Megatest

(include "task_records.scm")

;;======================================================================
;; Tasks db
;;======================================================================

;; wait up to aprox n seconds for a journal to go away
;;
(define (tasks:wait-on-journal path n #!key (remove #f)(waiting-msg #f))
  (if (not (string? path))
      (debug:print 0 "ERROR: Called tasks:wait-on-journal with path=" path " (not a string)")
      (let ((fullpath (conc path "-journal")))
	(handle-exceptions
	 exn
	 (begin
	   (print-call-chain (current-error-port))
	   (debug:print 0 " message: " ((condition-property-accessor 'exn 'message) exn))
	   (debug:print 0 " exn=" (condition->list exn))
	   (debug:print 0 "tasks:wait-on-journal failed. Continuing on, you can ignore this call-chain")
	   #t) ;; if stuff goes wrong just allow it to move on
	 (let loop ((journal-exists (file-exists? fullpath))
		    (count          n)) ;; wait ten times ...
	   (if journal-exists
	       (begin
		 (if (and waiting-msg
			  (eq? (modulo n 30) 0))
		     (debug:print 0 waiting-msg))
		 (if (> count 0)
		     (begin
		       (thread-sleep! 1)
		       (loop (file-exists? fullpath)
			     (- count 1)))
		     (begin
		       (if remove (system (conc "rm -rf " fullpath)))
		       #f)))
	       #t))))))

(define (tasks:get-task-db-path)
  (let* ((linktree     (configf:lookup *configdat* "setup" "linktree"))
	 (dbpath       (conc linktree "/.db")))
    dbpath))



;; If file exists AND
;;    file readable
;;         ==> open it
;; If file exists AND
;;    file NOT readable
;;         ==> open in-mem version
;; If file NOT exists
;;    ==> open in-mem version
;;
(define (tasks:open-db)
  (let* ((dbpath       (conc *toppath* "/monitor.db"))
	 (exists       (file-exists? dbpath))
	 (write-access (file-write-access? dbpath))
	 (mdb          (sqlite3:open-database dbpath)) ;; (never-give-up-open-db dbpath))
	 (handler      (make-busy-timeout 36000)))
    (if (and exists
	     (not write-access))
	(set! *db-write-access* write-access)) ;; only unset so other db's also can use this control
    (sqlite3:set-busy-handler! mdb handler)
    (sqlite3:execute mdb (conc "PRAGMA synchronous = 0;"))
    (if (not exists)
	(begin
	  (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS tasks_queue (id INTEGER PRIMARY KEY,
(define (tasks:open-db #!key (numretries 4))
  (if *task-db*
      *task-db*
      (handle-exceptions
       exn
       (if (> numretries 0)
	   (begin
	     (print-call-chain (current-error-port))
	     (debug:print 0 " message: " ((condition-property-accessor 'exn 'message) exn))
	     (debug:print 0 " exn=" (condition->list exn))
	     (thread-sleep! 1)
	     (tasks:open-db numretries (- numretries 1)))
	   (begin
	     (print-call-chain (current-error-port))
	     (debug:print 0 " message: " ((condition-property-accessor 'exn 'message) exn))
	     (debug:print 0 " exn=" (condition->list exn))))
       (let* ((dbpath       (tasks:get-task-db-path))
	      (dbfile       (conc dbpath "/monitor.db"))
	      (avail        (tasks:wait-on-journal dbpath 10)) ;; wait up to about 10 seconds for the journal to go away
	      (exists       (file-exists? dbpath))
	      (write-access (file-write-access? dbpath))
	      (mdb          (cond ;; what the hek is *toppath* doing here?
			     ((and (string? *toppath*)(file-write-access? *toppath*))
			      (sqlite3:open-database dbfile))
			     ((file-read-access? dbpath)    (sqlite3:open-database dbfile))
			     (else (sqlite3:open-database ":memory:")))) ;; (never-give-up-open-db dbpath))
	      (handler      (make-busy-timeout 36000)))
	 (if (and exists
		  (not write-access))
	     (set! *db-write-access* write-access)) ;; only unset so other db's also can use this control
	 (sqlite3:set-busy-handler! mdb handler)
	 (db:set-sync mdb) ;; (sqlite3:execute mdb (conc "PRAGMA synchronous = 0;"))
	 ;;  (if (or (and (not exists)
	 ;; 	      (file-write-access? *toppath*))
	 ;; 	 (not (file-read-access? dbpath)))
	 ;;      (begin
	 (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS tasks_queue (id INTEGER PRIMARY KEY,
                                action TEXT DEFAULT '',
                                owner TEXT,
                                state TEXT DEFAULT 'new',
                                target TEXT DEFAULT '',
                                name TEXT DEFAULT '',
                                test TEXT DEFAULT '',
                                testpatt TEXT DEFAULT '',
                                item TEXT DEFAULT '',
                                keylock TEXT,
                                params TEXT,
                                creation_time TIMESTAMP,
                                execution_time TIMESTAMP);")
	  (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS monitors (id INTEGER PRIMARY KEY,
	 (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS monitors (id INTEGER PRIMARY KEY,
                                pid INTEGER,
                                start_time TIMESTAMP,
                                last_update TIMESTAMP,
                                hostname TEXT,
                                username TEXT,
                               CONSTRAINT monitors_constraint UNIQUE (pid,hostname));")
	  (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS servers (id INTEGER PRIMARY KEY,
	 (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS servers (id INTEGER PRIMARY KEY,
                                  pid INTEGER,
                                  interface TEXT,
                                  hostname TEXT,
                                  port INTEGER,
                                  pubport INTEGER,
                                  start_time TIMESTAMP,
                                  priority INTEGER,
                                  state TEXT,
                                  mt_version TEXT,
                                  heartbeat TIMESTAMP,
                                  transport TEXT,
                                  run_id INTEGER);")
                               CONSTRAINT servers_constraint UNIQUE (pid,hostname,port));")
	  (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS clients (id INTEGER PRIMARY KEY,
	 ;;                               CONSTRAINT servers_constraint UNIQUE (pid,hostname,port));")
	 (sqlite3:execute mdb "CREATE TABLE IF NOT EXISTS clients (id INTEGER PRIMARY KEY,
                                  server_id INTEGER,
                                  pid INTEGER,
                                  hostname TEXT,
                                  cmdline TEXT,
                                  login_time TIMESTAMP,
                                  logout_time TIMESTAMP DEFAULT -1,
                                CONSTRAINT clients_constraint UNIQUE (pid,hostname));")
                                  
	  ))
    mdb))
    
	       
	       ;))
	 (sqlite3:execute mdb "DELETE FROM tasks_queue WHERE state='done' AND creation_time < ?;" (- (current-seconds)(* 24 60 60))) ;; remove older than 24 hrs
	 (set! *task-db* (cons mdb dbpath))
	 *task-db*))))

;;======================================================================
;; Server and client management
;;======================================================================

;; make-vector-record tasks hostinfo id interface port pubport transport pid hostname
(define (tasks:hostinfo-get-id          vec)    (vector-ref  vec 0))
(define (tasks:hostinfo-get-interface   vec)    (vector-ref  vec 1))
(define (tasks:hostinfo-get-port        vec)    (vector-ref  vec 2))
(define (tasks:hostinfo-get-pubport     vec)    (vector-ref  vec 3))
(define (tasks:hostinfo-get-transport   vec)    (vector-ref  vec 4))
(define (tasks:hostinfo-get-pid         vec)    (vector-ref  vec 5))
(define (tasks:hostinfo-get-hostname    vec)    (vector-ref  vec 6))

;; state: 'live, 'shutting-down, 'dead
(define (tasks:server-register mdb pid interface port priority state transport #!key (pubport -1))
  (debug:print-info 11 "tasks:server-register " pid " " interface " " port " " priority " " state)
(define (tasks:server-lock-slot mdb run-id)
  (tasks:server-clean-out-old-records-for-run-id mdb run-id " tasks:server-lock-slot")
  (if (< (tasks:num-in-available-state mdb run-id) 4)
      (begin 
	(tasks:server-set-available mdb run-id)
	;; (thread-sleep! 2) ;; Try removing this. It may not be needed.
	(tasks:server-am-i-the-server? mdb run-id))
      #f))
	
;; register that this server may come online (first to register goes though with the process)
(define (tasks:server-set-available mdb run-id)
  (sqlite3:execute 
   mdb 
   "INSERT OR REPLACE INTO servers (pid,hostname,port,pubport,start_time,priority,state,mt_version,heartbeat,interface,transport)
                             VALUES(?,  ?,       ?,   ?,  strftime('%s','now'), ?, ?, ?, strftime('%s','now'),?,?);"
   pid (get-host-name) port pubport priority (conc state) 
   (common:version-signature)
   interface 
   (conc transport))
   "INSERT INTO servers (pid,hostname,port,pubport,start_time,      priority,state,mt_version,heartbeat,   interface,transport,run_id)
                   VALUES(?, ?,       ?,   ?, strftime('%s','now'), ?,       ?,    ?,-1,?,        ?,        ?);"
   (current-process-id)       ;; pid
   (get-host-name)            ;; hostname
   -1                         ;; port
   -1                         ;; pubport
   (random 1000)              ;; priority (used a tiebreaker on get-available)
   "available"                ;; state
   (common:version-signature) ;; mt_version
   -1                         ;; interface
   "http"                     ;; transport
  (vector 
   (tasks:server-get-server-id mdb (get-host-name) interface port pid)
   interface
   port
   run-id
   pubport
   transport
   ))

(define (tasks:num-in-available-state mdb run-id)
  (let ((res 0))
    (sqlite3:for-each-row
     (lambda (num-in-queue)
       (set! res num-in-queue))
     mdb
     "SELECT count(id) FROM servers WHERE run_id=? AND state = 'available' AND (strftime('%s','now') - start_time) < 30 ;"
     run-id)
    res))
;; NB// two servers with same pid on different hosts will be removed from the list if pid: is used!
(define (tasks:server-deregister mdb hostname #!key (port #f)(pid #f)(action 'delete))
  (debug:print-info 11 "server-deregister " hostname ", port " port ", pid " pid)
  (if *db-write-access*
      (if pid
	  (case action
	    ((delete)(sqlite3:execute mdb "DELETE FROM servers WHERE pid=?;" pid))
	    (else    (sqlite3:execute mdb "UPDATE servers SET state='dead' WHERE pid=?;" pid)))
	  (if port
	      (case action
	    ((delete)(sqlite3:execute mdb "DELETE FROM servers WHERE (interface=? or hostname=?) AND port=?;" hostname hostname port))
	    (else    (sqlite3:execute mdb "UPDATE servers SET state='dead' WHERE (interface=? or hostname=?) AND port=?;" hostname hostname port)))
	      (debug:print 0 "ERROR: tasks:server-deregister called with neither pid nor port specified")))))

(define (tasks:num-servers-non-zero-running mdb)
  (let ((res 0))
    (sqlite3:for-each-row
     (lambda (num-running)
       (set! res num-running))
     mdb
     "SELECT count(id) FROM servers WHERE run_id != 0 AND state = 'running';")
    res))

(define (tasks:server-clean-out-old-records-for-run-id mdb run-id tag)
  (sqlite3:execute mdb "UPDATE servers SET state=?,heartbeat=strftime('%s','now') WHERE state in ('available','dbprep','shutting-down') AND (strftime('%s','now') - start_time) > 50 AND run_id=?;"
		   (conc "defunct" tag) run-id))

(define (tasks:server-force-clean-running-records-for-run-id mdb run-id tag)
  (sqlite3:execute mdb "UPDATE servers SET state=?,heartbeat=strftime('%s','now') WHERE state = 'running' AND run_id=?;"
		   (conc "defunct" tag) run-id))

(define (tasks:server-force-clean-run-record mdb run-id iface port tag)
  (sqlite3:execute mdb "UPDATE servers SET state=?,heartbeat=strftime('%s','now') WHERE state = 'running' AND run_id=? AND interface=? AND port=?;"
		   (conc "defunct" tag) run-id iface port))

(define (tasks:server-delete-records-for-this-pid mdb tag)
  (sqlite3:execute mdb "UPDATE servers SET state=?,heartbeat=strftime('%s','now') WHERE hostname=? AND pid=?;"
		   (conc "defunct" tag) (get-host-name) (current-process-id)))

(define (tasks:server-delete-record mdb server-id tag) 
  (sqlite3:execute mdb "UPDATE servers SET state=?,heartbeat=strftime('%s','now') WHERE id=?;"
		   (conc "defunct" tag) server-id)
  ;; use this opportuntity to clean out records over one month old or over 10 minutes old with port = -1 (i.e. a never used placeholder)
  (sqlite3:execute mdb "DELETE FROM servers WHERE state not in ('running','shutting-down','dbprep') AND (strftime('%s','now') - start_time) > 2628000;")
  (sqlite3:execute mdb "DELETE FROM servers WHERE state like 'defunct%' AND port=-1 AND (strftime('%s','now') - start_time) > 600;")
  )

(define (tasks:server-set-state! mdb server-id state)
  (sqlite3:execute mdb "UPDATE servers SET state=?,heartbeat=strftime('%s','now') WHERE id=?;" state server-id))

(define (tasks:server-set-interface-port mdb server-id interface port)
  (sqlite3:execute mdb "UPDATE servers SET interface=?,port=?,heartbeat=strftime('%s','now') WHERE id=?;" interface port server-id))

;; Get random port not used in long time
;;
(define (tasks:server-get-next-port mdb)
  (let* ((lownum        30000)
	(highnum        64000)
	(used-ports     '())
	(get-rand-port  (lambda ()
			  (+ lownum (random (- highnum lownum)))))
	(port-param     (if (and (args:get-arg "-port")
				 (string->number (args:get-arg "-port")))
			    (string->number (args:get-arg "-port"))
			    #f))
	;; (config-port    (if (and (config-lookup  *configdat* "server" "port")
	;; 			 (string->number (config-lookup  *configdat* "server" "port")))
	;; 		    (string->number (config-lookup  *configdat* "server" "port"))
	;; 		    #f))
	)
    (sqlite3:for-each-row
     (lambda (port)
       (set! used-ports (cons port used-ports)))
     mdb
     "SELECT port FROM servers;")
    (cond
     ((and port-param res)   (if (> res port-param) res port-param))
     (port-param             port-param)
     ;; ((and config-port res)  (if (> res config-port) res config-port))
     ;; (config-port            config-port)
     (else
      (let loop ((port     (get-rand-port))
		 (remtries 100))
	(if (member port used-ports)
	    (if (> remtries 0)
		(loop (get-rand-port)(- remtries 1))
		(get-rand-port))
	    port))))))

(define (tasks:server-deregister-self mdb hostname)
  (tasks:server-deregister mdb hostname pid: (current-process-id)))

;; need a simple call for robustly removing records given host and port
(define (tasks:server-delete mdb hostname port)
  (tasks:server-deregister mdb hostname port: port action: 'delete))
(define (tasks:server-am-i-the-server? mdb run-id)
  (let* ((all    (tasks:server-get-servers-vying-for-run-id mdb run-id))
	 (first  (if (null? all)
		     (begin (debug:print 0 "ERROR: no servers listed, should be at least one by now.") 
			    (sqlite3:finalize! mdb)
			    (exit 1))
		     (car (db:get-rows all))))
	 (header   (db:get-header all))
	 (id       (db:get-value-by-header first header "id"))
	 (hostname (db:get-value-by-header first header "hostname"))
	 (pid      (db:get-value-by-header first header "pid"))
	 (priority (db:get-value-by-header first header "priority")))
    (debug:print 0 "INFO: am-i-the-server got record " first)
    ;; for now a basic check. add tiebreaking by priority later
    (if (and (equal? hostname (get-host-name))
	     (equal? pid      (current-process-id)))
	id
	#f)))
	     
;; Use: (db:get-value-by-header (car (db:get-rows dat)) (db:get-header dat) "fieldname")
;;  to extract info from the structure returned
;;
(define (tasks:server-get-servers-vying-for-run-id mdb run-id)
   (let* ((header (list "id" "hostname" "pid" "interface" "port" "pubport" "state" "run_id" "priority" "start_time"))
	  (selstr (string-intersperse header ","))
	  (res    '()))
    (sqlite3:for-each-row
     (lambda (a . b)
       (set! res (cons (apply vector a b) res)))
     mdb
     (conc "SELECT " selstr " FROM servers WHERE run_id=? AND state in ('available','running','dbprep') ORDER BY start_time DESC;")
     run-id)
    (vector header res)))

(define (tasks:get-server mdb run-id #!key (retries 10))
  (let ((res  #f)
	(best #f))
    (handle-exceptions
     exn
     (begin
       (print-call-chain (current-error-port))
       (debug:print 0 "WARNING: tasks:get-server db access error.")
       (debug:print 0 " message: " ((condition-property-accessor 'exn 'message) exn))
       (debug:print 0 " for run " run-id)
       (print-call-chain (current-error-port))
       (if (> retries 0)
	   (begin
	     (debug:print 0 " trying call to tasks:get-server again in 10 seconds")
	     (thread-sleep! 10)
	     (tasks:get-server mdb run-id retries: (- retries 0)))
	   (debug:print 0 "10 tries of tasks:get-server all crashed and burned. Giving up and returning \"no server found\"")))
     (sqlite3:for-each-row
      (lambda (id interface port pubport transport pid hostname)
	(set! res (vector id interface port pubport transport pid hostname)))
      mdb
      ;; removed:
      ;; strftime('%s','now')-heartbeat < 10 AND mt_version = ?
      "SELECT id,interface,port,pubport,transport,pid,hostname FROM servers
          WHERE run_id=? AND state='running'
          ORDER BY start_time DESC LIMIT 1;" run-id) ;; (common:version-signature) run-id)
     res)))

(define (tasks:server-running-or-starting? mdb run-id)
  (let ((res #f))
    (sqlite3:for-each-row
     (lambda (id)
       (set! res id))
     mdb ;; NEEDS dbprep ADDED
     "SELECT id FROM servers WHERE run_id=? AND (state = 'running' OR (state = 'dbprep' AND  (strftime('%s','now') - start_time) < 60));" run-id)
    res))

(define (tasks:server-get-server-id mdb hostname iface port pid)
(define (tasks:server-running? mdb run-id)
  (debug:print-info 12 "tasks:server-get-server-id " mdb " " hostname " " iface " " port " " pid)
  (let ((res #f))
    (sqlite3:for-each-row
     (lambda (id)
       (set! res id))
     mdb
     mdb ;; NEEDS dbprep ADDED
     (cond
      ((and hostname  pid)  "SELECT id FROM servers WHERE hostname=?  AND pid=?;")
      ((and iface     port) "SELECT id FROM servers WHERE interface=? AND port=?;")
      ((and hostname  port) "SELECT id FROM servers WHERE hostname=?  AND port=?;")
      (else
       (begin
	 (debug:print 0 "ERROR: tasks:server-get-server-id needs (hostname and pid) OR (iface and port) OR (hostname and port)")
	 "SELECT id FROM servers WHERE pid=-999;")))
     "SELECT id FROM servers WHERE run_id=? AND state = 'running';" run-id)
     (if hostname hostname iface)(if pid pid port))
    res))

(define (tasks:server-update-heartbeat mdb server-id)
(define (tasks:need-server run-id)
  (debug:print-info 1 "Heart beat update of server id=" server-id)
  (handle-exceptions
   exn
   (begin
     (debug:print 0 "WARNING: probable timeout on monitor.db access")
     (thread-sleep! 1)
     (tasks:server-update-heartbeat mdb server-id))
  (let ((forced (configf:lookup *configdat* "server" "required"))
   (sqlite3:execute mdb "UPDATE servers SET heartbeat=strftime('%s','now') WHERE id=?;" server-id)))

	(maxqry (cdr (rmt:get-max-query-average run-id)))
;; alive servers keep the heartbeat field upto date with seconds every 6 or so seconds
(define (tasks:server-alive? mdb server-id #!key (iface #f)(hostname #f)(port #f)(pid #f))
  (let* ((server-id  (if server-id 
			 server-id
	(threshold   (string->number (or (configf:lookup *configdat* "server" "server-query-threshold") "10"))))
			 (tasks:server-get-server-id mdb hostname iface port pid)))
	 (heartbeat-delta 99e9))
    (sqlite3:for-each-row
    (cond
     (forced 
     (lambda (delta)
       (set! heartbeat-delta delta))
     mdb "SELECT strftime('%s','now')-heartbeat FROM servers WHERE id=?;" server-id)
    (< heartbeat-delta 10)))

      (if (common:low-noise-print 60 run-id "server required is set")
(define (tasks:client-register mdb pid hostname cmdline)
  (sqlite3:execute
   mdb
   "INSERT OR REPLACE INTO clients (server_id,pid,hostname,cmdline,login_time) VALUES(?,?,?,?,strftime('%s','now'));")
  (tasks:server-get-server-id mdb hostname #f #f pid)
  pid hostname cmdline)

	  (debug:print-info 0 "Server required is set, starting server."))
(define (tasks:client-logout mdb pid hostname cmdline)
  (sqlite3:execute
      #t)
   mdb
   "UPDATE clients SET logout_time=strftime('%s','now') WHERE pid=? AND hostname=? AND cmdline=?;"
   pid hostname cmdline))

     ((> maxqry threshold)
(define (tasks:get-logged-in-clients mdb server-id)
  (let ((res '()))
    (sqlite3:for-each-row 
     (lambda (id server-id pid hostname cmdline login-time logout-time)
       (set! res (cons (vector id server-id pid hostname cmdline login-time lougout-time) res)))
     mdb
     "SELECT id,server_id,pid,hostname,cmdline,login_time,logout_time FROM clients WHERE server_id=?;"
     server-id)))

      (if (common:low-noise-print 60 run-id "Max query time execeeded")
(define (tasks:have-clients? mdb server-id)
  (null? (tasks:get-logged-in-clients mdb server-id)))

	  (debug:print-info 0 "Max avg query time of " maxqry "ms exceeds limit of " threshold "ms, starting server."))
;; ping each server in the db and return first found that responds. 
;; remove any others. will not necessarily remove all!
(define (tasks:get-best-server mdb)
  (let ((res '())
	(best #f)
      #t)
	(transport (if (and *transport-type*
			    (not (eq? *transport-type* 'fs)))
     (else
		       (conc *transport-type*)
		       "%")))
      #f))))
    (sqlite3:for-each-row
     (lambda (id interface port pubport transport pid hostname)
       (set! res (cons (vector id interface port pubport transport pid hostname) res))
       ;;(debug:print-info 2 "Found existing server " hostname ":" port " registered in db"))
       )

     mdb
     
;; try to start a server and wait for it to be available
     "SELECT id,interface,port,pubport,transport,pid,hostname FROM servers
          WHERE strftime('%s','now')-heartbeat < 10 
          AND mt_version=? AND transport LIKE ? 
          ORDER BY start_time DESC LIMIT 1;" (common:version-signature) transport)
    ;; for now we are keeping only one server registered in the db, return #f or first server found
    (if (null? res) #f (car res))))

;; BUG: This logic is probably needed unless methodology changes completely...
;;
;;     (if (null? res) #f
;; 	(let loop ((hed (car res))
;; 		   (tal (cdr res)))
;; 	  ;; (print "hed=" hed ", tal=" tal)
;; 	  (let* ((host     (list-ref hed 0))
;; 		 (iface    (list-ref hed 1))
;; 		 (port     (list-ref hed 2))
;; 		 (pid      (list-ref hed 4))
;; 		 (alive    (open-run-close tasks:server-alive? tasks:open-db #f hostname: host port: port)))
;; 	    (if alive
;; 		(begin
;; 		  (debug:print-info 2 "Found an existing, alive, server " host ", " port ".")
;; 		  (list host iface port))
;; 		(begin
;; 		  (debug:print-info 1 "Marking " host ":" port " as dead in server registry.")
;; 		  (if port
;; 		      (open-run-close tasks:server-deregister tasks:open-db host port: port)
;; 		      (open-run-close tasks:server-deregister tasks:open-db host pid:  pid))
;; 		  (if (null? tal)
;; 		      #f
;; 		      (loop (car tal)(cdr tal))))))))))

(define (tasks:start-and-wait-for-server tdbdat run-id delay-max-tries)
(define (tasks:remove-server-records mdb)
  (sqlite3:execute mdb "DELETE FROM servers;"))

  ;; ensure a server is running for this run
(define (tasks:mark-server hostname port pid state transport)
  (if port
      (open-run-close tasks:server-deregister tasks:open-db hostname port: port)
      (open-run-close tasks:server-deregister tasks:open-db hostname pid:  pid)))


  (let loop ((server-dat (tasks:get-server (db:delay-if-busy tdbdat) run-id))
	     (delay-time 0))
(define (tasks:kill-server status hostname port pid transport)
  (debug:print-info 1 "Removing defunct server record for " hostname ":" port)
  (if port
      (if (and (not server-dat)
      (open-run-close tasks:server-deregister tasks:open-db hostname port: port)
      (open-run-close tasks:server-deregister tasks:open-db hostname pid:  pid))
  (if status ;; #t means alive
      (begin
	       (< delay-time delay-max-tries))
	  (begin
	(if (equal? hostname (get-host-name))
	    (handle-exceptions
	     exn
	     (debug:print-info 0 "server may or may not be dead, check for megatest -server running as pid " pid "\n"
			       "  EXCEPTION: " ((condition-property-accessor 'exn 'message) exn))
	     (debug:print 1 "Sending signal/term to " pid " on " hostname)
	    (if (common:low-noise-print 60 "tasks:start-and-wait-for-server" run-id)
		(debug:print 0 "Try starting server for run-id " run-id))
	     (process-signal pid signal/term)
	     (thread-sleep! 5) ;; give it five seconds to die peacefully then do a brutal kill
	     ;;(process-signal pid signal/kill)
	     ) ;; local machine, send sig term
	    (begin
	      ;;(debug:print-info 1 "Stopping remote servers not yet supported."))))
	      (debug:print-info 1 "Telling alive server on " hostname ":" port " to commit servercide")
	      (let ((serverdat (list hostname port)))
	    (server:kind-run run-id)
	      	(case (if (string? transport) (string->symbol transport) transport)
	      	  ((http)(http-transport:client-connect hostname port))
	      	  (else  (debug:print "ERROR: remote stopping servers of type " transport " not supported yet")))
	      	(cdb:kill-server serverdat pid)))))    ;; remote machine, try telling server to commit suicide
      (begin
	(if status 
	    (if (equal? hostname (get-host-name))
		(begin
		  (debug:print-info 1 "Sending signal/term to " pid " on " hostname)
		  (process-signal pid signal/term)  ;; local machine, send sig term
		  (thread-sleep! 5)                 ;; give it five seconds to die peacefully then do a brutal kill
		  (process-signal pid signal/kill)) 
		(debug:print 0 "WARNING: Can't kill frozen server on remote host " hostname))))))


	    (thread-sleep! (min delay-time 5))
	    (loop (tasks:get-server (db:delay-if-busy tdbdat) run-id)(+ delay-time 1))))))

(define (tasks:get-all-servers mdb)
  (let ((res '()))
    (sqlite3:for-each-row
     (lambda (id pid hostname interface port pubport start-time priority state mt-version last-update transport)
       (set! res (cons (vector id pid hostname interface port pubport start-time priority state mt-version last-update transport) res)))
     (lambda (id pid hostname interface port pubport start-time priority state mt-version last-update transport run-id)
       ;;                       0  1     2         3      4     5          6        7     8          9          10        11     12
       (set! res (cons (vector id pid hostname interface port pubport start-time priority state mt-version last-update transport run-id) res)))
     mdb
     "SELECT id,pid,hostname,interface,port,pubport,start_time,priority,state,mt_version,strftime('%s','now')-heartbeat AS last_update,transport FROM servers ORDER BY start_time DESC;")
     "SELECT id,pid,hostname,interface,port,pubport,start_time,priority,state,mt_version,strftime('%s','now')-heartbeat AS last_update,transport,run_id FROM servers WHERE state NOT LIKE 'defunct%' ORDER BY start_time DESC;")
    res))
       

;; no elegance here ...
;;
(define (tasks:kill-server hostname pid)
  (debug:print-info 0 "Attempting to kill server process " pid " on host " hostname)
  (setenv "TARGETHOST" hostname)
  (setenv "TARGETHOST_LOGF" "server-kills.log")
  (system (conc "nbfake kill " pid))
  (unsetenv "TARGETHOST_LOGF")
  (unsetenv "TARGETHOST"))
 
;; look up a server by run-id and send it a kill, also delete the record for that server
;;
(define (tasks:kill-server-run-id run-id #!key (tag "default"))
  (let* ((tdbdat  (tasks:open-db))
	 (sdat    (tasks:get-server (db:delay-if-busy tdbdat) run-id)))
    (if sdat
	(let ((hostname (vector-ref sdat 6))
	      (pid      (vector-ref sdat 5))
	      (server-id (vector-ref sdat 0)))
	  (tasks:server-set-state! (db:delay-if-busy tdbdat) server-id "killed")
	  (debug:print-info 0 "Killing server " server-id " for run-id " run-id " on host " hostname " with pid " pid)
	  (tasks:kill-server hostname pid)
	  (tasks:server-delete-record (db:delay-if-busy tdbdat) server-id tag) )
	(debug:print-info 0 "No server found for run-id " run-id ", nothing to kill"))
    ;; (sqlite3:finalize! tdb)
    ))
    

;;======================================================================
;; Tasks and Task monitors
;;======================================================================


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

       (set! res count))
     mdb
     "SELECT count(id) FROM monitors WHERE last_update < (strftime('%s','now') - 300) AND username=?;"
     (car (user-information (current-user-id))))
    res))

;; register a task
(define (tasks:add mdb action owner target runname test item params)
  (sqlite3:execute mdb "INSERT INTO tasks_queue (action,owner,state,target,name,test,item,params,creation_time,execution_time)
                       VALUES (?,?,'new',?,?,?,?,?,strftime('%s','now'),0);" 
(define (tasks:add mdb action owner target runname testpatt params)
  (sqlite3:execute mdb "INSERT INTO tasks_queue (action,owner,state,target,name,testpatt,params,creation_time,execution_time)
                       VALUES (?,?,'new',?,?,?,?,strftime('%s','now'),0);" 
		   action
		   owner
		   target
		   runname
		   test
		   testpatt
		   item
		   (if params params "")))

(define (keys:key-vals-hash->target keys key-params)
  (let ((tmp (hash-table-ref/default key-params (vector-ref (car keys) 0) "")))
    (if (> (length keys) 1)
	(for-each (lambda (key)
		    (set! tmp (conc tmp "/" (hash-table-ref/default key-params (vector-ref key 0) ""))))

  (sqlite3:execute mdb (conc "DELETE FROM tasks_queue WHERE id IN (" task-ids ");")))

;; 
(define (tasks:start-monitor db mdb)
  (if (> (tasks:get-num-alive-monitors mdb) 2) ;; have two running, no need for more
      (debug:print-info 1 "Not starting monitor, already have more than two running")
      (let* ((megatestdb     (conc *toppath* "/megatest.db"))
	     (monitordbf     (conc *toppath* "/monitor.db"))
	     (monitordbf     (conc (configf:lookup *configdat* "setup" "linktree") "/.db/monitor.db"))
	     (last-db-update 0)) ;; (file-modification-time megatestdb)))
	(task:register-monitor mdb)
	(let loop ((count      0)
		   (next-touch 0)) ;; next-touch is the time where we need to update last_update
	  ;; if the db has been modified we'd best look at the task queue
	  (let ((modtime (file-modification-time megatestdbpath )))
	    (if (> modtime last-db-update)

		   (current-process-id)
		   (get-host-name)))

(define (tasks:set-state mdb task-id state)
  (sqlite3:execute mdb "UPDATE tasks_queue SET state=? WHERE id=?;" 
		   state 
		   task-id))

;;======================================================================
;; Access using task key (stored in params; (hash-table->alist flags) hostname pid
;;======================================================================

(define (tasks:param-key->id mdb task-params)
  (handle-exceptions
   exn
   #f
   (sqlite3:first-result mdb "SELECT id FROM tasks_queue WHERE params LIKE ?;" task-params)))

(define (tasks:set-state-given-param-key mdb param-key new-state)
  (sqlite3:execute mdb "UPDATE tasks_queue SET state=? WHERE params LIKE ?;" new-state param-key))

(define (tasks:get-records-given-param-key mdb param-key state-patt action-patt test-patt)
  (handle-exceptions
   exn
   '()
   (sqlite3:first-row mdb "SELECT id,action,owner,state,target,name,testpatt,keylock,params WHERE
         params LIKE ? AND state LIKE ? AND action LIKE ? AND testpatt LIKE ?;"
		      param-key state-patt action-patt test-patt)))


;;======================================================================
;; Rogue items, no place to put these yet
;;======================================================================

(define (tasks:find-task-queue-records mdb target run-name test-patt state-patt action-patt)
  ;; (handle-exceptions
  ;;  exn
  ;;  '()
  ;;  (sqlite3:first-row
  (let ((res '()))
    (sqlite3:for-each-row 
     (lambda (a . b)
       (set! res (cons (cons a b) res)))
     mdb "SELECT id,action,owner,state,target,name,testpatt,keylock,params FROM tasks_queue 
       WHERE
         target = ? AND name = ? AND state LIKE ? AND action LIKE ? AND testpatt LIKE ?;"
     target run-name state-patt action-patt test-patt)
    res)) ;; )


(define (tasks:kill-runner mdb target run-name)
  (let ((records    (tasks:find-task-queue-records mdb target run-name "%" "running" "run-tests"))
	(hostpid-rx (regexp "\\s+(\\w+)\\s+(\\d+)$"))) ;; host pid is at end of param string
    (if (null? records)
	(debug:print 0 "No run launching processes found for " target " / " run-name)
	(debug:print 0 "Found " (length records) " run(s) to kill."))
    (for-each 
     (lambda (record)
       (let* ((param-key (list-ref record 8))
	      (match-dat (string-search hostpid-rx param-key)))
	 (if match-dat
	     (let ((hostname  (cadr match-dat))
		   (pid       (string->number (caddr match-dat))))
	       (debug:print 0 "Sending SIGINT to process " pid " on host " hostname)
	       (if (equal? (get-host-name) hostname)
		   (if (process:alive? pid)
		       (begin
			 (handle-exceptions
			  exn
			  (begin
			    (debug:print 0 "Kill of process " pid " on host " hostname " failed.")
			    (debug:print 0 " message: " ((condition-property-accessor 'exn 'message) exn))
			    #t)
			  (process-signal pid signal/int)
			  (thread-sleep! 5)
			  (if (process:alive? pid)
			      (process-signal pid signal/kill)))))
		   ;;  (call-with-environment-variables
		   (let ((old-targethost (getenv "TARGETHOST")))
		     (setenv "TARGETHOST" hostname)
		     (setenv "TARGETHOST_LOGF" "server-kills.log")
		     (system (conc "nbfake kill " pid))
		     (if old-targethost (setenv "TARGETHOST" old-targethost))
		     (unsetenv "TARGETHOST")
		     (unsetenv "TARGETHOST_LOGF"))))
	     (debug:print 0 "ERROR: no record or improper record for " target "/" run-name " in tasks_queue in monitor.db"))))
     records)))


;;======================================================================
;; The routines to process tasks
;;======================================================================

;; NOTE: It might be good to add one more layer of checking to ensure
;;       that no task gets run in parallel.