Megatest

server.scm at [653a2890a7]
Login

File server.scm artifact 37237a3d19 part of check-in 653a2890a7


;; Copyright 2006-2017, Matthew Welland.
;; 
;; This file is part of Megatest.
;; 
;;     Megatest is free software: you can redistribute it and/or modify
;;     it under the terms of the GNU General Public License as published by
;;     the Free Software Foundation, either version 3 of the License, or
;;     (at your option) any later version.
;; 
;;     Megatest is distributed in the hope that it will be useful,
;;     but WITHOUT ANY WARRANTY; without even the implied warranty of
;;     MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
;;     GNU General Public License for more details.
;; 
;;     You should have received a copy of the GNU General Public License
;;     along with Megatest.  If not, see <http://www.gnu.org/licenses/>.
;;

(require-extension (srfi 18) extras tcp s11n)

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

(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))

(declare (uses commonmod))
(import commonmod)

(declare (uses dbmod))
(import dbmod)

(declare (uses configfmod))
(import configfmod)

(declare (uses servermod))
(import servermod)

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

(define  *server-loop-heart-beat* (current-seconds))

;;======================================================================
;; P K T S   S T U F F 
;;======================================================================

;; ???

;;======================================================================
;; P K T S   S T U F F 
;;======================================================================

;; ???

;;======================================================================
;; S E R V E R
;;======================================================================

;; Call this to start the actual server
;;

;; all routes though here end in exit ...
;;
;; start_server
;;
(define (server:launch run-id transport-type)
  (case transport-type
    ((http)(http-transport:launch))
    ;;((nmsg)(nmsg-transport:launch run-id))
    ;;((rpc)  (rpc-transport:launch run-id))
    (else (debug:print-error 0 *default-log-port* "unknown server type " transport-type))))

;; Given a run id start a server process    ### NOTE ### > file 2>&1 
;; if the run-id is zero and the target-host is set 
;; try running on that host
;;   incidental: rotate logs in logs/ dir.
;;
(define  (server:run areapath) ;; areapath is *toppath* for a given testsuite area
  (let* ((curr-host   (get-host-name))
         ;; (attempt-in-progress (server:start-attempted? areapath))
         ;; (dot-server-url (server:check-if-running areapath))
	 (curr-ip     (server:get-best-guess-address curr-host))
	 (curr-pid    (current-process-id))
	 (homehost    (common:get-homehost)) ;; configf:lookup *configdat* "server" "homehost" ))
	 (target-host (car homehost))
	 (testsuite   (common:get-testsuite-name))
	 (logfile     (conc areapath "/logs/server.log")) ;; -" curr-pid "-" target-host ".log"))
	 (profile-mode (or (configf:lookup *configdat* "misc" "profilesw")
			   ""))
	 (cmdln (conc (common:get-megatest-exe)
		      " -server " (or target-host "-") (if (equal? (configf:lookup *configdat* "server" "daemonize") "yes")
							   " -daemonize "
							   "")
		      ;; " -log " logfile
		      " -m testsuite:" testsuite
		      " " profile-mode
		      )) ;; (conc " >> " logfile " 2>&1 &")))))
	 (log-rotate  (make-thread common:rotate-logs  "server run, rotate logs thread")) ;; why are we rotating logs here? This is a sensitive location with a lot going on!?
         (load-limit  (configf:lookup-number *configdat* "jobtools" "max-server-start-load" default: 3.0)))
    ;; we want the remote server to start in *toppath* so push there
    (push-directory areapath)
    (debug:print 0 *default-log-port* "INFO: Trying to start server (" cmdln ") ...")
    (thread-start! log-rotate)
    
    ;; host.domain.tld match host?
    (if (and target-host 
	     ;; look at target host, is it host.domain.tld or ip address and does it 
	     ;; match current ip or hostname
	     (not (string-match (conc "("curr-host "|" curr-host"\\..*)") target-host))
	     (not (equal? curr-ip target-host)))
	(begin
	  (debug:print-info 0 *default-log-port* "Starting server on " target-host ", logfile is " logfile)
	  (setenv "TARGETHOST" target-host)))
      
    (setenv "TARGETHOST_LOGF" logfile)
    (thread-sleep! (/ (random 5000) 1000)) ;; add about a random (up to 5 seconds) initial delay. It seems pretty common that many running tests request a server at the same time
    ;; (common:wait-for-normalized-load load-limit " delaying server start due to load" target-host) ;; do not try starting servers on an already overloaded machine, just wait forever
    #;(common:wait-for-homehost-load load-limit (conc " delaying server start due to load on homehost. limit is " load-limit))
    (system (conc "nbfake " cmdln))
    (unsetenv "TARGETHOST_LOGF")
    (if (get-environment-variable "TARGETHOST")(unsetenv "TARGETHOST"))
    (thread-join! log-rotate)
    (pop-directory)))











(define (server:ping host-port-in server-id #!key (do-exit #f))
  (let ((host:port (if (not host-port-in) ;; use read-dotserver to find
		       #f ;; (server:check-if-running *toppath*)
		;; (if (number? host-port-in) ;; we were handed a server-id
		;; 	   (let ((srec (tasks:get-server-by-id (db:delay-if-busy (tasks:open-db)) host-port-in)))
		;; 	     ;; (print "srec: " srec " host-port-in: " host-port-in)
		;; 	     (if srec
		;; 		 (conc (vector-ref srec 3) ":" (vector-ref srec 4))
		;; 		 (conc "no such server-id " host-port-in)))
		       host-port-in))) ;; )
    (let* ((host-port (if host:port
			  (let ((slst (string-split   host:port ":")))
			    (if (eq? (length slst) 2)
				(list (car slst)(string->number (cadr slst)))
				#f))
			  #f)))
;;	   (toppath       (launch:setup)))
      ;; (print "host-port=" host-port)
      (if (not host-port)
	  (begin
	    (if host-port-in
		(debug:print 0 *default-log-port*  "ERROR: bad host:port"))
	    (if do-exit (exit 1))
	    #f)
	  (let* ((iface      (car host-port))
		 (port       (cadr host-port))
		 (server-dat (http-transport:client-connect iface port server-id))
		 (login-res  (rmt:login-no-auto-client-setup server-dat)))
	    (if (and (list? login-res)
		     (car login-res))
		(begin
		  ;; (print "LOGIN_OK")
		  (if do-exit (exit 0))
		  #t)
		(begin
		  ;; (print "LOGIN_FAILED")
		  (if do-exit (exit 1))
		  #f)))))))


;; kind start up of servers, wait 40 seconds before allowing another server for a given
;; run-id 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 3 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* ((last-run-dat (hash-table-ref/default *server-kind-run* areapath '(0 0))) ;; callnum, whenrun
	     (call-num     (car last-run-dat))
	     (when-run     (cadr last-run-dat))
	     (run-delay    (+ (case call-num
				((0)    0)
				((1)   20)
				((2)  300)
				(else 600))
			      (random 5)))   ;; add a small random number just in case a lot of jobs hit the work hosts simultaneously
	     (lock-file    (conc areapath "/logs/server-start.lock")))
	(if	(> (- (current-seconds) when-run) run-delay)
		(let* ((start-flag (conc areapath "/logs/server-start-last")))
		  (common:simple-file-lock-and-wait lock-file expire-time: 15)
                  (debug:print-info  0 *default-log-port* "server:kind-run: touching " start-flag)
		  (system (conc "touch " start-flag)) ;; lazy but safe
		  (server:run areapath)
		  (thread-sleep! 2) ;; don't release the lock for at least a few seconds
		  (common:simple-file-release-lock lock-file)))
	(hash-table-set! *server-kind-run* areapath (list (+ call-num 1)(current-seconds))))))

;; 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))
      (if (or server-info
	      (> (current-seconds) give-up-time)) ;; server-url will be #f if no server available.
	  (server:record->url server-info)
	  (let ((num-ok (length (server:get-best (server:get-list areapath)))))
	    (if (and (> try-num 0)  ;; first time through simply wait a little while then try again
		     (< num-ok 1))  ;; if there are no decent candidates for servers then try starting a new one
		(server:kind-run areapath))
	    (thread-sleep! 5)
	    (loop (server:check-if-running areapath)
		  (+ try-num 1)))))))


;; no longer care if multiple servers are started by accident. older servers will drop off in time.
;;
(define (server:check-if-running areapath) ;;  #!key (numservers "2"))
  (let* ((ns            (server:get-num-servers))
	 (servers       (server:get-best (server:get-list areapath))))
    (if (or (and servers
		 (null? servers))
	    (not servers)
	    (and (list? servers)
		 (< (length servers) (random ns)))) ;; somewhere between 0 and numservers
        #f
        (let loop ((hed (car servers))
                   (tal (cdr servers)))
          (let ((res (server:check-server hed)))
            (if res
                hed
                (if (null? tal)
                    #f
                    (loop (car tal)(cdr tal)))))))))

;; ping the given server
;;
(define (server:check-server server-record)
  (let* ((server-url (server:record->url server-record))
         (server-id (server:record->id server-record)) 
         (res        (case *transport-type*
                       ((http)(server:ping server-url server-id))
                       ;; ((nmsg)(nmsg-transport:ping (tasks:hostinfo-get-interface server)
                       )))
    (if res
        server-url
	#f)))


(define server:try-running server:run) ;; there is no more per-run servers ;; REMOVE ME. BUG.















(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)))))))