issues in exiting
			   (if (eq? signum signal/stop)
			       (debug:print-error 0 *default-log-port* "attempt to STOP process. Exiting."))
			   (set! *time-to-exit* #t)
			   (debug:print 0 *default-log-port* "Received signal " signum ", cleaning up before exit (set this test to COMPLETED/ABORT) . Please wait...")
			   (let ((th1 (make-thread (lambda ()
                                                     (debug:print 0 *default-log-port* "set test to COMPLETED/ABORT begin.")
						     (rmt:test-set-state-status run-id test-id "COMPLETED" "ABORT" "received kill signal")
                                                     (debug:print 0 *default-log-port* "set test to COMPLETED/ABORT complete.")
						     (debug:print 0 *default-log-port* "Killed by signal " signum ". Exiting")
						     (exit 1))))
				 (th2 (make-thread (lambda ()
						     (thread-sleep! 20)
						     (debug:print 0 *default-log-port* "Done")
						     (exit 4)))))
			     (thread-start! th2)
			     (thread-start! th1)
			     (thread-join! th2)))))
	    (set-signal-handler! signal/int sighand)
	    (set-signal-handler! signal/term sighand)
	    ) ;; (set-signal-handler! signal/stop sighand)
	  
	  ;; Do not run the test if it is REMOVING, RUNNING, KILLREQ or REMOTEHOSTSTART,
	  ;; Mark the test as REMOTEHOSTSTART *IMMEDIATELY*
	  ;;
	  (let* ((test-info (let loop ((tries 0))
			      (let ((tinfo (rmt:get-test-info-by-id run-id test-id)))
				(if tinfo
				    tinfo
				    (if (> tries 5)
					#f
					(begin
					  (thread-sleep! (+ 1 (* tries 10)))
					  (loop (+ tries 1))))))))
		 (test-host (if test-info
				(db:test-get-host        test-info)
				(begin
				  (debug:print 0 *default-log-port* "ERROR: failed to find a record for test-id " test-id ", exiting.")
				  (exit))))
		 (test-pid  (db:test-get-process_id  test-info)))
	    (cond
             ;; -mrw- I'm removing KILLREQ from this list so that a test in KILLREQ state is treated as a "do not run" flag.
	     ((member (db:test-get-state test-info) '("INCOMPLETE" "KILLED" "UNKNOWN" "STUCK")) ;; prior run of this test didn't complete, go ahead and try to rerun
	      (debug:print 0 *default-log-port* "INFO: test is INCOMPLETE or KILLED, treat this execute call as a rerun request")
	      ;; (tests:test-force-state-status! run-id test-id "REMOTEHOSTSTART" "n/a")

              (rmt:general-call 'set-test-start-time run-id test-id)
              (rmt:test-set-state-status run-id test-id "REMOTEHOSTSTART" "n/a" #f)
	      ) ;; prime it for running
	     ((member (db:test-get-state test-info) '("RUNNING" "REMOTEHOSTSTART"))
	      (if (process:alive-on-host? test-host test-pid)
		  (debug:print-error 0 *default-log-port* "test state is "  (db:test-get-state test-info) " and process " test-pid " is still running on host " test-host ", cannot proceed")
		  (exit 1)))
	     ((member (db:test-get-state test-info) '("COMPLETED"))  ;; we do NOT want to re-run COMPLETED jobs. Mark as NOT_STARTED to run!
	      (debug:print 0 *default-log-port* "test state is " (db:test-get-state test-info) ", cannot proceed")
	      (debug:print 0 *default-log-port* "exiting with status 1")
	      (exit 1))
	     ((not (member (db:test-get-state test-info) '("REMOVING" "REMOTEHOSTSTART" "RUNNING" "KILLREQ")))
	      ;; (tests:test-force-state-status! run-id test-id "REMOTEHOSTSTART" "n/a")
              (rmt:general-call 'set-test-start-time run-id test-id)
	      (rmt:test-set-state-status run-id test-id "REMOTEHOSTSTART" "n/a" #f))
	     (else ;; (member (db:test-get-state test-info) '("REMOVING" "REMOTEHOSTSTART" "RUNNING" "KILLREQ"))
	      (debug:print 0 *default-log-port* "test state is " (db:test-get-state test-info) ", cannot proceed")
	      (debug:print 0 *default-log-port* "exiting with status 1")
	      (exit 1))))

          ;; cleanup prior execution's steps
          (rmt:delete-steps-for-test! run-id test-id)
          
	  (debug:print 2 *default-log-port* "Executing " test-name " (id: " test-id ") on " (get-host-name))
	  (set! keys       (rmt:get-keys))
	  ;; (runs:set-megatest-env-vars run-id inkeys: keys inkeyvals: keyvals) ;; these may be needed by the launching process
	  ;; one of these is defunct/redundant ...
	  (if (not (launch:setup force-reread: #t))
	      (begin
		(debug:print 0 *default-log-port* "Failed to setup, exiting") 
		;; (sqlite3:finalize! db)
		;; (sqlite3:finalize! tdb)
		(exit 1)))
          ;; validate that the test run area is available
          (check-work-area)
          
          ;; still need to go back to run area home for next couple steps
	  (change-directory *toppath*) 

	  ;; NOTE: Current order is to process runconfigs *before* setting the MT_ vars. This 
	  ;;       seems non-ideal but could well break stuff
	  ;;    BUG? BUG? BUG?
	  
	  (let ((rconfig (full-runconfigs-read)) ;; (read-config (conc  *toppath* "/runconfigs.config") #f #t sections: (list "default" target))))
		(wconfig (read-config "waivers.config" #f #t sections: `( "default" ,target )))) ;; read the waivers config if it exists
	    ;; (setup-env-defaults (conc *toppath* "/runconfigs.config") run-id (make-hash-table) keyvals target)
	    ;; (set-run-config-vars run-id keyvals target) ;; (db:get-target db run-id))
	    ;; Now have runconfigs data loaded, set environment vars
	    (for-each
	     (lambda (section)
	       (for-each
		(lambda (varval)
		  (let ((var (car varval))
			(val (cadr varval)))
		    (if (and (string? var)(string? val))
			(begin
			  (safe-setenv var (configf:eval-string-in-environment val))) ;; val)
			(debug:print-error 0 *default-log-port* "bad variable spec, " var "=" val))))
		(configf:get-section rconfig section)))
	     (list "default" target)))
          ;;(bb-check-path msg: "launch:execute post block 1")

	  ;; NFS might not have propagated the directory meta data to the run host - give it time if needed
	  (let loop ((count 0))
	    (if (or (common:file-exists? work-area)
		    (> count 10))
		(change-directory work-area)
		(begin
		  (debug:print 0 *default-log-port* "INFO: Not starting job yet - directory " work-area " not found")
		  (thread-sleep! 10)
		  (loop (+ count 1)))))

          ;; now we can switch to the work-area?
          (change-directory work-area)
          ;;(bb-check-path msg: "launch:execute post block 1.5")
	  ;; (change-directory work-area) 
	  (set! keyvals    (keys:target->keyval keys target))
	  ;; apply pre-overrides before other variables. The pre-override vars must not
	  ;; clobbers things from the official sources such as megatest.config and runconfigs.config
	  (if (string? set-vars)
	      (let ((varpairs (string-split set-vars ",")))
		(debug:print 4 *default-log-port* "varpairs: " varpairs)
		(map (lambda (varpair)
		       (let ((varval (string-split varpair "=")))
			 (if (eq? (length varval) 2)
			     (let ((var (car varval))
				   (val (cadr varval)))
			       (debug:print 1 *default-log-port* "Adding pre-var/val " var " = " val " to the environment")
			       (setenv var val)))))
		     varpairs)))
          ;;(bb-check-path msg: "launch:execute post block 2")
	  (for-each
	   (lambda (varval)
	     (let ((var (car varval))
		   (val (cadr varval)))
	       (if val
		   (setenv var val)
		   (begin
		     (debug:print-error 0 *default-log-port* "required variable " var " does not have a valid value. Exiting")
		     (exit)))))
	     (list 
	      (list  "MT_TEST_RUN_DIR" work-area)
	      (list  "MT_TEST_NAME" test-name)
	      (list  "MT_ITEM_INFO" (conc itemdat))
	      (list  "MT_ITEMPATH"  item-path)
	      (list  "MT_RUNNAME"   runname)
	      (list  "MT_MEGATEST"  megatest)
	      (list  "MT_TARGET"    target)
	      (list  "MT_LINKTREE"  (common:get-linktree)) ;; (configf:lookup *configdat* "setup" "linktree"))
	      (list  "MT_TESTSUITENAME" (common:get-testsuite-name))))
          ;;(bb-check-path msg: "launch:execute post block 3")

	  (let ((tmppath (getenv "PATH")))
	    (if (string-search tmppath " ")
		(debug:print 0 *default-log-port* "WARNING: spaces in PATH are not supported."))
	    (if mt-bindir-path (setenv "PATH" (conc tmppath":"mt-bindir-path))))
          ;;(bb-check-path msg: "launch:execute post block 4")
	  ;; (change-directory top-path)
	  ;; Can setup as client for server mode now
	  ;; (client:setup)
  
	  ;; environment overrides are done *before* the remaining critical envars.
	  (alist->env-vars env-ovrd)
          ;;(bb-check-path msg: "launch:execute post block 41")
	  (runs:set-megatest-env-vars run-id inkeys: keys inkeyvals: keyvals)
          ;;(bb-check-path msg: "launch:execute post block 42")
	  (set-item-env-vars itemdat)
          ;;(bb-check-path msg: "launch:execute post block 43")
          (let ((blacklist (configf:lookup *configdat* "setup" "blacklistvars")))
            (if blacklist
		(let ((vars (string-split blacklist)))
		  (save-environment-as-files "megatest" ignorevars: vars)
		  (for-each (lambda (var)
			      (unsetenv var))
			    vars))
                (save-environment-as-files "megatest")))
          ;;(bb-check-path msg: "launch:execute post block 44")
	  ;; open-run-close not needed for test-set-meta-info
	  ;; (tests:set-full-meta-info #f test-id run-id 0 work-area)
	  ;; (tests:set-full-meta-info test-id run-id 0 work-area)
	  (tests:set-full-meta-info #f test-id run-id 0 work-area 10)

	  ;; (thread-sleep! 0.3) ;; NFS slowness has caused grief here

	  (if (args:get-arg "-xterm")
	      (set! fullrunscript "xterm")
	      (if (and fullrunscript 
		       (common:file-exists? fullrunscript)
		       (not (file-execute-access? fullrunscript)))
		  (system (conc "chmod ug+x " fullrunscript))))
	  (launch:extract-scripts-logpro work-area test-name item-path tconfigreg)

;;;;;	  ;; We are about to actually kick off the test
;;;;;	  ;; so this is a good place to remove the records for 
;;;;;	  ;; any previous runs
;;;;;	  ;; (db:test-remove-steps db run-id testname itemdat)
;;;;;	  ;; now is also a good time to write the .testconfig file
;;;;;	  (let* ((tconfig-fname   (conc work-area "/.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"))
;;;;;		 (precmd          (configf:lookup tconfig )
;;;;;	    ;; 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)
;;;;;		       (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))
	  ;;

	  (let* ((m            (make-mutex))
		 (kill-job?    #f)
		 (exit-info    (make-launch:einf pid: #t exit-status: #t exit-code: #t rollup-status: 0)) ;; pid exit-status exit-code (i.e. process was successfully run) rollup-status
		 (job-thread   #f)
		 ;; (keep-going   #t)
		 (misc-flags   (let ((ht (make-hash-table)))
				 (hash-table-set! ht 'keep-going #t)
				 ht))
		 (runit        (lambda ()
				 (launch:manage-steps run-id test-id item-path fullrunscript ezsteps subrun test-name tconfigreg exit-info m)))
		 (monitorjob   (lambda ()
				 (launch:monitor-job  run-id test-id item-path fullrunscript ezsteps test-name tconfigreg exit-info m work-area runtlim misc-flags)))
		 (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")
		 (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.
	    (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 ...")
            (debug:print-info 2 *default-log-port* "exit-info = " exit-info)
	    (hash-table-set! misc-flags 'keep-going #f)
	    (thread-join! th1)
	    (thread-sleep! 1)       ;; givbe thread th1 a chance to be done TODO: Verify this is needed. At 0.1 I was getting fail to stop, increased to total of 1.1 sec.
	    (mutex-lock! m)
	    (let* ((item-path (item-list->path itemdat))
		   ;; only state and status needed - use lazy routine
		   (testinfo  (rmt:get-testinfo-state-status run-id test-id)))
	      ;; Am I completed?
	      (if (member (db:test-get-state testinfo) '("REMOTEHOSTSTART" "RUNNING"))
                 (let ((new-state  (if kill-job? "KILLED" "COMPLETED"))
		        (new-status (cond
				     ((not (launch:einf-exit-status exit-info)) "FAIL") ;; job failed to run ... (vector-ref exit-info 1)
				     ((eq? (launch:einf-rollup-status exit-info) 0)     ;; (vector-ref exit-info 3)
				      ;; if the current status is AUTO then defer to the calculated value (i.e. leave this AUTO)
				      (if (equal? (db:test-get-status testinfo) "AUTO") "AUTO" "PASS"))
				     ((eq? (launch:einf-rollup-status exit-info) 1) "FAIL")  ;; (vector-ref exit-info 3)
				     ((eq? (launch:einf-rollup-status exit-info) 2)	     ;;	(vector-ref exit-info 3)
				      ;; if the current status is AUTO the defer to the calculated value but qualify (i.e. make this AUTO-WARN)
				      (if (equal? (db:test-get-status testinfo) "AUTO") "AUTO-WARN" "WARN"))
				     ((eq? (launch:einf-rollup-status exit-info) 3) "CHECK")
				     ((eq? (launch:einf-rollup-status exit-info) 4) "WAIVED")
				     ((eq? (launch:einf-rollup-status exit-info) 5) "ABORT")
				     ((eq? (launch:einf-rollup-status exit-info) 6) "SKIP")
				     (else "FAIL")))
                        ) ;; (db:test-get-status testinfo)))
		    (debug:print-info 0 *default-log-port* "Test exited in state=" (db:test-get-state testinfo) ", setting state/status based on exit code of " (launch:einf-exit-status exit-info) " and rollup-status of " (launch:einf-rollup-status exit-info))
   
                    ;; Leave a .final-status file for each sub-test
                    (tests:save-final-status run-id test-id)

		    (tests:test-set-status! run-id 
					    test-id 
					    new-state
					    new-status
					    (args:get-arg "-m") #f)
		    ;; need to update the top test record if PASS or FAIL and this is a subtest
		    ;; NO NEED TO CALL set-state-status-and-roll-up-items HERE, THIS IS DONE IN set-state-status-and-roll-up-items called by tests:test-set-status!
		 )
              )


	      ;; for automated creation of the rollup html file this is a good place...
	      (if (not (equal? item-path ""))
		  (tests:summarize-items run-id test-id test-name #f))
	      ;; BUG was this meant to be the antecnt of the if above?
	      ;; BUG was this meant to be the antecnt of the if above?
	      (tests:summarize-test run-id test-id)  ;; don't force - just update if no
              ;; Leave a .final-status file for the top level test
              (tests:save-final-status run-id test-id)
	      (rmt:update-run-stats run-id (rmt:get-raw-run-stats run-id))) ;; end of let*

	    (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)))

            ;; 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 (not (launch:einf-exit-status exit-info))
		(exit 4))))
        )))

;; DO NOT USE - caching of configs is handled in launch:setup now.
;;
(define (launch:cache-config)
  ;; if we have a linktree and -runtests and -target and the directory exists dump the config
  ;; to megatest-(current-seconds).cfg and symlink it to megatest.cfg
  (if (and *configdat* 
	   (or (args:get-arg "-run")
	       (args:get-arg "-runtests")
	       (args:get-arg "-execute")))
      (let* ((linktree (common:get-linktree)) ;; (get-environment-variable "MT_LINKTREE"))
	     (target   (common:args-get-target exit-if-bad: #t))
	     (runname  (or (args:get-arg "-runname")
			   (args:get-arg ":runname")
			   (getenv "MT_RUNNAME")))
	     (fulldir  (conc linktree "/"
			     target "/"
			     runname)))
	(if (and linktree (common:file-exists? linktree)) ;; can't proceed without linktree
	    (begin
	      (debug:print-info 0 *default-log-port* "Have -run with target=" target ", runname=" runname ", fulldir=" fulldir ", testpatt=" (or (args:get-arg "-testpatt") "%"))
	      (if (not (common:file-exists? fulldir))
		  (create-directory fulldir #t)) ;; need to protect with exception handler 
	      (if (and target
		       runname
		       (common:file-exists? fulldir))
		  (let ((tmpfile  (conc fulldir "/.megatest.cfg." (current-seconds)))
			(targfile (conc fulldir "/.megatest.cfg-"  megatest-version "-" megatest-fossil-hash))
			(rconfig  (conc fulldir "/.runconfig." megatest-version "-" megatest-fossil-hash)))
		    (if (common:file-exists? rconfig) ;; only cache megatest.config AFTER runconfigs has been cached
			(begin
			  (debug:print-info 0 *default-log-port* "Caching megatest.config in " tmpfile)
                          (if (not (common:in-running-test?))
                              (configf:write-alist *configdat* tmpfile))
			  (system (conc "ln -sf " tmpfile " " targfile))))
		    )))
	    (debug:print-info 1 *default-log-port* "No linktree yet, no caching configs.")))))

(define (launch:handle-zombie-tests run-id)
  (let* ((key (conc "zombiescan-runid-"run-id))
         (now (current-seconds))
         (threshold (- (current-seconds)  (* 2 (or (configf:lookup-number *configdat* "setup" "deadtime") 120))))
         (val (rmt:get-var key))
         (do-scan?
          (cond
           ((not val)
            #t)
           ((< val threshold)
            #t)
           (else #f))))
    (when do-scan?
      (debug:print 1 *default-log-port* "INFO: search and mark zombie tests")
      (rmt:set-var key (current-seconds))
      (rmt:find-and-mark-incomplete run-id #f))))

;; recover a test where the top controlling mtest may have died
;;
(define (launch:recover-test run-id test-id)
  ;; this function is called on the test run host via ssh
  ;;
  ;; 1. look at the process from pid
  ;;    - is it owned by calling user
  ;;    - it it's run directory correct for the test
  ;;    - is there a controlling mtest (maybe stuck)
  ;; 2. if recovery is needed watch pid
  ;;    - when it exits take the exit code and do the needful
  ;;
  (let* ((pid (rmt:test-get-top-process-pid run-id test-id))
	 (psres (with-input-from-pipe
		 (conc "ps -F -u " (current-user-name) " | grep -E '" pid " ' | grep -v 'grep -E " pid "'")
		 (lambda ()
		   (read-line))))
	 (rundir (if (string? psres) ;; real process owned by user
		     (read-symbolic-link (conc "/proc/" pid "/cwd"))
		     #f)))
    ;; now wait on that process if all is correct
    ;; periodically update the db with runtime
    ;; when the process exits look at the db, if still RUNNING after 10 seconds set
    ;; state/status appropriately
    (process-wait pid)))

;;======================================================================
;; Maintenance
;;======================================================================

(define (rmt:find-and-mark-incomplete run-id ovr-deadtime)
  (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)))

;;  select end_time-now from
;;      (select testname,item_path,event_time+run_duration as
;;                          end_time,strftime('%s','now') as now from tests where state in
;;      ('RUNNING','REMOTEHOSTSTART','LAUNCHED'));
;;
;; NOT EASY TO MIGRATE TO db{file,mod}
;;
(define (rmt:find-and-mark-incomplete-engine run-id ovr-deadtime cfg-deadtime test-stats-update-period)
  (let* ((incompleted '())
	 (oldlaunched '())
	 (toplevels   '())
          ;; The default running-deadtime is 720 seconds = 12 minutes.
          ;; "(running-deadtime-default (+ server-start-allowance (* 2 launch-monitor-period)))" = 200 + (2 * (200 + 30 + 30))
         (deadtime-trim (or ovr-deadtime cfg-deadtime))
         (server-start-allowance 200)
         (server-overloaded-budget 200)
         (launch-monitor-off-time (or test-stats-update-period 30))
         (launch-monitor-on-time-budget 30)
         (launch-monitor-period (+ launch-monitor-off-time launch-monitor-on-time-budget server-overloaded-budget))
         (remotehoststart-deadtime-default (+ server-start-allowance server-overloaded-budget 30))
         (remotehoststart-deadtime (or deadtime-trim remotehoststart-deadtime-default))
         (running-deadtime-default (+ server-start-allowance (* 2 launch-monitor-period)))
         (running-deadtime (or deadtime-trim running-deadtime-default))) ;; two minutes (30 seconds between updates, this leaves 3x grace period)

    (debug:print-info 4  *default-log-port* "running-deadtime = " running-deadtime)
    (debug:print-info 4  *default-log-port* "deadtime-trim = " deadtime-trim)

    (let* ((dat (rmt:get-toplevels-and-incompletes run-id running-deadtime remotehoststart-deadtime)))
      (set! oldlaunched (list-ref dat 1))
      (set! toplevels   (list-ref dat 2))
      (set! incompleted (list-ref dat 0)))

    (debug:print-info 18 *default-log-port* "Found " (length oldlaunched) " old LAUNCHED items, "
		      (length toplevels) " old LAUNCHED toplevel tests and "
		      (length incompleted) " tests marked RUNNING but apparently dead.")
  
    ;; These are defunct tests, do not do all the overhead of set-state-status. Force them to INCOMPLETE.
    ;;
    ;; (db:delay-if-busy dbdat)
    (let* ((min-incompleted-ids (map car incompleted)) ;; do 'em all
	   (all-ids             (append min-incompleted-ids (map car oldlaunched))))
      (if (> (length all-ids) 0)
	  (begin
	    ;; (launch:is-test-alive "localhost" 435)
	    (debug:print 0 *default-log-port* "WARNING: Marking test(s); " (string-intersperse (map conc all-ids) ", ")
			 " as DEAD")
	    (for-each
             (lambda (test-id)
               (let* ((tinfo   (rmt:get-test-info-by-id run-id test-id))
		      (run-dir (db:test-get-rundir     tinfo))
		      (host    (db:test-get-host       tinfo))
		      (pid     (db:test-get-process_id tinfo))
		      (result (rmt:get-status-from-final-status-file run-dir)))
		 (if (and (list? result) (> (length result) 1) (equal? "PASS" (cadr result)) (equal? "COMPLETED" (car result))) 
		     (begin
		       (debug:print 0 *default-log-port* "INFO: test " test-id " actually passed, so marking PASS not DEAD")
		       (rmt:set-state-status-and-roll-up-items
			run-id test-id 'foo "COMPLETED" "PASS"
			"Test stopped responding but it has PASSED; marking it PASS in the DB."))
		     (let ((is-alive (and (not (eq? pid 0))  ;; 0 is default in re-used field "attemptnum" where pid stored.
					  (commonmod:is-test-alive host pid))))
		       (if is-alive
			   (debug:print 0 *default-log-port* "INFO: test " test-id " on host " host
					" has a process on pid " pid ", NOT setting to DEAD.")
			   (begin
			     (debug:print 0 *default-log-port* "INFO: test " test-id
					  " final state/status is not COMPLETED/PASS. It is " result)
			     (rmt:set-state-status-and-roll-up-items
			      run-id test-id 'foo "COMPLETED" "DEAD"
			      "Test stopped responding while in RUNNING or REMOTEHOSTSTART; presumed dead.")))))))
	     ;; call end of eud of run detection for posthook - from merge, is it needed?
	     ;; (launch:end-of-run-check run-id)
	     all-ids)
	    )))))

;; set up needed environment variables given a run-id and optionally a target, itempath etc.
;;
(define (runs:set-megatest-env-vars run-id #!key (inkeys #f)(inrunname #f)(inkeyvals #f)(intarget #f)(testname #f)(itempath #f))
  ;;(bb-check-path msg: "runs:set-megatest-env-vars entry")
  (let* ((target    (or intarget 
			(common:args-get-target)
			(get-environment-variable "MT_TARGET")))
	 (keys      (if inkeys    inkeys    (rmt:get-keys)))
	 (keyvals   (if inkeyvals inkeyvals (keys:target->keyval keys target)))
	 (vals      (hash-table-ref/default *env-vars-by-run-id* run-id #f))
	 (link-tree (common:get-linktree))) ;; (configf:lookup *configdat* "setup" "linktree")))
    (if testname (setenv "MT_TEST_NAME" testname))
    (if itempath (setenv "MT_ITEMPATH"  itempath))

    ;; get the info from the db and put it in the cache
    (if link-tree
	(setenv "MT_LINKTREE" link-tree)
	(debug:print-error 0 *default-log-port* "linktree not set, should be set in megatest.config in [setup] section."))
    (if (not vals)
	(let ((ht (make-hash-table)))
	  (hash-table-set! *env-vars-by-run-id* run-id ht)
	  (set! vals ht)
	  (for-each
	   (lambda (key)
	     (hash-table-set! vals (car key) (cadr key)))
	   keyvals)))
    ;; from the cached data set the vars
    
    (hash-table-for-each
     vals
     (lambda (key val)
       (debug:print 2 *default-log-port* "setenv " key " " val)
       (safe-setenv key val)))
    ;;(bb-check-path msg: "runs:set-megatest-env-vars block 1")
    ;;(BB> "*env-vars-by-run-id*/runid("run-id" vals="(hash-table->alist vals))

    (if (not (get-environment-variable "MT_TARGET"))(setenv "MT_TARGET" target))
    ;; we had a case where there was an exception generated by the hash-table-ref
    ;; due to *configdat* being #f Adding a handle and exit
    (let fatal-loop ((count 0)) 
      (handle-exceptions
	  exn
	  (let ((call-chain (get-call-chain))
		(msg        ((condition-property-accessor 'exn 'message) exn)))
	    (if (< count 5)
		(begin ;; this call is colliding, do some crude stuff to fix it.
		  (debug:print 0 *default-log-port* "ERROR: *configdat* was inaccessible! This should never happen. Retry #" count
			       ", exn=" exn)
		  (launch:setup force-reread: #t)
		  (fatal-loop (+ count 1))) 
		(begin
		  (debug:print 0 *default-log-port* "FATAL: *configdat* was inaccessible! This should never happen. Retried " count
			       " times. Message: " msg)
		  (debug:print 0 *default-log-port* "Call chain:")
		  (with-output-to-port *default-log-port*
		    (lambda ()
		      (print "*configdat* is >>"*configdat*"<<")
                      (pp *configdat*)
                      (pp call-chain)))
                  
		  (exit 1))))
          ;;(bb-check-path msg: "runs:set-megatest-env-vars block 1.5")
          (when (or (not *configdat*) (not (hash-table? *configdat*)))
              (debug:print 0 *default-log-port* "WARNING: *configdat* was inaccessible! This should never happen.  Brute force reread.")
              ;;(BB> "ERROR: *configdat* was inaccessible! This should never happen.  Brute force reread.")
              (thread-sleep! 2) ;; assuming nfs lag.
              (launch:setup force-reread: #t))
          (alist->env-vars (hash-table-ref/default *configdat* "env-override" '())))) ;;;; environment is tainted HERE in this let block.
    ;;(bb-check-path msg: "runs:set-megatest-env-vars block 2")
    ;; Lets use this as an opportunity to put MT_RUNNAME in the environment
    (let ((runname  (if inrunname inrunname (rmt:get-run-name-from-id run-id))))
      (if runname
	  (setenv "MT_RUNNAME" runname)
	  (debug:print-error 0 *default-log-port* "no value for runname for id " run-id)))
    (setenv "MT_RUN_AREA_HOME" *toppath*)
    ;; if a testname and itempath are available set the remaining appropriate variables
    (if testname (setenv "MT_TEST_NAME" testname))
    (if itempath (setenv "MT_ITEMPATH"  itempath))
    ;;(bb-check-path msg: "runs:set-megatest-env-vars block 3")
    (if (and testname link-tree)
	(setenv "MT_TEST_RUN_DIR" (conc (getenv "MT_LINKTREE")  "/"
					(getenv "MT_TARGET")    "/"
					(getenv "MT_RUNNAME")   "/"
					(getenv "MT_TEST_NAME")
					(if (and itempath
						 (not (equal? itempath "")))
					    (conc "/" itempath)
					    ""))))))
(define (set-item-env-vars itemdat)
  (for-each (lambda (item)
	      (debug:print 2 *default-log-port* "setenv " (car item) " " (cadr item))
	      (setenv (car item) (cadr item)))
	    itemdat))


;; TODO: deprecate me in favor of ezsteps.scm
;;
(define (ezsteps:runstep ezstep run-id test-id exit-info m tal testconfig all-steps-dat)
  (let* ((stepname       (car ezstep))  ;; do stuff to run the step
	 (stepinfo       (cadr ezstep))
	;; (let ((info (cadr ezstep)))
	;; 		   (if (proc? info) "" info)))
	;; (stepproc       (let ((info (cadr ezstep)))
	;; 		   (if (proc? info) info #f)))
	 (stepparts      (string-match (regexp "^(\\{([^\\}\\{]*)\\}\\s*|)(.*)$") stepinfo))
	 (stepparams     (if (and (list? stepparts)
				  (> (length stepparts) 1))
			     (list-ref stepparts 2)
			     #f)) ;; for future use, {VAR=1,2,3}, run step for each
	 (paramparts     (if (string? stepparams)
			     (map (lambda (x)(string-split x "=")) (string-split-fields "[^;]*=[^;]*" stepparams))
			     '()))
	 (subrun         (alist-ref "subrun" paramparts equal?))
	 (stepcmd        (if (and (list? stepparts)
				  (> (length stepparts) 2))
			     (list-ref stepparts 3)
			     (conc "# error, no command for step "stepname)))
	 (script         "") ; "#!/bin/bash\n") ;; yep, we depend on bin/bash FIXME!!!\
	 (logpro-file    (conc stepname ".logpro"))
	 (html-file      (conc stepname ".html"))
	 (dat-file       (conc stepname ".dat"))
	 (tconfig-logpro (configf:lookup testconfig "logpro" stepname))
	 (logpro-used    (common:file-exists? logpro-file))
	 (mtexepath      (common:get-megatest-exe-path)))
    (setenv "MT_STEP_NAME" stepname)
    (hash-table-set! all-steps-dat stepname `((params . ,paramparts)))
    (debug:print 0 *default-log-port* "stepparts: " stepparts ", stepparams: " stepparams
                 ", paramparts: " paramparts ", subrun: " subrun ", stepcmd: " stepcmd)
    
    (if (and tconfig-logpro
	     (not logpro-used)) ;; no logpro file found but have a defn in the testconfig
	(begin
	  (with-output-to-file logpro-file
	    (lambda ()
	      (print ";; logpro file extracted from testconfig\n"
		     ";;")
	      (print tconfig-logpro)))
	  (set! logpro-used #t)))
    
    ;; NB// can safely assume we are in test-area directory
    (debug:print 4 *default-log-port* "ezsteps:\n stepname: " stepname " stepinfo: " stepinfo " stepparts: " stepparts
		 " stepparams: " stepparams " stepcmd: " stepcmd)
    
    ;; ;; first source the previous environment
    ;; (let ((prev-env (conc ".ezsteps/" prevstep (if (string-search (regexp "csh") 
    ;;      							 (get-environment-variable "SHELL")) ".csh" ".sh"))))
    ;;   (if (and prevstep (common:file-exists? prev-env))
    ;;       (set! script (conc script "source " prev-env))))
    
    ;; call the command using mt_ezstep
    ;; (set! script (conc "mt_ezstep " stepname " " (if prevstep prevstep "x") " " stepcmd))
    
    (debug:print 4 *default-log-port* "script: " script)
    (rmt:teststep-set-status! run-id test-id stepname "start" "-" #f #f)
    ;; now launch the actual process
    (call-with-environment-variables 
     (list (cons "PATH" mtexepath))
     (lambda () ;; (process-run "/bin/bash" "-c" "exec ls -l /tmp/foobar > /tmp/delme-more.log 2>&1")
       (let* ((cmd (conc stepcmd " > " stepname ".log 2>&1")) ;; >outfile 2>&1 
	      (pid #f))
	 (let ((proc (lambda ()
		       (set! pid (process-run "/bin/bash" (list "-c" cmd))))))
	   (if subrun
               (begin
                 (debug:print-info 0 *default-log-port* "Running "cmd" without MT_.* environment variables.")
                 (common:without-vars proc "^MT_.*"))
	       (proc)))
	 
         (with-output-to-file "Makefile.ezsteps"
           (lambda ()
             (print stepname ".log :")
             (print "\t" cmd)
             (if (common:file-exists? (conc stepname ".logpro"))
                 (print "\tlogpro " stepname ".logpro " stepname ".html < " stepname ".log"))
             (print)
             (print stepname " : " stepname ".log")
             (print))
           #:append)

	 (rmt:test-set-top-process-pid run-id test-id pid)
	 (let processloop ((i 0))
	   (let-values (((pid-val exit-status exit-code)(process-wait pid #t)))
		       (mutex-lock! m)
		       (launch:einf-pid-set!         exit-info pid)         ;; (vector-set! exit-info 0 pid)
		       (launch:einf-exit-status-set! exit-info exit-status) ;; (vector-set! exit-info 1 exit-status)
		       (launch:einf-exit-code-set!   exit-info exit-code)   ;; (vector-set! exit-info 2 exit-code)
		       (mutex-unlock! m)
		       (if (eq? pid-val 0)
			   (begin
			     (thread-sleep! 2)
			     (processloop (+ i 1))))
		       )))))
    (debug:print-info 0 *default-log-port* "step " stepname " completed with exit code " (launch:einf-exit-code exit-info)) ;; (vector-ref exit-info 2))
    ;; now run logpro if needed
    (if logpro-used
	(let* ((logpro-exe (or (getenv "LOGPRO_EXE") "logpro"))
               (pid        (process-run (conc "/bin/sh -c '"logpro-exe" "logpro-file " " (conc stepname ".html") " < " stepname ".log > /dev/null'"))))
	  (let processloop ((i 0))
	    (let-values (((pid-val exit-status exit-code)(process-wait pid #t)))
			(mutex-lock! m)
			;; (make-launch:einf pid: pid exit-status: exit-status exit-code: exit-code)
			(launch:einf-pid-set!         exit-info pid)         ;; (vector-set! exit-info 0 pid)
			(launch:einf-exit-status-set! exit-info exit-status) ;; (vector-set! exit-info 1 exit-status)
			(launch:einf-exit-code-set!   exit-info exit-code)   ;; (vector-set! exit-info 2 exit-code)
			(mutex-unlock! m)
			(if (eq? pid-val 0)
			    (begin
			      (thread-sleep! 2)
			      (processloop (+ i 1)))))
	    (debug:print-info 0 *default-log-port* "logpro for step " stepname " exited with code " (launch:einf-exit-code exit-info))))) ;; (vector-ref exit-info 2)))))
    
    (let ((exinfo (launch:einf-exit-code exit-info)) ;; (vector-ref exit-info 2))
	  (logfna (if logpro-used (conc stepname ".html") ""))
	  (comment #f))
      (if logpro-used
	  (let ((datfile (conc stepname ".dat")))
	    ;; load the .dat file into the test_data table if it exists
	    (if (common:file-exists? datfile)
		(set! comment (launch:load-logpro-dat run-id test-id stepname)))
	    (rmt:test-set-log! run-id test-id (conc stepname ".html"))))
      (rmt:teststep-set-status! run-id test-id stepname "end" exinfo comment logfna))
    ;; set the test final status
    (let* ((process-exit-status (launch:einf-exit-code exit-info)) ;; (vector-ref exit-info 2))
	   (this-step-status (cond
			      ((and (eq? process-exit-status 2) logpro-used) 'warn)   ;; logpro 2 = warnings
			      ((and (eq? process-exit-status 3) logpro-used) 'check)  ;; logpro 3 = check
			      ((and (eq? process-exit-status 4) logpro-used) 'waived) ;; logpro 4 = waived
			      ((and (eq? process-exit-status 5) logpro-used) 'abort)  ;; logpro 5 = abort
			      ((and (eq? process-exit-status 6) logpro-used) 'skip)   ;; logpro 6 = skip
			      ((eq? process-exit-status 0)                   'pass)   ;; logpro 0 = pass
			      (else 'fail)))
	   (overall-status   
                             (cond
			      ((eq? (launch:einf-rollup-status exit-info) 2) 'warn) ;; rollup-status (vector-ref exit-info 3)
			      ((eq? (launch:einf-rollup-status exit-info) 3) 'check) 
			      ((eq? (launch:einf-rollup-status exit-info) 4) 'waived)
			      ((eq? (launch:einf-rollup-status exit-info) 5) 'abort)
			      ((eq? (launch:einf-rollup-status exit-info) 6) 'skip)
			      ((eq? (launch:einf-rollup-status exit-info) 0) 'pass) 
			      (else 'fail)))
	   (next-status      (common:worse-status-sym this-step-status overall-status))

	   (next-state       ;; "RUNNING") ;; WHY WAS THIS CHANGED TO NOT USE (null? tal) ??
	    (cond
	     ((null? tal) ;; more to run?
	      "COMPLETED")
	     (else "RUNNING"))))
      (debug:print 4 *default-log-port* "Exit value received: " (launch:einf-exit-code exit-info) " logpro-used: " logpro-used 
		   " this-step-status: " this-step-status " overall-status: " overall-status 
		   " next-status: " next-status " rollup-status: "  (launch:einf-rollup-status exit-info)) ;; (vector-ref exit-info 3))
      (case next-status
	((warn)
	 (launch:einf-rollup-status-set! exit-info 2) ;; (vector-set! exit-info 3 2) ;; rollup-status
	 ;; NB// test-set-status! does rdb calls under the hood
	 (tests:test-set-status! run-id test-id next-state "WARN" 
				 (if (eq? this-step-status 'warn) "Logpro warning found" #f)
				 #f))
	((check)
	 (launch:einf-rollup-status-set! exit-info 3) ;; (vector-set! exit-info 3 3) ;; rollup-status
	 ;; NB// test-set-status! does rdb calls under the hood
	 (tests:test-set-status! run-id test-id next-state "CHECK" 
				 (if (eq? this-step-status 'check) "Logpro check found" #f)
				 #f))
	((waived)
	 (launch:einf-rollup-status-set! exit-info 4) ;; (vector-set! exit-info 3 3) ;; rollup-status
	 ;; NB// test-set-status! does rdb calls under the hood
	 (tests:test-set-status! run-id test-id next-state "WAIVED" 
				 (if (eq? this-step-status 'check) "Logpro waived found" #f)
				 #f))
	((abort)
	 (launch:einf-rollup-status-set! exit-info 5) ;; (vector-set! exit-info 3 4) ;; rollup-status
	 ;; NB// test-set-status! does rdb calls under the hood
	 (tests:test-set-status! run-id test-id next-state "ABORT" 
				 (if (eq? this-step-status 'abort) "Logpro abort found" #f)
				 #f))
	((skip)
	 (launch:einf-rollup-status-set! exit-info 6) ;; (vector-set! exit-info 3 4) ;; rollup-status
	 ;; NB// test-set-status! does rdb calls under the hood
	 (tests:test-set-status! run-id test-id next-state "SKIP" 
				 (if (eq? this-step-status 'skip) "Logpro skip found" #f)
				 #f))
	((pass)
	 (tests:test-set-status! run-id test-id next-state "PASS" #f #f))
	(else ;; 'fail
	 (launch:einf-rollup-status-set! exit-info 1) ;; (vector-set! exit-info 3 1) ;; force fail, this used to be next-state but that doesn't make sense. should always be "COMPLETED" 
	 (tests:test-set-status! run-id test-id "COMPLETED" "FAIL" (conc "Failed at step " stepname) #f)
	 )))
    logpro-used))

)