Megatest

;; 
;;     You should have received a copy of the GNU General Public License
;;     along with Megatest.  If not, see <http://www.gnu.org/licenses/>.

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

(declare (unit launchmod))

(declare (uses commonmod))
(declare (uses configfmod))
(declare (uses csv-xml))
(declare (uses dbmod))
(declare (uses debugprint))
(declare (uses keysmod))
(declare (uses mtargs))
(declare (uses mtmod))
(declare (uses mtver))
(declare (uses processmod))
(declare (uses rmtmod))
(declare (uses servermod))
(declare (uses testsmod))
(declare (uses ezstepsmod))

(module launchmod
	*
	
(import scheme
	(prefix sqlite3 sqlite3:)
	chicken.base
	chicken.condition
	chicken.file
	chicken.file.posix
	chicken.format
	chicken.io
	chicken.pathname
	chicken.port
	chicken.pretty-print
	chicken.process
	chicken.process-context
	chicken.process-context.posix
	chicken.process.signal
	chicken.sort
	chicken.string
	chicken.time
	chicken.time.posix

	(prefix base64 base64:)
	csv-xml
	directory-utils
	matchable
	regex
	s11n
	srfi-1
	srfi-13
	srfi-18
	srfi-69
	stack
	system-information
       
	typed-records
	z3
	sxml-serializer
	sxml-modifications
	(prefix sxml-modifications sxml-)
	sxml-transforms
	chicken.bitwise
	
	)

(import	(prefix mtargs args:))
(import	commonmod)
(import	configfmod)
(import	dbmod)
(import	debugprint)
(import	keysmod)
(import	mtmod)
(import	mtver)
(import	processmod)
(import	rmtmod)
(import	servermod)
(import	testsmod)
(import ezstepsmod)

(include "db_records.scm")
(include "key_records.scm")

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

;; ezsteps were going to be coded as
;; stepname[,predstep1,predstep2 ...] [{VAR1=first,second,third}] command to execute
;;   BUT

		      (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?
        (runs:end-of-run-check run-id)
	      (exit)))
	(if (hash-table-ref/default misc-flags 'keep-going #f)
	    (begin
	      (thread-sleep! 3) ;; (+ 3 (pseudo-random-integer 6))) ;; add some jitter to the call home time to spread out the db accesses
	      (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
		  (loop (calc-minutes)
                        (or new-cpu-load cpu-load)

	      (if (not (equal? item-path ""))
		      (tests:summarize-items run-id test-id test-name #f))
	      (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)))
	    (mutex-unlock! m)
            (runs: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")
	    (if (not (launch:einf-exit-status exit-info))
		(exit 4))))
        )))

;; launch:end-of-run-check moved to runs:end-of-run-check

























;; 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")

    ;; 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)))


(define (runs:remove-test-directory test mode) ;; remove-data-only)
  (let* ((run-dir       (db:test-get-rundir test))    ;; run dir is from the link tree
	 (real-dir      (if (common:file-exists? run-dir)
			    ;; (resolve-pathname run-dir)
			    (common:nice-path run-dir)
			    #f))
         (clean-mode    (or mode 'remove-all))
         (test-id       (db:test-get-id test))
        ;; (lock-key      (conc "test-" test-id))
        ;; (got-lock      (let loop ((lock        (rmt:no-sync-get-lock lock-key))
	;; 			     (expire-time (+ (current-seconds) 30))) ;; give up on getting the lock and steal it after 15 seconds
	;; 		    (if (car lock)
	;; 			#t
	;; 			(if (> (current-seconds) expire-time)
	;; 			    (begin
	;; 			      (debug:print-info 0 *default-log-port* "Timed out waiting for a lock to clean test with id " test-id)
	;; 			      (rmt:no-sync-del! lock-key) ;; destroy the lock
	;; 			      (loop (rmt:no-sync-get-lock lock-key) expire-time)) ;; 
	;; 			    (begin
	;; 			      (thread-sleep! 1)
	 ;; 			      (loop (rmt:no-sync-get-lock lock-key) expire-time)))))))
	 )
    (case clean-mode
      ((remove-data-only)(mt:test-set-state-status-by-id (db:test-get-run_id test)(db:test-get-id test) "CLEANING" "LOCKED" #f))
      ((remove-all)      (mt:test-set-state-status-by-id (db:test-get-run_id test)(db:test-get-id test) "REMOVING" "LOCKED" #f))
      ((archive-remove)  (mt:test-set-state-status-by-id (db:test-get-run_id test)(db:test-get-id test) "ARCHIVE_REMOVING" #f #f)))
    (debug:print-info 1 *default-log-port* "Attempting to remove " (if real-dir (conc " dir " real-dir " and ") "") " link " run-dir)
    (if (and real-dir 
	     (> (string-length real-dir) 5)
	     (common:file-exists? real-dir)) ;; bad heuristic but should prevent /tmp /home etc.
	(let* ((realpath (realpath run-dir)))
	  (debug:print-info 1 *default-log-port* "Recursively removing " realpath)
	  (if (common:file-exists? realpath)
	      (runs:safe-delete-test-dir realpath)
	      (debug:print 0 *default-log-port* "WARNING: test dir " realpath " appears to not exist or is not readable")))
	(if real-dir 
	    (debug:print 0 *default-log-port* "WARNING: directory " real-dir " does not exist")
	    (debug:print 0 *default-log-port* "WARNING: no real directory corrosponding to link " run-dir ", nothing done")))
    (if (symbolic-link? run-dir)
	(begin
	  (debug:print-info 1 *default-log-port* "Removing symlink " run-dir)
	  (handle-exceptions
	   exn
	   (debug:print-error 0 *default-log-port* " Failed to remove symlink " run-dir ((condition-property-accessor 'exn 'message) exn) ", attempting to continue, exn=" exn)
	   (delete-file run-dir)))
	(if (directory? run-dir)
	    (if (> (directory-fold (lambda (f x)(+ 1 x)) 0 run-dir) 0)
		(debug:print 0 *default-log-port* "WARNING: refusing to remove " run-dir " as it is not empty")
		(handle-exceptions
		 exn
		 (debug:print-error 0 *default-log-port* " Failed to remove directory " run-dir ((condition-property-accessor 'exn 'message) exn) ", attempting to continue, exn=" exn)
		 (delete-directory run-dir)))
	    (if (and run-dir
		     (not (member run-dir (list "n/a" "/tmp/badname"))))
		(debug:print 0 *default-log-port* "WARNING: not removing " run-dir " as it either doesn't exist or is not a symlink")
		(debug:print 0 *default-log-port* "NOTE: the run dir for this test is undefined. Test may have already been deleted."))
	    ))
    ;; Only delete the records *after* removing the directory. If things fail we have a record 
    (case clean-mode
      ((remove-data-only)(mt:test-set-state-status-by-id (db:test-get-run_id test)(db:test-get-id test) (db:test-get-state test)(db:test-get-status test) #f))
      ((archive-remove)  (mt:test-set-state-status-by-id (db:test-get-run_id test)(db:test-get-id test) "ARCHIVED" #f #f))
      (else (rmt:delete-test-records (db:test-get-run_id test) (db:test-get-id test))))
    ;; (rmt:no-sync-del! lock-key)
    ))

(define (runs:recursive-delete-with-error-msg real-dir)
  (if (> (system (conc "rm -rf " real-dir)) 0)
      (begin
	;; FAILED, possibly due to permissions, do chmod a+rwx then try one more time
	(system (conc "chmod -R a+rwx " real-dir))
	(if (> (system (conc "rm -rf " real-dir)) 0)
	    (debug:print-error 0 *default-log-port* "There was a problem removing " real-dir " with rm -f")))))

(define (runs:safe-delete-test-dir real-dir)
  ;; first delete all sub-directories
  (directory-fold 
   (lambda (f x)
     (let ((fullname (conc real-dir "/" f)))
       (if (directory? fullname)(runs:recursive-delete-with-error-msg fullname)))
     (+ 1 x))
   0 real-dir)
  ;; then files other than *testdat.db*
  (directory-fold 
   (lambda (f x)
     (let ((fullname (conc real-dir "/" f)))
       (if (not (string-search (regexp "testdat.db") f))
	   (runs:recursive-delete-with-error-msg fullname)))
     (+ 1 x))
   0 real-dir #t)
  ;; then the entire directory
  (runs:recursive-delete-with-error-msg real-dir))


;;======================================================================
;; ideally put all this info into the db, no need to preserve it across moving homehost
;;
;; return list of
;;  ( reachable? cpuload update-time )
(define (common:get-host-info hostname)
  (let* ((loadinfo                         (rmt:get-latest-host-load hostname)) ;; if this host happens to have been recently used by a test reuse the load data
         (load                             (car loadinfo))
         (load-sample-time                 (cdr loadinfo))
         (load-sample-age                  (- (current-seconds) load-sample-time))
         (loadinfo-timeout-seconds         6) ;; this was 20 seconds, seems way too lax. Switch to 6 seconds
         (host-last-update-timeout-seconds 4)
         (host-rec (hash-table-ref/default *host-loads* hostname #f))
         )
    (cond
     ((< load-sample-age loadinfo-timeout-seconds)
      (list #t
            load-sample-time
            load))
     ((and host-rec
           (< (current-seconds) (+ (host-last-update host-rec) host-last-update-timeout-seconds)))
      (list #t
            (host-last-update host-rec)
            (host-last-cpuload host-rec )))
     ((common:unix-ping hostname)
      (list #t
            (current-seconds)
            (alist-ref 'adj-core-load (common:get-normalized-cpu-load hostname)))) ;; this is cheaper than you might think. get-normalized-cpu-load is cached for up to 5 seconds
     (else
      (list #f 0 -1) ;; bad host, don't use!
      ))))

;;======================================================================
;; see defstruct host at top of file.
;;    host: reachable last-update last-used last-cpuload
;;
(define (common:update-host-loads-table hosts-raw)
  (let* ((hosts (filter (lambda (x)
                          (string-match (regexp "^\\S+$") x))
                        hosts-raw)))
    (for-each
     (lambda (hostname)
       (let* ((rec       (let ((h (hash-table-ref/default *host-loads* hostname #f)))
                          (if h
                              h
                              (let ((h (make-host)))
                                (hash-table-set! *host-loads* hostname h)
                                h))))
              (host-info         (common:get-host-info hostname))
              (is-reachable      (car host-info))
              (last-reached-time (cadr host-info))
              (load              (caddr host-info)))
         (host-reachable-set!    rec is-reachable)
         (host-last-update-set!  rec last-reached-time)
         (host-last-cpuload-set! rec load)))
     hosts)))


;;======================================================================
;;  T E S T   L A U N C H I N G   P E R   I T E M   W I T H   H O S T   T Y P E S
;;======================================================================
;; 
;; [hosts]
;; arm cubie01 cubie02
;; x86_64 zeus xena myth01
;; allhosts #{g hosts arm} #{g hosts x86_64}
;; 
;; [host-types]
;; general #MTLOWESTLOAD #{g hosts allhosts}
;; arm     #MTLOWESTLOAD #{g hosts arm}
;; nbgeneral nbjob run JOBCOMMAND -log $MT_LINKTREE/$MT_TARGET/$MT_RUNNAME.$MT_TESTNAME-$MT_ITEM_PATH.lgo
;;
;; [host-rules]
;; # maxnload   => max normalized load
;; # maxnjobs   => max jobs per cpu
;; # maxjobrate => max jobs per second
;; general maxnload=1.1; maxnjobs=1.2; maxjobrate=0.1 
;; 
;; [launchers]
;; envsetup general
;; xor/%/n 4C16G
;; % nbgeneral
;; 
;; [jobtools]
;; # if defined and not "no" flexi-launcher will bypass "launcher" unless no match.
;; flexi-launcher yes  
;; launcher nbfake
;;
(define (common:get-launcher configdat testname itempath)
  (let ((fallback-launcher (configf:lookup configdat "jobtools" "launcher")))
    (if (and (configf:lookup configdat "jobtools" "flexi-launcher") ;; overrides launcher
	     (not (equal? (configf:lookup configdat "jobtools" "flexi-launcher") "no")))
	(let* ((launchers         (hash-table-ref/default configdat "launchers" '())))
	  (if (null? launchers)
	      fallback-launcher
	      (let loop ((hed (car launchers))
			 (tal (cdr launchers)))
		(let ((patt      (car hed))
		      (host-type (cadr hed)))
		  (if (tests:match patt testname itempath)
		      (begin
			(debug:print-info 2 *default-log-port* "Have flexi-launcher match for " testname "/" itempath " = " host-type)
			(let ((launcher (configf:lookup configdat "host-types" host-type)))
			  (if launcher
			      (let* ((launcher-parts (string-split launcher))
				     (launcher-exe   (car launcher-parts)))
				(if (equal? launcher-exe "#MTLOWESTLOAD") ;; this is our special case, we will find the lowest load and craft a nbfake commandline
				    (let host-loop ((targ-host (common:get-least-loaded-host (cdr launcher-parts) host-type configdat))
						    (count     100))
				      (if targ-host
					  (conc "remrun " targ-host)
					  (if (> count 0)
					      (begin
						(debug:print 0 *default-log-port* "INFO: Waiting for a host for host-type " host-type)
						(thread-sleep! (- 101 count))
						(host-loop (common:get-least-loaded-host (cdr launcher-parts) host-type configdat)
							   (- count 1)))
					      (begin
						(debug:print 0 *default-log-port* "FATAL: Failed to find a host from #MTLOWESTLOAD for host-type " host-type)
						(exit)))))
				    launcher))
			      (begin
				(debug:print-info 0 *default-log-port* "WARNING: no launcher found for host-type " host-type)
				(if (null? tal)
				    fallback-launcher
				    (loop (car tal)(cdr tal)))))))
		      ;; no match, try again
		      (if (null? tal)
			  fallback-launcher
			  (loop (car tal)(cdr tal))))))))
	fallback-launcher)))

;;======================================================================
;; go through the hosts from least recently used to most recently used, pick the first that meets the load criteral from the
;; [host-rules] section.
;;
(define (common:get-least-loaded-host hosts-raw host-type configdat)
  (let* ((rdat       (configf:lookup configdat "host-rules" host-type))
	 (rules      (common:val->alist (or rdat "") convert: #t))   ;; maxnload, maxnjobs, maxjobrate
	 (maxnload   (common:alist-ref/default 'maxnload rules 1.5)) ;; max normalized load
	 (maxnjobs   (common:alist-ref/default 'maxnjobs rules 1.5)) ;; max normalized number of jobs
	 (maxjobrate (common:alist-ref/default 'maxjobrate rules (/ 1 6))) ;; max rate of submitting jobs to a given host in jobs/second
	 (hosts      (filter (lambda (x)
			       (string-match (regexp "^\\S+$") x))
			     hosts-raw))
         ;; (best-host #f)
	 (get-rec    (lambda (hostname)
		       ;; (print "get-rec hostname=" hostname)
		       (let ((h (hash-table-ref/default *host-loads* hostname #f)))
			 (if h
			     h
			     (let ((h (make-host)))
			       (hash-table-set! *host-loads* hostname h)
			       h)))))
         (best-load 99999)
         (curr-time (current-seconds))
	 (get-hosts-sorted (lambda (hosts)
			     (sort hosts (lambda (a b)
					   (let ((a-rec (get-rec a))
						 (b-rec (get-rec b)))
					     ;; (print "a=" a " a-rec=" a-rec " host-last-used=" (host-last-used a-rec))
					     ;; (print "b=" b " b-rec=" b-rec " host-last-used=" (host-last-used b-rec))
					     (< (host-last-used a-rec)
						(host-last-used b-rec))))))))
    (debug:print 0 *default-log-port* "INFO: hosts-sorted=" (get-hosts-sorted hosts))
    (if (null? hosts)
	#f ;; no hosts to select from. All done and giving up now.
	(let ((hosts-sorted (get-hosts-sorted hosts)))
	  (common:update-host-loads-table hosts)
	  (let loop ((hostname  (car hosts-sorted))
		     (tal       (cdr hosts-sorted))
		     (best-host #f))
	    (let* ((rec       (get-rec hostname))
		   (reachable (host-reachable     rec))
		   (load      (host-last-cpuload  rec))
		   (last-used (host-last-used     rec))
		   (delta     (- curr-time last-used))
		   (job-rate  (if (> delta 0)
				  (/ 1 delta)
				  999)) ;; jobs per second
		   (new-best  
		    (cond
		     ((not reachable)
		      (debug:print 0 *default-log-port* "Skipping host " hostname " as it cannot be reached.")
		      best-host)
		     ((and (< load maxnload)        ;; load is acceptable
			   (< job-rate maxjobrate)) ;; job rate is acceptable
		      (set! best-load load)
		      hostname)
		     (else best-host))))
	      (debug:print 0 *default-log-port* "INFO: Trying host " hostname " with load " load ", last used " delta " seconds ago, with job-rate " job-rate " for running a test." )
	      (if new-best
		  (begin ;; found a host, return it
		    (debug:print 0 *default-log-port* "INFO: Found host: " new-best " load: " load " last-used: " delta " seconds ago, with job-rate: " job-rate)
		    (host-last-used-set! rec curr-time)
		    new-best)
		  (if (null? tal) #f (loop (car tal)(cdr tal) best-host)))))))))


;; Spec for End of test
;; At end of each test call, after marking self as COMPLETED do run-state-status-rollup
;; At transition to run COMPLETED/X do hooks
;; Definition: test_dead if event_time + duration + 1 minute? < current_time AND
;; we can prove the process is not alive (ssh host pstree -A pid)
;; if dead safe to mark the test as killed in the db
;; State/status table
;; new
;; 100% COMPLETED/ (PASS,FAIL,ABORT etc.) ==> COMPLETED / X where X is same as itemized rollup
;; > 3 RUNNING with not test_dead do nothing (run should already be RUNNING/ na
;; > 0 RUNNING and test_dead then send KILLREQ ==> COMPLETED
;; 0 RUNNING ==> this is actually the first condition, should not get here

(define (runs:end-of-run-check run-id )
  (let* ((not-completed-cnt (rmt:get-not-completed-cnt run-id))  
	 (running-cnt (rmt:get-count-tests-running-for-run-id run-id))
	 (all-test-launched (rmt:get-var (conc "lunch-complete-" run-id)))
	 (current-state (rmt:get-run-state run-id))
	 (current-status (rmt:get-run-status run-id)))
    ;;get-vars run-id to query metadata table to check if all completed. if all-test-launched = yes then only not-completed-cnt = 0 means everyting is completed if no entry found in the table do nothing 
    (debug:print 0 *default-log-port* "Running test cnt :" running-cnt)                      
    (rmt:set-state-status-and-roll-up-run  run-id current-state current-status)
    (runs:update-junit-test-reporter-xml run-id) 
    (cond 
     ((and all-test-launched (eq? not-completed-cnt 0) (equal? all-test-launched "yes" ))
      (if (and (equal? (rmt:get-var (conc "end-of-run-" run-id)) "no") (common:simple-lock (conc "endOfRun" run-id)))
	  (begin
	    (debug:print 4 *default-log-port* "look for  post hook. currseconds: " (current-seconds) " EOR " (rmt:get-var (conc "end-of-run-" run-id)))
	    (debug:print 0 *default-log-port* "End of Run Detected.")
	    (rmt:set-var (conc "end-of-run-" run-id) "yes")
					;(thread-sleep! 10)
	    (runs:run-post-hook run-id)
	    (debug:print 4 *default-log-port* "currseconds: " (current-seconds)" eor: " (rmt:get-var (conc "end-of-run-" run-id)))
	    (common:simple-unlock (conc "endOfRun" run-id)))
	  (debug:print 0 *default-log-port* "End of Run Detected but not running post hook. This should happen when eor is set to yes. This will happen only when 2 tests exit at smae time. eor= " (rmt:get-var (conc "end-of-run-" run-id)))))
     ((> running-cnt 3) 
      (debug:print 0 *default-log-port* "There are " running-cnt " tests running." ))
     ((> running-cnt 0)
      (debug:print 0 *default-log-port* "running cnt > 0 but <= 3 kill-running-tests-if-dead" )
      (let ((kill-cnt (launch:kill-tests-if-dead run-id)))
	(if (and all-test-launched  (equal? all-test-launched "yes") (eq? kill-cnt running-cnt))
	    (runs:end-of-run-check run-id)))) ;;todo
     (else  (debug:print 0 *default-log-port* "Should it get here?? May be everything is not launched yet. Running test cnt:" running-cnt " Not completed test cnt:" not-completed-cnt)
	    (let* ((not-completed-tests (rmt:get-tests-for-run run-id "%" `("NOT_STARTED" "RUNNING" "LAUNCHED" "REMOTEHOSTSTART") `() #f #f #f #f #f #f #f #f)))
	      (if (> (length not-completed-tests) 0) 
		  (let loop ((running-test (car not-completed-tests))
			     (tal    (cdr not-completed-tests)))
		    (let* ((test-name (vector-ref running-test 2))
			   (item-path (vector-ref running-test 11)))
		      (debug:print 0 *default-log-port* "test " test-name "/" item-path " not completed")
		      (if (not (null? tal))
			  (loop (car tal) (cdr tal)))))))))))

(define (runs:find-and-mark-incomplete-and-check-end-of-run run-id ovr-deadtime)
  (rmt:find-and-mark-incomplete run-id ovr-deadtime)
  (runs:end-of-run-check run-id))



(define (launch:kill-tests-if-dead run-id)
  (let* ((running-tests (rmt:get-tests-for-run run-id "%" `("RUNNING" "LAUNCHED" "REMOTEHOSTSTART") `() #f #f #f #f #f #f #f #f)))
       (let loop ((running-test (car running-tests))
			     (tal    (cdr running-tests))
			     (kill-cnt 0))
		       (let* ((test-name (vector-ref running-test 2))
                 (item-path (vector-ref running-test 11))
								 (test-id (vector-ref running-test 0))
                 (host (vector-ref running-test 6))
                 (pid  (rmt:test-get-top-process-pid run-id test-id))   
                 (event-time (vector-ref running-test 5))
                 (duration (vector-ref running-test 12))
                 (flag 0)   
                 (curr-time (current-seconds)))
       (if (and (< (+ event-time duration 600) curr-time) (not (launch:is-test-alive host pid))) ;;test has not updated duration in last 10 min then likely its not running but confirm before marking it as killed
           (begin    
			       	(debug:print 0 *default-log-port* "test " test-name "/" item-path " needs to be killed")
              (set! flag 1) 
              (rmt:set-state-status-and-roll-up-items run-id test-name item-path "KILLREQ" "n/a" #f)))
               (if (not (null? tal))
				  (loop (car tal) (cdr tal) (+ kill-cnt flag))
                 (+ kill-cnt flag))))))


(define (runs:run-post-hook run-id)
    (let* ((run-post-hook   (configf:lookup *configdat* "runs" "post-hook"))
           (existing-tests (if run-post-hook
                               (rmt:get-tests-for-run run-id "%" '() '() ;; run-id testpatt states statuses
                                                      #f #f ;; offset limit
                                                      #f ;; not-in
                                                      #f ;; sort-by
                                                      #f ;; sort-order
                                                      #f ;; get full data (not 'shortlist)
                                                      0 ;; (runs:gendat-inc-results-last-update *runs:general-data*) ;; last update time
                                                      'dashboard)
                               '()))
           (log-dir         (conc *toppath* "/logs"))
           (log-file        (conc "post-hook-" (string-translate (getenv "MT_TARGET") "/" "-") "-" (getenv "MT_RUNNAME") ".log"))
           (full-log-fname  (conc log-dir "/" log-file)))
      (if run-post-hook
          ;; (if (null? existing-tests)
          ;;    (debug:print 0 *default-log-port* "Skipping post-hook call \"" run-post-hook "\" as there are existing tests for this run.")))))
	  (let* ((use-log-dir (if (not (directory-exists? log-dir))
				  (handle-exceptions
				      exn
				      (begin
					(debug:print 0 *default-log-port* "WARNING: Failed to create " log-dir ", exn=" exn)
					#f)
				    (create-directory log-dir #t)
				    #t)
				  #t))
		 (start-time   (current-seconds))
		 (actual-logf  (if use-log-dir full-log-fname log-file)))
	    (handle-exceptions
		exn
		(begin
		  (print-call-chain *default-log-port*)
		  (debug:print 0 *default-log-port* "Message: " ((condition-property-accessor 'exn 'message) exn) ", exn=" exn)
		  (debug:print 0 *default-log-port* "ERROR: failed to run post-hook " run-post-hook ", check the log " log-file))
	      (debug:print-info 0 *default-log-port* "running run-post-hook: \"" run-post-hook "\", log is " actual-logf)
	      (system (conc run-post-hook " >> " actual-logf " 2>&1"))
	      (debug:print-info 0 *default-log-port* "post-hook \"" run-post-hook "\" took " (- (current-seconds) start-time) " seconds to run."))))))


(define (runs:rerun-hook test-id new-test-path testdat rerunlst)
    (let* ((rerun-hook   (configf:lookup *configdat* "runs" "rerun-hook"))
           (log-dir         (conc *toppath* "/reruns/logs"))
           (target (getenv "MT_TARGET"))
           (runname (common:args-get-runname))
           (rundir (db:test-get-rundir testdat))
           (tarfiledir (conc *toppath* "/reruns"))
           (status (db:test-get-status testdat))
           (comment (conc "\"" (db:test-get-comment testdat) "\"" ))
           (testname (db:test-get-testname testdat))
           (itempath (db:test-get-item-path testdat))
           (file-body (conc status "-" (string-translate target "/" "-") "-" runname "." testname (if (not (string=? itempath "")) (conc  "." (string-translate itempath "/" "-")) "")))
           (log-file (conc file-body ".log"))
           ;; (log-file        (conc status "-" (string-translate target "/" "-") "-" runname "." testname (if (not (string=? itempath "")) (conc  "." (string-translate itempath "/" "-")) "") ".log"))
           (full-log-fname  (conc log-dir "/" log-file))
           (tarfilename (conc file-body ".tar"))
           ;; (tarfilename (conc status "." (string-translate target "/" "-") "." runname "." testname (if (not (string=? itempath "")) (conc  "." (string-translate itempath "/" "-")) "") ".tar"))
           )
      (if rerun-hook
	  (let* ((use-log-dir (if (not (directory-exists? log-dir))
				  (handle-exceptions
				      exn
				      (begin
					(debug:print 0 *default-log-port* "WARNING: Failed to create " log-dir ", exn=" exn)
					#f)
				    (create-directory log-dir #t)
				    #t)
				  #t))
		 (start-time   (current-seconds))
		 (actual-logf  (if use-log-dir full-log-fname log-file))
                 (sys-call-text (conc rerun-hook " " tarfilename " " rundir " " actual-logf " " runname " " tarfiledir " " status " " target " " comment " " testname " " itempath " >> " actual-logf " 2>&1"))
                 )
	    (debug:print 2 *default-log-port* "Found rerun-hook in config:" rerun-hook)
	    (handle-exceptions
		exn
		(begin
		  (print-call-chain *default-log-port*)
		  (debug:print 0 *default-log-port* "Message: " ((condition-property-accessor 'exn 'message) exn) ", exn=" exn)
		  (debug:print 0 *default-log-port* "ERROR: failed to run rerun-hook " rerun-hook ", check the log " log-file))
	      (debug:print-info 0 *default-log-port* "running rerun-hook: \"" rerun-hook "\", log is " actual-logf)
              ;; call the hook
              (debug:print-info 0 *default-log-port* "Calling rerun-hook for " test-id new-test-path testdat rerunlst)
              (debug:print-info 0 *default-log-port* "rerun hook: " rerun-hook)
              (debug:print-info 0 *default-log-port* "tarfilename: " tarfilename)
              (debug:print-info 0 *default-log-port* "rundir: " rundir)
              (debug:print-info 0 *default-log-port* "actual-logf: " actual-logf)
              (debug:print-info 0 *default-log-port* "runname: " runname)
              (debug:print-info 0 *default-log-port* "sys-call-text: " sys-call-text)
	      (system sys-call-text)
	      (debug:print-info 0 *default-log-port* "rerun-hook \"" rerun-hook "\" took " (- (current-seconds) start-time) " seconds to run."))))))


(define (runs:update-junit-test-reporter-xml run-id)
  (let*	((junit-test-reporter	(configf:lookup *configdat* "runs" "junit-test-reporter-xml"))
	 (junit-test-report-dir  (configf:lookup *configdat* "runs" "junit-test-report-dir"))
	 (xml-dir		(if (and junit-test-reporter (equal? junit-test-reporter "yes" ))
				    (if junit-test-report-dir
					junit-test-report-dir
					(conc (getenv "MT_LINKTREE") "/" (getenv "MT_TARGET") "/" (getenv "MT_RUNNAME")))
				    #f))
	 (xml-ts-name		(if xml-dir
				    (conc (getenv "MT_TESTSUITENAME")"."(string-translate (getenv "MT_TARGET") "/" ".") "." (getenv "MT_RUNNAME"))
				    #f))
         (keyname               (if xml-ts-name (common:get-signature xml-ts-name) #f))
	 (xml-path		(if xml-dir
				    (conc xml-dir "/" keyname ".xml")
				    #f))

	 (test-data		(if xml-dir
				    (rmt:get-tests-for-run run-id "%" '() '() ;; run-id testpatt states statuses
							   #f #f ;; offset limit
							   #f ;; not-in
							   #f ;; sort-by
							   #f ;; sort-order
							   #f ;; get full data (not 'shortlist)
							   0 ;; (runs:gendat-inc-results-last-update *runs:general-data*) ;; last update time
							   #f)
				    '()))
	 (tests-count		(if xml-dir (length test-data) #f)))
    (if (and junit-test-reporter (equal? junit-test-reporter "yes" ))
	(begin
					;((sxml-modify! `("testsuite" insert-into (@ (name ,xml-ts-name) (tests ,tests-count)))) doc)

	  (let loop	((test		(car test-data))
			 (tail		(cdr test-data))
			 (doc		doc-template)
			 (fail-cnt	0)
			 (error-cnt	0))
	    (let*	((test-name	(vector-ref test 2))
			 (test-itempath	(vector-ref test 11))
			 (tc-name	(conc test-name (if (and test-itempath (not (equal? test-itempath ""))) (conc "." (string-translate test-itempath "/" "." )) "")))
			 (test-state	(vector-ref test 3))
			 (comment	(vector-ref test 14))   
			 (test-status	(vector-ref test 4))
			 (exc-msg	(conc "No bucket for State " test-state " Status " test-status))
			 (new-doc	(cond 
					 ((member test-state (list "RUNNING" ))
					  ((sxml-modify `("testsuite" insert-into (testcase (@ (name ,tc-name)) (inProgress)))) doc))
					 ((member test-state (list "LAUNCHED" "REMOTEHOSTSTART"  "NOT_STARTED"))
					  ((sxml-modify `("testsuite" insert-into (testcase (@ (name ,tc-name)) (inQueue)))) doc))
					 ((member test-status (list "PASS" "WARN" "WAIVED"))
					  ((sxml-modify `("testsuite" insert-into (testcase (@ (name ,tc-name))))) doc))
					 ((member test-status (list "FAIL" "CHECK"))
					  ((sxml-modify `("testsuite" insert-into (testcase (@ (name ,tc-name)) (failure (@ (message ,comment) (type "failure")))))) doc)) 
					 ((member test-status (list "DEAD" "KILLED" "ABORT" "PREQ_FAIL" "PREQ_DISCARDED"))
					  ((sxml-modify `("testsuite" insert-into (testcase (@ (name ,tc-name)) (failure (@ (message ,comment) (type "error")))))) doc))
					 ((member test-status (list "SKIP"))
					  ((sxml-modify `("testsuite" insert-into (testcase (@ (name ,tc-name)) (skipped (@ (type "skipped")))))) doc))
					 (else 
					  (debug:print 0 *default-log-port* (conc "What do I do with State " test-state " Status " test-status))
					  ((sxml-modify `("testsuite" insert-into (testcase (@ (name ,tc-name)) (failure (@ (message ,exc-msg) (type "error")))))) doc))))
			 (new-error-cnt	(if (member test-status (list "DEAD" "KILLED" "ABORT" "PREQ_FAIL" "PREQ_DISCARDED"))
					    (+ error-cnt 1) 
					    error-cnt))
			 (new-fail-cnt	(if (member test-status (list "FAIL" "CHECK"))
					    (+ fail-cnt 1)
					    fail-cnt)))
 	      (if (null? tail)
		    (let* ((final-doc ((sxml-modify `("testsuite" insert-into (@ (name ,xml-ts-name) (tests ,tests-count) (errors ,error-cnt) (failures ,fail-cnt)))) new-doc)))
		    (debug:print 0 *default-log-port* "modify attrib error=" error-cnt " fail= " fail-cnt)
		    (handle-exceptions
		     exn
		     (let*	((msg	((condition-property-accessor 'exn 'message) exn)))
		       (debug:print 0 *default-log-port* (conc "WARNING: Failed to update file" xml-path". Message:" msg ", exn=" exn)))
		     		   
		     (if (not (file-exists? xml-dir)) 
			 (create-directory xml-dir #t))
                     (if (not (rmt:no-sync-get/default keyname #f)) 
                       (begin
			 (rmt:no-sync-set  keyname "on")
			 (debug:print 0 *default-log-port* "creating xml at " xml-path)
		         (with-output-to-file xml-path
		         (lambda ()
			   (print (sxml-serializer#serialize-sxml final-doc  ns-prefixes: (list (cons 'gnm "http://foo"))))))
                           (rmt:no-sync-del! keyname))
                          (debug:print 0 *default-log-port*  "Could not get the lock. Skip writing the xml file."))))
		  (loop (car tail) (cdr tail) new-doc new-fail-cnt new-error-cnt))))))))
	 
(define doc-template 
  '(*TOP*
    (*PI* xml "version='1.0'")
    (testsuite)))

(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))

;; 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    (common:get-fields *configdat*) #;(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)
					    ""))))))

  
;; cache the runconfigs in $MT_LINKTREE/$MT_TARGET/$MT_RUNNAME/.runconfig
;;
(define (full-runconfigs-read)
  ;; in the envprocessing branch the below code replaces the further below code
  ;;  (if (eq? *configstatus* 'fulldata)
  ;;      *runconfigdat*
  ;;      (begin
  ;;	(launch:setup)
  ;;	*runconfigdat*)))
  (let* ((rundir (if (and (getenv "MT_LINKTREE")(getenv "MT_TARGET")(getenv "MT_RUNNAME"))
		     (conc (getenv "MT_LINKTREE") "/" (getenv "MT_TARGET") "/" (getenv "MT_RUNNAME"))
		     #f))
	 (cfgf   (if rundir (conc rundir "/.runconfig." megatest-version "-" megatest-fossil-hash) #f)))
    (if (and cfgf
	     (common:file-exists? cfgf)
	     (file-writable? cfgf)
	     (common:use-cache?))
	(configf:read-alist cfgf)
	(let* ((keys   (common:get-fields cfgf)) ;; (rmt:get-keys))
	       (target (common:args-get-target))
	       (key-vals (if target (keys:target->keyval keys target) #f))
	       (sections (if target (list "default" target) #f))
	       (data     (begin
			   (setenv "MT_RUN_AREA_HOME" *toppath*)
			   (if key-vals
			       (for-each (lambda (kt)
					   (setenv (car kt) (cadr kt)))
					 key-vals))
			   ;; (read-config (conc *toppath* "/runconfigs.config") #f #t sections: sections))))
                           (runconfig:read (conc *toppath* "/runconfigs.config") target #f))))
	  (if (and rundir ;; have all needed variabless
		   (directory-exists? rundir)
		   (file-writable? rundir))
	      (begin
                (if (not (common:in-running-test?))
                    (configf:write-alist data cfgf))
		;; force re-read of megatest.config - this resolves circular references between megatest.config
		(launch:setup force-reread: #t)
		;; (launch:cache-config) ;; there are two independent config cache locations, turning this one off for now. MRW.
		)) ;; we can safely cache megatest.config since we have a valid runconfig
	  data))))
     


)