Megatest

(use srfi-1 posix regex regex-case srfi-69 hostinfo md5 message-digest
     directory-utils posix-extras matchable utils)

(use spiffy uri-common intarweb http-client spiffy-request-vars)

(declare (unit server))



(declare (uses common))
(declare (uses db))
(declare (uses tasks)) ;; tasks are where stuff is maintained about what is running.
;; (declare (uses synchash))
(declare (uses http-transport))
;;(declare (uses rpc-transport))
(declare (uses launch))
;; (declare (uses daemon))



(include "common_records.scm")
(include "db_records.scm")

(define (server:make-server-url hostport)
  (if (not hostport)
      #f

		  (condition-case
		   (create-directory (conc areapath "/logs") #t)
		   (exn (i/o file)(debug:print 0 *default-log-port* "ERROR: Cannot create directory at " (conc areapath "/logs")))
		   (exn ()(debug:print 0 *default-log-port* "ERROR: Unknown error attemtping to get server list. exn=" exn)))
		  (directory-exists? (conc areapath "/logs")))
		'()))

        ;; Get the list of server logs. First remove logs for servers that have exited.
	(let* (
               ;; For some reason, when I uncomment the below line, ext-tests sometimes starts 1000's of servers.
               ;; (exiting-servers (system (conc "bash -c 'rm -f `grep -il exiting " areapath "/logs/server-*-*.log 2> /dev/null`'")))









               (server-logs   (glob (conc areapath "/logs/server-*-*.log")))




	       (num-serv-logs (length server-logs)))
	  (if (or (null? server-logs) (= num-serv-logs 0))
              (let ()
                 (debug:print 2  *default-log-port* "There are no servers running at " (common:human-time))
	         '()
              )
	      (let loop ((hed  (string-chomp (car server-logs)))
			 (tal  (cdr server-logs))
			 (res '()))
		(let* ((mod-time  (handle-exceptions
				   exn
				   (begin
				     (debug:print 0 *default-log-port* "failed to get modification time on " hed ", exn=" exn)
				     (current-seconds)) ;; 0
				   (file-modification-time hed))) ;; default to *very* old so log gets ignored if deleted
		       (down-time (- (current-seconds) mod-time))
		       (serv-dat  (if (or (< num-serv-logs 10)
				  	  (< down-time 900)) ;; day-seconds))
				      (server:logf-get-start-info hed)
				      '())) ;; don't waste time processing server files not touched in the 15 minutes if there are more than ten servers to look at

	 ;;; THIS INTERACTS WITH [server] timeout. Suggest using 0.1 or above for timeout (6 seconds)
	 (idletime    (configf:lookup-number *configdat* "server" "idletime" default: 4))
	 (server-key (conc (get-host-name) "-" (current-process-id))))
    (if (file-exists? start-flag)
	(let* ((fmodtime (file-modification-time start-flag))
	       (delta    (- (current-seconds) fmodtime))
	       (old-enough   (> delta idletime))
               (new-server-key "")
              )

          ;; write start-flag file, wait 0.25s, then if previously the start-flag file was older than <idletime> seconds, and the new file still has the same server key as you just wrote, return #t.
	  ;; the intention is to make sure nfs can read the file we just wrote, and make sure it was written by us, and not another process.
           (if (and old-enough
		   (begin
                     (debug:print-info 2 *default-log-port* "Writing " start-flag)
		     (with-output-to-file start-flag (lambda () (print server-key)))
		     (thread-sleep! 0.25)
		     (set! new-server-key (with-input-from-file start-flag (lambda () (read-line))))
		     (equal? server-key new-server-key))
                )
	      #t

           ;; If either of the above conditions is not true, print a "Gating server start" message, wait <idle-time> + 1, then call this function recursively. 
	      (begin
		(debug:print-info 0 *default-log-port* "Gating server start, last start: "
				  (seconds->time-string fmodtime) ", time since last start: " delta ", required idletime: " idletime ", gating reason:" (if old-enough "another job started a server" "too soon to start another server"))

		(thread-sleep! ( + 1 idletime))
		(server:wait-for-server-start-last-flag areapath)))))))


        
;; kind start up of server, wait before allowing another server for a given
;; area to be launched
;;
(define (server:kind-run areapath)
  ;; look for $MT_RUN_AREA_HOME/logs/server-start-last
  ;; and wait for it to be at least <server idletime> seconds old
  (server:wait-for-server-start-last-flag areapath)
  (if (not (server:check-if-running areapath)) ;; why try if there is already a server running?
      (let* (
	     (lock-file    (conc areapath "/logs/server-start.lock")))
	(let* ((start-flag (conc areapath "/logs/server-start-last")))
	  (common:simple-file-lock-and-wait lock-file expire-time: 25)
	  (debug:print-info  2 *default-log-port* "server:kind-run: touching " start-flag)
	  (system (conc "touch " start-flag)) ;; lazy but safe
	  (server:run areapath)
	  (thread-sleep! 20) ;; don't release the lock for at least a few seconds. And allow time for the server startup to get to "SERVER STARTED".
	  (common:simple-file-release-lock lock-file)))

      (debug:print-info 0 *default-log-port* "Found server already running. NOT trying to start another.")
   )
)

;; this one seems to be the general entry point
;;
(define (server:start-and-wait areapath #!key (timeout 60))
  (let ((give-up-time (+ (current-seconds) timeout)))
    (let loop ((server-info (server:check-if-running areapath))
	       (try-num    0))

  (lambda (toppath)
    (set! *db-last-access* (current-seconds)) ;; might not be needed.
    (if (equal? *toppath* toppath)
	#t
	#f)))

;; timeout is hms string: 1h 5m 3s, default is 1 minute


;;
(define (server:expiration-timeout)
  (let ((tmo (configf:lookup *configdat* "server" "timeout")))
    (if (and (string? tmo)
	     (common:hms-string->seconds tmo)) ;; BUG: hms-string->seconds is broken, if given "10" returns 0. Also, it doesn't belong in this logic unless the string->number is changed below
        (* 3600 (string->number tmo))
	60)))

(define (server:get-best-guess-address hostname)
  (let ((res #f))
    (for-each 
     (lambda (adr)
       (if (not (eq? (u8vector-ref adr 0) 127))
	   (set! res adr)))

;;       (server:have-sync-lock?))
;;      (else #f))))

;; 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))

                    ))
              (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))

(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 *time-to-exit*) (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* = " *time-to-exit*" pid="(current-process-id)
			    )))))

(define (server:writable-watchdog-deltasync dbstruct)
  (thread-sleep! 0.05) ;; 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 *time-to-exit*))
	(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 *time-to-exit*)
                                          (> (- (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 *time-to-exit*)       ;; 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 *time-to-exit*)
		(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*="*time-to-exit*)
                                                            
		  (if (and (not *time-to-exit*)
			   (< count 6)) ;; was 11, changing to 4. 
		      (begin
			(thread-sleep! 1)
			(delay-loop (+ count 1))))
		  (if (not *time-to-exit*) (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* = " *time-to-exit*" pid="(current-process-id) ))))))) ;;" this-wd-num="this-wd-num)))))))