Megatest

                    (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))))
     
;;======================================================================
;; TODO: for multiple areas, we will have multiple watchdogs; and multiple threads to manage
(define (common:watchdog)
  (debug:print-info 13 *default-log-port* "common:watchdog entered.")
  (if (launch:setup)
      (if (common:on-homehost?)
	  (let ((dbstruct (db:setup #t)))
	    (debug:print-info 13 *default-log-port* "after db:setup with dbstruct=" dbstruct)
	    (cond
	     ((dbr:dbstruct-read-only dbstruct)
	      (debug:print-info 13 *default-log-port* "loading read-only watchdog")
	      (common:readonly-watchdog dbstruct))
	     (else
	      (debug:print-info 13 *default-log-port* "loading writable-watchdog.")
              (let* ((syncer (or (configf:lookup *configdat* "server" "sync-method") "brute-force-sync")))
                (cond
                 ((equal? syncer "brute-force-sync")
                  (server:writable-watchdog-bruteforce dbstruct))
                 ((equal? syncer "delta-sync")
                  (server:writable-watchdog-deltasync dbstruct))
                 (else
                  (debug:print-error 0 *default-log-port* "Unknown server/sync-method specified ("syncer") - valid values are brute-force-sync and delta-sync.")
                  (exit 1)))
                ;;(debug:print 1 *default-log-port* "INFO: ["(common:human-time)"] Syncer started (method="syncer")")
                )))
	    (debug:print-info 13 *default-log-port* "watchdog done."))
	  (debug:print-info 13 *default-log-port* "no need for watchdog on non-homehost"))))

;;======================================================================
;; currently the primary job of the watchdog is to run the sync back to megatest.db from the db in /tmp
;; if we are on the homehost and we are a server (by definition we are on the homehost if we are a server)
;;
(define (common:readonly-watchdog dbstruct)
  (let ((just-testing 0.0501))
    (thread-sleep! just-testing)) ;; (/ 1 20)) ;; 0.051) ;; delay for startup
  (debug:print-info 13 *default-log-port* "common:readonly-watchdog entered.")
  ;; sync megatest.db to /tmp/.../megatst.db
  (let* ((sync-cool-off-duration   3)
        (golden-mtdb     (dbr:dbstruct-mtdb dbstruct))
        (golden-mtpath   (db:dbdat-get-path golden-mtdb))
        (tmp-mtdb        (dbr:dbstruct-tmpdb dbstruct))
        (tmp-mtpath      (db:dbdat-get-path tmp-mtdb)))
    (debug:print-info 0 *default-log-port* "Read-only periodic sync thread started.")
    (let loop ((last-sync-time 0))
      (debug:print-info 13 *default-log-port* "loop top tmp-mtpath="tmp-mtpath" golden-mtpath="golden-mtpath)
      (let* ((duration-since-last-sync (- (current-seconds) last-sync-time)))
        (debug:print-info 13 *default-log-port* "duration-since-last-sync="duration-since-last-sync)
        (if (and (not (bdat-time-to-exit *bdat*))
                 (< duration-since-last-sync sync-cool-off-duration))
            (thread-sleep! (- sync-cool-off-duration duration-since-last-sync)))
        (if (not (bdat-time-to-exit *bdat*))
            (let ((golden-mtdb-mtime (file-modification-time golden-mtpath))
                  (tmp-mtdb-mtime    (file-modification-time tmp-mtpath)))
	      (if (> golden-mtdb-mtime tmp-mtdb-mtime)
		  (if (< golden-mtdb-mtime (- (current-seconds) 3)) ;; file has NOT been touched in past three seconds, this way multiple servers won't fight to sync back
		      (let ((res (db:multi-db-sync dbstruct 'old2new)))
			(debug:print-info 13 *default-log-port* "rosync called, " res " records transferred."))))
              (loop (current-seconds)))
            #t)))
    (debug:print-info 0 *default-log-port* "Exiting readonly-watchdog timer, (bdat-time-to-exit *bdat*) = " (bdat-time-to-exit *bdat*)" pid="(current-process-id)" mtpath="golden-mtpath)))


;; TODO: for multiple areas, we will have multiple watchdogs; and multiple threads to manage
(define (init-watchdog)
  (set! (bdat-watchdog-set! *bdat*)
	(make-thread
	 (lambda ()
	   (handle-exceptions
	    exn
	    (begin
	      (print-call-chain)
	      (print " message: " ((condition-property-accessor 'exn 'message) exn) ", exn=" exn))
	    (common:watchdog)))
	 "Watchdog thread"))
  (start-watchdog))

(define (start-watchdog)
  ;;(if (not (args:get-arg "-server"))
  ;;    (thread-start! *watchdog*)) ;; if starting a server; wait till we get to running state before kicking off watchdog
  (let* ((no-watchdog-args
	  '("-list-runs"
	    "-testdata-csv"
	    "-list-servers"
	    "-server"
	    "-adjutant"
	    "-list-disks"
	    "-list-targets"
	    "-show-runconfig"
	    ;;"-list-db-targets"
	    "-show-runconfig"
	    "-show-config"
	    "-show-cmdinfo"
	    "-cleanup-db"
	    ))
	 (no-watchdog-argvals (list '("-archive" . "replicate-db")))
	 (start-watchdog-specail-arg-val (let loop ((hed  (car no-watchdog-argvals))
						    (tail (cdr   no-watchdog-argvals)))
					   ;; (print "hed" hed " arg " (args:get-arg (car hed)) " val:" (cdr hed)  " eql" (equal? (args:get-arg (car hed)) (cdr hed)))  
					   (if (equal? (args:get-arg (car hed)) (cdr hed))
					       #f
					       (if (null? tail)
						   #t
						   (loop (car tail) (cdr tail))))))      
	 (no-watchdog-args-vals (filter (lambda (x) x)
					(map args:get-arg no-watchdog-args)))
	 (start-watchdog (and (null? no-watchdog-args-vals) start-watchdog-specail-arg-val)))
					;(print  "no-watchdog-args="no-watchdog-args "no-watchdog-args-vals="no-watchdog-args-vals " start-watchdog-specail-arg-val:" start-watchdog-specail-arg-val " start-watchdog:" start-watchdog) 
    (if start-watchdog
	(thread-start! (bdat-watchdog *bdat*)))))
     
(define (server:writable-watchdog-deltasync dbstruct)
  (thread-sleep! 0.054) ;; delay for startup
  (let ((legacy-sync  (common:run-sync?))
        (sync-stale-seconds (configf:lookup-number *configdat* "server" "sync-stale-seconds" default: 300))
	(debug-mode   (debug:debug-mode 1))
	(last-time    (current-seconds))
	(no-sync-db   (db:open-no-sync-db))
	(stmt-cache   (dbr:dbstruct-stmt-cache dbstruct))
        (sync-duration 0) ;; run time of the sync in milliseconds
        )
    (set! *no-sync-db* no-sync-db) ;; make the no sync db available to api calls
    (debug:print-info 2 *default-log-port* "Periodic sync thread started.")
    (debug:print-info 3 *default-log-port* "watchdog starting. legacy-sync is " legacy-sync" pid="(current-process-id)  );;  " this-wd-num="this-wd-num)
    (if (and legacy-sync (not (bdat-time-to-exit *bdat*)))
	(let* (;;(dbstruct (db:setup))
	       (mtdb       (dbr:dbstruct-mtdb dbstruct))
	       (mtpath     (db:dbdat-get-path mtdb))
	       (tmp-area   (common:get-db-tmp-area))
	       (start-file (conc tmp-area "/.start-sync"))
	       (end-file   (conc tmp-area "/.end-sync")))
	  (debug:print-info 0 *default-log-port* "Server running, periodic sync started.")
	  (let loop ()
	    ;; sync for filesystem local db writes
	    ;;
	    (mutex-lock! *db-multi-sync-mutex*)
	    (let* ((need-sync        (>= *db-last-access* *db-last-sync*)) ;; no sync since last write
		   (sync-in-progress *db-sync-in-progress*)
                   (min-intersync-delay (configf:lookup-number *configdat* "server" "minimum-intersync-delay" default: 5))
		   (should-sync      (and (not (bdat-time-to-exit *bdat*))
                                          (> (- (current-seconds) *db-last-sync*) min-intersync-delay))) ;; sync every five seconds minimum, deprecated logic, can probably be removed
		   (start-time       (current-seconds))
                   (cpu-load-adj     (alist-ref 'adj-proc-load (common:get-normalized-cpu-load #f)))
		   (mt-mod-time      (file-modification-time mtpath))
		   (last-sync-start  (if (common:file-exists? start-file)
					 (file-modification-time start-file)
					 0))
		   (last-sync-end    (if (common:file-exists? end-file)
					 (file-modification-time end-file)
					 10))
                   (sync-period      (+ 3 (* cpu-load-adj 30))) ;; as adjusted load increases increase the sync period
		   (recently-synced  (and (< (- start-time mt-mod-time) sync-period) ;; not useful if sync didn't modify megatest.db!
					  (< mt-mod-time last-sync-start)))
		   (sync-done        (<= last-sync-start last-sync-end))
                   (sync-stale       (> start-time (+ last-sync-start sync-stale-seconds)))
		   (will-sync        (and (not (bdat-time-to-exit *bdat*))       ;; do not start a sync if we are in the process of exiting
                                          (or need-sync should-sync)
					  (or sync-done sync-stale)
					  (not sync-in-progress)
					  (not recently-synced))))
              (debug:print-info 13 *default-log-port* "WD writable-watchdog top of loop.  need-sync="need-sync" sync-in-progress=" sync-in-progress
				" should-sync="should-sync" start-time="start-time" mt-mod-time="mt-mod-time" recently-synced="recently-synced" will-sync="will-sync
				" sync-done=" sync-done " sync-period=" sync-period)
              (if (and (> sync-period 5)
                       (common:low-noise-print 30 "sync-period"))
                  (debug:print-info 0 *default-log-port* "Increased sync period due to long sync times, sync took: " sync-period " seconds."))
	      ;; (if recently-synced (debug:print-info 0 *default-log-port* "Skipping sync due to recently-synced flag=" recently-synced))
	      ;; (debug:print-info 0 *default-log-port* "need-sync: " need-sync " sync-in-progress: " sync-in-progress " should-sync: " should-sync " will-sync: " will-sync)
	      (if will-sync (set! *db-sync-in-progress* #t))
	      (mutex-unlock! *db-multi-sync-mutex*)
	      (if will-sync
                  (let (;; (max-sync-duration  (configf:lookup-number *configdat* "server" "max-sync-duration")) ;; KEEPING THIS AVAILABLE BUT SHOULD NOT USE, I'M PRETTY SURE IT DOES NOT WORK!
                        (sync-start         (current-milliseconds)))
		    (with-output-to-file start-file (lambda ()(print (current-process-id))))
		    
		    ;; put lock here
		    
                    ;; (if (or (not max-sync-duration)
                    ;;        (< sync-duration max-sync-duration)) ;; NOTE: db:sync-to-megatest.db keeps track of time of last sync and syncs incrementally
                        (let ((res        (db:sync-to-megatest.db dbstruct no-sync-db: no-sync-db))) ;; did we sync any data? If so need to set the db touched flag to keep the server alive
                          (set! sync-duration (- (current-milliseconds) sync-start))
                          (if (> res 0) ;; some records were transferred, keep the db alive
                              (begin
                                (mutex-lock! *heartbeat-mutex*)
                                (set! *db-last-access* (current-seconds))
                                (mutex-unlock! *heartbeat-mutex*)
                                (debug:print-info 0 *default-log-port* "sync called, " res " records transferred."))
                              (debug:print-info 2 *default-log-port* "sync called but zero records transferred")))))
;;                         ;; TODO: factor this next routine out into a function
;;                         (with-input-from-pipe ;; this should not block other threads but need to verify this
;;                          (conc "megatest -sync-to-megatest.db -m testsuite:" (common:get-area-name) ":" *toppath*)
;;                          (lambda ()
;;                            (let loop ((inl (read-line))
;;                                       (res #f))
;;                              (if (eof-object? inl)
;;                                  (begin
;;                                    (set! sync-duration (- (current-milliseconds) sync-start))
;;                                    (cond
;;                                     ((not res)
;;                                      (debug:print 0 *default-log-port* "ERROR: sync from /tmp db to megatest.db appears to have failed. Recommended that you stop your runs and run \"megatest -cleanup-db\""))
;;                                     ((> res 0)
;;                                      (mutex-lock! *heartbeat-mutex*)
;;                                      (set! *db-last-access* (current-seconds))
;;                                      (mutex-unlock! *heartbeat-mutex*))))
;;                                  (let ((num-synced (let ((matches (string-match "^Synced (\\d+).*$" inl)))
;;                                                      (if matches
;;                                                          (string->number (cadr matches))
;;                                                          #f))))
;;                                    (loop (read-line)
;;                                          (or num-synced res))))))))))
	      (if will-sync
		  (begin
		    (mutex-lock! *db-multi-sync-mutex*)
		    (set! *db-sync-in-progress* #f)
		    (set! *db-last-sync* start-time)
		    (with-output-to-file end-file (lambda ()(print (current-process-id))))

		    ;; release lock here

		    (mutex-unlock! *db-multi-sync-mutex*)))
	      (if (and debug-mode
		       (> (- start-time last-time) 60))
		  (begin
		    (set! last-time start-time)
		    (debug:print-info 4 *default-log-port* "timestamp -> " (seconds->time-string (current-seconds)) ", time since start -> " (seconds->hr-min-sec (- (current-seconds) *time-zero*))))))
	    
	    ;; keep going unless time to exit
	    ;;
	    (if (not (bdat-time-to-exit *bdat*))
		(let delay-loop ((count 0))
                  ;;(debug:print-info 13 *default-log-port* "delay-loop top; count="count" pid="(current-process-id)" this-wd-num="this-wd-num" *time-to-exit*="(bdat-time-to-exit *bdat*))
                                                            
		  (if (and (not (bdat-time-to-exit *bdat*))
			   (< count 6)) ;; was 11, changing to 4. 
		      (begin
			(thread-sleep! 1)
			(delay-loop (+ count 1))))
		  (if (not (bdat-time-to-exit *bdat*)) (loop))))
	    ;; time to exit, close the no-sync db here
	    (db:no-sync-close-db no-sync-db stmt-cache)
	    (if (common:low-noise-print 30)
		(debug:print-info 0 *default-log-port* "Exiting watchdog timer, *time-to-exit* = " (bdat-time-to-exit *bdat*)" pid="(current-process-id) ))))))) ;;" this-wd-num="this-wd-num)))))))

(define (server:writable-watchdog-bruteforce dbstruct)
  (thread-sleep! 1) ;; delay for startup
  (let* ((do-a-sync  (server:get-bruteforce-syncer dbstruct))
         (final-sync (server:get-bruteforce-syncer dbstruct fork-to-background: #t persist-until-sync: #t)))
    (when (and (not (args:get-arg "-sync-to-megatest.db")) ;; conditions under which we do not run the sync
	       (args:get-arg "-server"))
      
      (let loop ()
	(do-a-sync)
        (if (not (bdat-time-to-exit *bdat*)) (loop))) ;; keep going unless time to exit

      ;; time to exit, close the no-sync db here
      (final-sync)

      (if (common:low-noise-print 30)
	  (debug:print-info 0 *default-log-port* "Exiting watchdog timer, *time-to-exit* = "(bdat-time-to-exit *bdat*)" pid="(current-process-id)
			    )))))

;; moving this here as it needs access to db and cannot be in common.
;;

(define (server:get-bruteforce-syncer dbstruct #!key (fork-to-background #f) (persist-until-sync #f))
  (let* ((sqlite-exe   (or (get-environment-variable "MT_SQLITE3_EXE"))) ;; defined in cfg.sh
         (sync-log     (or (args:get-arg "-sync-log") (conc *toppath* "/logs/sync-" (current-process-id) "-" (get-host-name) ".log")))
	 (tmp-area     (common:get-db-tmp-area))
	 (tmp-db       (conc tmp-area "/megatest.db"))
	 (staging-file (conc *toppath* "/.megatest.db"))
	 (mtdbfile     (conc *toppath* "/megatest.db"))
	 (lockfile     (common:get-sync-lock-filepath))
         (sync-cmd-core     (conc sqlite-exe" " tmp-db " .dump | "sqlite-exe" " staging-file "&>"sync-log))
         (sync-cmd     (if fork-to-background 
                           (conc "/usr/bin/env NBFAKE_LOG="*toppath*"/logs/last-server-sync-"(current-process-id)".log nbfake \""sync-cmd-core" && /bin/mv -f " staging-file " " mtdbfile" \"")
                           sync-cmd-core))
         (default-min-intersync-delay 2)
	 (min-intersync-delay (configf:lookup-number *configdat* "server" "minimum-intersync-delay" default: default-min-intersync-delay))
         (default-duty-cycle 0.1)
         (duty-cycle   (configf:lookup-number *configdat* "server" "sync-duty-cycle" default: default-duty-cycle))
         (last-sync-seconds 10) ;; we will adjust this to a measurement and delay last-sync-seconds * (1 - duty-cycle)
         (calculate-off-time (lambda (work-duration duty-cycle)
                                  (* (/ (- 1 duty-cycle) duty-cycle) last-sync-seconds)))
         (off-time min-intersync-delay) ;; adjusted in closure below.
         (do-a-sync
          (lambda ()
            ;; (BB> "Start do-a-sync with fork-to-background="fork-to-background" persist-until-sync="persist-until-sync)
            (let* ((finalres
                    (let retry-loop ((num-tries 0))
                         (if (common:simple-file-lock lockfile)
	                     (begin
                               (cond
                                ((not (or fork-to-background persist-until-sync))
                                 (debug:print 0 *default-log-port* "INFO: syncer thread sleeping for max of (server.minimum-intersync-delay="min-intersync-delay
                                              " , off-time="off-time" seconds ]")
                                 (thread-sleep! (max off-time min-intersync-delay)))
                                (else
                                 (debug:print 0 *default-log-port* "INFO: syncer thread NOT sleeping ; maybe time-to-exit...")))

                               (if (not (configf:lookup *configdat* "server" "disable-db-snapshot"))
                                   (common:snapshot-file mtdbfile subdir: ".db-snapshot"))
		               (delete-file* staging-file)
		               (let* ((start-time (current-milliseconds))
                                      (res (system sync-cmd))
                                      (dbbackupfile (conc mtdbfile ".backup"))
                                      (res2 
                                       (cond
                                        ((eq? 0 res )
                                         (handle-exceptions
                                            exn
                                            #f
                                         (if (file-exists? dbbackupfile)
		                           (delete-file* dbbackupfile)
                                         )
                                         (if (eq? 0 (file-size sync-log))
                                             (delete-file* sync-log))
		                         (system (conc "/bin/mv " staging-file " " mtdbfile))
                                         
                                         (set! last-sync-seconds (/ (- (current-milliseconds) start-time) 1000))
                                         (set! off-time (calculate-off-time
                                                         last-sync-seconds
                                                         (cond
                                                          ((and (number? duty-cycle) (> duty-cycle 0) (< duty-cycle 1))
                                                           duty-cycle)
                                                          (else
                                                           (debug:print 0 *default-log-port* "WARNING: ["(common:human-time)"] server.sync-duty-cycle is invalid.  Should be a number between 0 and 1, but "duty-cycle" was specified.  Using default value: "default-duty-cycle)
                                                           default-duty-cycle))))
                                         
                                         (debug:print 1 *default-log-port* "INFO: ["(common:human-time)"] pid="(current-process-id)" SYNC took "last-sync-seconds" sec")
                                         (debug:print 1 *default-log-port* "INFO: ["(common:human-time)"] pid="(current-process-id)" SYNC took "last-sync-seconds" sec ; with duty-cycle of "duty-cycle" off time is now "off-time)
                                         'sync-completed))
                                        (else
                                         (system (conc "/bin/cp "sync-log" "sync-log".fail"))
                                         (debug:print 0 *default-log-port* "ERROR: ["(common:human-time)"] Sync failed. See log at "sync-log".fail")
                                         (if (file-exists? (conc mtdbfile ".backup"))
                                             (system (conc "/bin/cp "mtdbfile ".backup " mtdbfile)))
                                         #f))))
                                 (common:simple-file-release-lock lockfile)
                                 ;; (BB> "released lockfile: " lockfile)
                                 ;; (when (common:file-exists? lockfile)
                                 ;;   (BB> "DID NOT ACTUALLY RELEASE LOCKFILE"))
                                 res2) ;; end let
                               );; end begin
                             ;; else
                             (cond
                              (persist-until-sync
                               (thread-sleep! 1)
                               (debug:print 1 *default-log-port* "INFO: ["(common:human-time)"] pid="(current-process-id)" other SYNC in progress; we're in a fork-to-background so we need to succeed.  Let's wait a jiffy and and try again. num-tries="num-tries" (waiting for lockfile="lockfile" to disappear)")
                               (retry-loop (add1 num-tries)))
                              (else
                               (thread-sleep! (max off-time (+ last-sync-seconds min-intersync-delay)))
                               (debug:print 1 *default-log-port* "INFO: ["(common:human-time)"] pid="(current-process-id)" other SYNC in progress; not syncing.")
                               'parallel-sync-in-progress))
                             ) ;; end if got lockfile
                         )
                    ))
              ;; (BB> "End do-a-sync with fork-to-background="fork-to-background" persist-until-sync="persist-until-sync" and result="finalres)
              finalres)
            ) ;; end lambda
          ))
    do-a-sync))


)