Megatest

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

TODO
====

23WW45 - the revolution branch
. Add "fast" db start option (no handshaking over NFS)
. Add server-ro to server types (just "server" is fine for read/write).
. Create pause-server and resume-server calls
. Create rsync or cp sync to MTRAH function
. Change rmt:send-receive to divert calls to read-only server when possible
. Change start server to call main.db server for 1..N.db servers, block until server is read for use.

23WW21
. Dashboard needs its own cache db in /tmp

23WW07
. Remove use of *dbstruct-dbs*

WW15

        (print-prefix      "Running: ") 
        (archive-info (archive:allocate-new-archive-block blockid-cache *toppath* tsname min-space target-patt run-patt "megatest-db"))
	(archive-dir  (if archive-info (cdr archive-info) #f))
	(archive-id   (if archive-info (car archive-info) -1))
        (home-host    (server:choose-server *toppath* 'homehost))
        (archive-time (seconds->std-time-str (current-seconds)))
        (archive-staging-db (conc *toppath* "/.db-snapshot/archive_" archive-time))
        (tmp-db-path (conc (dbfile:make-tmpdir-name *toppath* "") "/megatest.db"))
        (dbfile             (conc  archive-staging-db "/megatest.db"))) 
        (create-directory archive-staging-db #t)
        (let-values (((pid-val exit-status exit-code) (run-n-wait rsync-exe params: (list "-v" (conc (car home-host) ":"tmp-db-path) archive-staging-db) print-cmd: print-prefix)))
            (if (eq? exit-code 0)   
              (case archiver
	        ((bup) ;; Archive using bup
	          (let* ((bup-init-params  (list "-d" archive-dir "init"))

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

(declare (unit common))
(declare (uses commonmod))
(declare (uses rmtmod))
(declare (uses debugprint))
(declare (uses mtargs))
        

(use srfi-1 data-structures posix regex-case (prefix base64 base64:)
     format dot-locking csv-xml z3 udp ;; sql-de-lite
     hostinfo md5 message-digest typed-records directory-utils stack
     matchable regex posix (srfi 18) extras ;; tcp 
     (prefix nanomsg nmsg:)
     (prefix sqlite3 sqlite3:)

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

(define *common:this-exe-fullpath* (common:get-this-exe-fullpath))
(define *common:this-exe-dir*      (pathname-directory *common:this-exe-fullpath*))
(define *common:this-exe-name*     (pathname-strip-directory *common:this-exe-fullpath*))

(define (common:get-sync-lock-filepath)
  (let* ((tmp-area     (common:make-tmpdir-name *toppath* ""))
         (lockfile     (conc tmp-area "/megatest.db.lock")))
    lockfile))

(define *common:logpro-exit-code->status-sym-alist*
  '( ( 0 . pass )
     ( 1 . fail )
     ( 2 . warn )

;;======================================================================
;; lazy-safe get file mod time. on any error (file not existing etc.) return 0
;;
(define (common:lazy-modification-time fpath)
  (handle-exceptions
      exn
    (begin
      (debug:print 2 *default-log-port* "Failed to get modification time for " fpath ", treating it as zero. exn=" exn)
      0)
    (if (file-exists? fpath)
	(file-modification-time fpath)
	0)))

;;======================================================================
;; find timestamp of newest file associated with a sqlite db file

;; returns: ok/not dbspace required-space
;;
(define (common:check-db-dir-space)
  (let* ((required (string->number 
                    ;; default is 1GB (or actually a billion bytes) This is the number of 1 kB blocks.
		    (or (configf:lookup *configdat* "setup" "dbdir-space-required")
			"1000000")))
	 (dbdir    (common:make-tmpdir-name *toppath* "")) ;; (db:get-dbdir))
	 (tdbspace (common:check-space-in-dir dbdir required))
	 (mdbspace (common:check-space-in-dir *toppath* required)))
    (sort (list tdbspace mdbspace) (lambda (a b)
				     (< (cadr a)(cadr b))))))

;;======================================================================
;; check available space in dbdir, exit if insufficient

    (if valstr
	(val->alist valstr)
	'()))) ;; should it return empty list or #f to indicate not set?


(define (get-section cfgdat section)
  (hash-table-ref/default cfgdat section '()))

(define (common:make-tmpdir-name areapath tmpadj)
  (let* ((area (pathname-file areapath))
         (dname (conc "/tmp/"(current-user-name)"/megatest_localdb/" area "/" (string-translate areapath "/" ".") tmpadj "/.mtdb")))
    (unless (directory-exists? dname)
      (create-directory dname #t))
    dname))

;; dot-locking egg seems not to work, using this for now
;; if lock is older than expire-time then remove it and try again
;; to get the lock
;;
(define (common:simple-file-lock fname #!key (expire-time 300))
  (let* ((lock-exists (file-exists? fname))

    dlog))


;;======================================================================
;;
;;======================================================================
(define (dashboard-tests:examine-test run-id test-id) ;; run-id run-key origtest)
  (let* ((db-path       (common:make-tmpdir-name *toppath* "")) ;; (conc (configf:lookup *configdat* "setup" "linktree") "/db/" run-id ".db"))
	 (dbstruct      #f) ;; NOT USED
	 (testdat        (rmt:get-test-info-by-id run-id test-id)) ;; (db:get-test-info-by-id dbstruct run-id test-id))
	 (db-mod-time   0) ;; (file-modification-time db-path))
	 (last-update   0) ;; (current-seconds))
	 (request-update #t))
    (if (not testdat)
	(begin

(define (dboard:tabdat-make-data)
  (let ((dat (make-dboard:tabdat)))
    (dboard:setup-tabdat dat)
    (dboard:setup-num-rows dat)
    dat))

(define (dboard:setup-tabdat tabdat)
  (dboard:tabdat-dbdir-set! tabdat (common:make-tmpdir-name *toppath* "")) ;; (conc (configf:lookup *configdat* "setup" "linktree") "/.db"))
  (dboard:tabdat-dbfpath-set! tabdat (common:make-tmpdir-name *toppath* ""))
  (dboard:tabdat-monitor-db-path-set! tabdat (conc (dboard:tabdat-dbdir tabdat) "/monitor.db"))


  ;; HACK ALERT: this is a hack, please fix.
  (dboard:tabdat-ro-set! tabdat (not (file-read-access? (dboard:tabdat-dbfpath tabdat))))
  (dboard:tabdat-keys-set! tabdat (rmt:get-keys))
  (dboard:tabdat-dbkeys-set! tabdat (append (dboard:tabdat-keys tabdat) (list "runname")))

	      (if (null? all-test-ids)
		  (hash-table-delete! (dboard:tabdat-allruns-by-id tabdat) run-id)
		  (hash-table-set!    (dboard:tabdat-allruns-by-id tabdat) run-id run-struct))
	      (if (or (null? tal)
		      (> elapsed-time 2)) ;; stop loading data after 5 seconds, on the next call more data *should* be loaded since get-tests-for-run uses last update
		  (begin
		    (when (> elapsed-time 2)   
                      (debug:print 2 *default-log-port* "NOTE: updates are taking a long time, " elapsed-time "s elapsed.")
                      (let* ((old-val (iup:attribute *tim* "TIME"))
                             (new-val (number->string (inexact->exact (floor (* 2  (string->number old-val)))))))
                        (if (< (string->number new-val) 5000)
                            (begin
			      (debug:print 2 *default-log-port* "NOTE: increasing poll interval from "old-val" to "new-val)
			      (iup:attribute-set! *tim* "TIME" new-val)))))
		    (dboard:tabdat-allruns-set! tabdat new-res)
		    maxtests)
		  (if (> (dboard:rundat-run-data-offset run-struct) 0)
		      (loop run tal new-res newmaxtests) ;; not done getting data for this run
		      (loop (car tal)(cdr tal) new-res newmaxtests)))))))
    (dboard:tabdat-filters-changed-set! tabdat #f)

   exn
   (begin
     (debug:print 2 *default-log-port* "WARNING: error in accessing databases in get-youngest-run-db-mod-time: "
		  ((condition-property-accessor 'exn 'message) exn) " db-dir="dbdir ", exn=" exn)
     (current-seconds)) ;; something went wrong - just print an error and return current-seconds
   (common:max (map (lambda (filen)
		      (file-modification-time filen))
		    (cons (conc dbdir "/main.db") (glob (conc dbdir "/?.db")))))))

(define (dashboard:monitor-changed? commondat tabdat)
  (let* ((run-update-time (current-seconds))
	 (monitor-db-path (dboard:tabdat-monitor-db-path tabdat))
	 (monitor-modtime (if (and monitor-db-path (common:file-exists? monitor-db-path))
			      (file-modification-time monitor-db-path)
			      -1)))

	   (debug:print-error 0 *default-log-port* " query " stmt " failed, params: " params ", error: " ((condition-property-accessor 'exn 'message) exn) ", exn=" exn)
	   (print-call-chain (current-error-port))
	   default)))
   (apply sqlite3:first-result db stmt params)))

(define (db:setup do-sync)
  (assert *toppath* "FATAL: db:setup called before launch:setup has been run.")
  (let* ((tmpdir (common:make-tmpdir-name *toppath* "")))
    (if (not *dbstruct-dbs*)
	(dbfile:setup do-sync *toppath* tmpdir)
	*dbstruct-dbs*)))

;; moved from dbfile
;;
;; ADD run-id SUPPORT

  (debug:print-error 0 *default-log-port* "   params: " params
		     ", error: "     ((condition-property-accessor 'exn 'message)   exn)
		     ", arguments: " ((condition-property-accessor 'exn 'arguments) exn)
		     ", location: "  ((condition-property-accessor 'exn 'location)  exn)
		     ))









(define (db:set-sync db)
  (let ((syncprag (configf:lookup *configdat* "setup" "sychronous")))
    (sqlite3:execute db (conc "PRAGMA synchronous = " (or syncprag 0) ";")))) 


(define (db:get-last-update-time db)
  (let ((last-update-time #f))

    (max (get-mtime fname)
	 (get-mtime wal-file)
	 (get-mtime shm-file))))
	 
;; (define (db:all-db-sync dbstruct)
;;   (let* ((dbdat (db:open-db dbstruct #f db:initialize-main-db))
;; 	 (data-synced       0) ;; count of changed records
;;     (tmp-area       (common:make-tmpdir-name *toppath*))
;;     (dbfiles        (glob (conc tmp-area"/.mtdb/*.db")))
;;     (sync-durations (make-hash-table))
;;     (no-sync-db        (db:open-no-sync-db)))
;;     (for-each
;;      (lambda (file) ;; tmp db file
;;        (debug:print-info 3 *default-log-port* "file: " file)
;;        (let* ((fname       (conc (pathname-file file) ".db")) ;; fname is tmp db file

;;  'closeall     - close all opened dbs
;;  'schema       - attempt to apply schema changes
;;  run-ids: '(1 2 3 ...) or #f (for all)
;;
(define (db:multi-db-sync dbstruct . options)
  (let* (;; (dbdat       (db:open-db dbstruct #f dbfile:db-init-proc))
	 (data-synced 0) ;; count of changed records
	 (tmp-area       (common:make-tmpdir-name *toppath* ""))
	 (old2new (member 'old2new options))
	 (dejunk (member 'dejunk options))
	 (killservers (member 'killservers options))
	 (src-area (if old2new *toppath* tmp-area))
	 (dest-area (if old2new tmp-area *toppath*))
	 (dbfiles        (if old2new (glob (conc *toppath* "/.mtdb/*.db"))
			     (glob (conc tmp-area "/.mtdb/*.db"))))

;;======================================================================
;; no-sync.db - small bits of data to be shared between servers
;;======================================================================

(define (db:get-dbsync-path)
  (case (rmt:transport-mode)
    ((http)(common:make-tmpdir-name *toppath* ""))
    ((tcp) (conc *toppath*"/.mtdb"))
    ((nfs) (conc *toppath*"/.mtdb"))
    (else "/tmp/dunno-this-gonna-exist")))

;; This is needed for api.scm
(define (db:open-no-sync-db)
   (dbfile:open-no-sync-db (db:get-dbsync-path)))

    res))

;; TODO: Switch this to use max(update_time) from each run db? Then if using a server there is no disk traffic (using cachedb db) ???
;;
;; NOTE: This DOESN'T (necessarily) get the real run ids, but the number of the <number>.db!!

(define (db:get-changed-run-ids since-time)
  (let* ((dbdir      (common:make-tmpdir-name *toppath* "")) ;; (configf:lookup *configdat* "setup" "dbdir"))
	 (alldbs     (glob (conc *toppath* "/.mtdb/[0-9]*.db*")))
	 (changed    (filter (lambda (dbfile)
			       (> (file-modification-time dbfile) since-time))
			     alldbs)))
    (delete-duplicates
     (map (lambda (dbfile)
	    (let* ((res (string-match ".*\\/(\\d+)\\.db*" dbfile)))

          (debug:print 0 *default-log-port* "could not get lock for " from-db " from no-sync-db")
	  #f
        ))))

;; sync for filesystem local db writes
;;
(define (db:run-lock-and-sync no-sync-db)
  (let* ((tmp-area       (common:make-tmpdir-name *toppath* ""))
	 (dbfiles        (glob (conc tmp-area"/.mtdb/*.db")))
	 (sync-durations (make-hash-table)))
    ;; (debug:print-info 0 *default-log-port* "lock-and-sync, dbfiles: "dbfiles)
    (for-each
     (lambda (file)
       (let* ((fname (conc (pathname-file file) ".db"))
	      (fulln (conc *toppath*"/.mtdb/"fname))

  (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))     ;; last time through the sync loop
 	(no-sync-db         (db:open-no-sync-db))
 	(sync-duration      0)  ;; run time of the sync in milliseconds
	(tmp-area           (common:make-tmpdir-name *toppath* "")))
    ;; Sync moved to http-transport keep-running loop
    (debug:print-info 2 *default-log-port* "Periodic copy-based sync thread started. syncer is copy-sync, tmp-area is " tmp-area)
    (debug:print-info 3 *default-log-port* "watchdog starting. syncer is copy-sync pid="(current-process-id));;  " this-wd-num="this-wd-num)
    
    (if (and legacy-sync (not *time-to-exit*))
 	(begin
 	  (debug:print-info 0 *default-log-port* "Server running, periodic copy-based sync started.")

 		      ;;
 		     
 		       (for-each
 			(lambda (subdb)
 			  (let* (;;(dbstruct (db:setup))
 				 (mtdb       (dbr:subdb-mtdb subdb))
 				 (mtpath     (db:dbdat-get-path mtdb))
 				 (tmp-area   (common:make-tmpdir-name *toppath* ""))
 				 (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*)

           #t
          )
          #f
  )
)

(define (dbfile:make-tmpdir-name areapath tmpadj)
  (let* ((area (pathname-file areapath))
         (dname (conc "/tmp/"(current-user-name)"/megatest_localdb/" area "/" (string-translate areapath "/" ".") tmpadj "/.mtdb")))
    (unless (directory-exists? dname)
      (create-directory dname #t))
    dname))

(define (dbfile:run-id->path apath run-id)
  (conc apath"/"(dbfile:run-id->dbname run-id)))

;; opens and returns handle and nothing else
;;
;; NOTE: this is already protected by mutex *no-sync-db-mutex*
;;
(define (dbfile:raw-open-no-sync-db dbpath)
  (if (not (file-exists? dbpath))
      (create-directory dbpath #t))
  (debug:print-info 2 *default-log-port* "(dbfile:raw-open-no-sync-db: Opening "dbpath"/no-sync.db")
  (let* ((dbname    (conc dbpath "/no-sync.db"))
	 (db-exists (file-exists? dbname))
	 (init-proc (lambda (db)
		      (sqlite3:with-transaction
		       db
		       (lambda ()
			 ;; I have been having trouble with init of no-sync.db so

;; transaction protected lock aquisition
;; either:
;;    fails    returns  (#f lock-creation-time identifier)
;;    succeeds (returns (#t lock-creation-time identifier)
;; use (db:no-sync-del! db keyname) to release the lock
;;
(define (db:no-sync-get-lock-with-id db keyname identifier)
  (debug:print 2 *default-log-port* "db:no-sync-get-lock-with-id: db: " db " keyname: " keyname " identifier: " identifier)
  (sqlite3:with-transaction
   db
   (lambda ()
     (condition-case
      (let* ((curr-val (db:no-sync-get/default db keyname #f)))
        (debug:print 2 *default-log-port* "db:no-sync-get-lock-with-id: curr-val: " curr-val)
	(if curr-val
	    (match (db:extract-time-identifier curr-val) ;; result->timestamp, identifier
	       ((timestamp . ident)
		(cons (equal? ident identifier) timestamp))
	       (else 
                (debug:print 2 *default-log-port* "db:no-sync-get-lock-with-id: malformed lock")
                (cons #f 'malformed-lock)
                )
            )  ;; lock malformed
	    (let ((curr-sec (current-seconds))
		  (lock-value (if identifier
				  (conc (current-seconds)"+"identifier)
				  (current-seconds))))
	      (sqlite3:execute db "INSERT OR REPLACE INTO no_sync_metadat (var,val) VALUES(?,?);" keyname lock-value)
	      (cons #t curr-sec))))
      (exn (io-error)  (dbfile:print-err "ERROR: i/o error with no-sync db. Check permissions, disk space etc. and try again."))

			    (tmpadj  "")       ;; add to tmp path
			    (syncdir 'todisk)) ;; todisk is used when caching in /tmp and writing data back to MTRAH
  (let* ((dbstruct     (or dbstruct-in (make-dbr:dbstruct areapath: areapath)))
	 (dbfname      (or dbfname-in (dbmod:run-id->dbfname run-id)))
	 (dbpath       (dbmod:get-dbdir dbstruct))             ;; directory where all the .db files are kept
	 (dbfullname   (conc dbpath"/"dbfname)) ;; (dbmod:run-id->full-dbfname dbstruct run-id))
	 (dbexists     (file-exists? dbfullname))
	 (tmpdir       (common:make-tmpdir-name areapath tmpadj))
	 (tmpdb        (let* ((fname (conc tmpdir"/"dbfname)))
			 fname))
	 (cachedb        (dbmod:open-cachedb-db init-proc
					    ;; (if (eq? (dbfile:cache-method) 'cachedb)
					    ;; 	#f
					    tmpdb
					    ;; )

;;     along with Megatest.  If not, see <http://www.gnu.org/licenses/>.

;; Always use two or four digit decimal
;; 1.01, 1.02...1.10,1.11,1.1101 ... 1.99,2.00..

;; (declare (unit megatest-version))

(define megatest-version 1.8018)

    (let* (;; (run-id     (args:get-arg "-run-id"))
	   (dbfname    (args:get-arg "-db"))
	   (tl         (launch:setup))
	   (keys       (keys:config-get-fields *configdat*)))
      (case (rmt:transport-mode)
	((tcp)
	 (let* ((timeout    (server:expiration-timeout)))
	   (debug:print 0 *default-log-port* "INFO: Starting server for " dbfname " using tcp method with server timeout of "timeout)
	   (tt-server-timeout-param timeout)
	   (if dbfname
	       (tt:start-server tl #f dbfname api:tcp-dispatch-request-make-handler keys)
	       (begin
		 (debug:print 0 *default-log-port* "ERROR: transport mode is tcp - -db is required.")
		 (exit 1)))))
	(else (debug:print 0 *default-log-port* "ERROR: rmt:transport-mode value not recognised "(rmt:transport-mode))))

                ) 
              )
              sfiles
            )
          )
       )
       dbfiles
     )
     ;; remove this db, because otherwise metadata contains records for old servers, and this causes a problem with db:no-sync-get-lock-with-id.
     (if (file-exists? (conc *toppath* "/.mtdb/no-sync.db"))
       (delete-file (conc *toppath* "/.mtdb/no-sync.db"))
     )
     (set! *didsomething* #t)
     (exit)  
  )
)

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

             (begin 
             (debug:print-info 1 *default-log-port* "Missing required argument -source <archive path>")
             (exit 1)))
         (if (common:file-exists? (conc  *toppath* "/megatest.db"))
             (begin  
               (debug:print-info 1 *default-log-port* "File " (conc  *toppath* "/megatest.db") " already exists. Please remove it before trying to replicate db")
               (exit 1)))
         (if (and (common:make-tmpdir-name *toppath* "") (> (length (directory   (common:make-tmpdir-name *toppath* "") #f)) 0))
           (begin
           (debug:print-info 1 *default-log-port* (common:make-tmpdir-name *toppath* "") " not empty. Please remove it before trying to replicate db")
           (exit 1)))    
          ;; check if timestamp 
          (let* ((source (args:get-arg "-source"))
                (src     (if (not (equal? (substring source 0 1) "/"))
                             (conc (current-directory) "/" source)
                             source))
                (ts (if (args:get-arg "-time-stamp")   (args:get-arg "-time-stamp") "latest")))

     ((tcp) (tt:make-remote areapath))
     (else #f)))

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

(define *send-receive-mutex* (make-mutex)) ;; should have separate mutex per run-id


;; how to make area-dat


(define (rmt:set-ttdat areapath ttdat)
  (if (not ttdat)



      (let* ((newremote  (make-and-init-remote areapath)))




	(set! *ttdat* newremote)











	ttdat)))




;; NB// area-dat replaced by ttdat
;; 
(define (rmt:send-receive cmd run-id params #!key (attemptnum 1)(ttdat #f))
  (assert *toppath* "FATAL: rmt:send-receive called with *toppath* not set.")
  (assert ttdat     "FATAL: rmt:send-receive must receive initialized area-dat")
  (let* ((start-time    (current-seconds)) ;; snapshot time so all use cases get same value
         (areapath      *toppath*);; TODO - resolve from dbstruct to be compatible with multiple areas


         (attemptnum    (+ 1 attemptnum))
	 (readonly-mode (rmtmod:calc-ro-mode ttdat *toppath*))
	 (testsuite     (common:get-testsuite-name))
	 (mtexe         (common:find-local-megatest))
	 (dbfname       (conc (dbfile:run-id->dbnum run-id)".db")))

    (rmt:set-ttdat areapath ttdat)
    (tt:handler (rmt:set-ttdat ttdat) cmd run-id params
		attemptnum readonly-mode dbfname
		testsuite mtexe)))

;; KEEP THIS HERE FOR A WHILE, WE MAY WANT TO RESURECT IT
;; (define (nfs-transport-handler  runremote cmd run-id params attemptnum area-dat areapath readonly-mode testsuite mtexe)
;;   (let* ((keys     (common:get-fields *configdat*))
;; 	 (dbstruct (dbmod:nfs-get-dbstruct run-id keys (dbfile:db-init-proc) areapath tmpadj: "/dashboard")))
;;     (api:dispatch-request dbstruct cmd run-id params)))








	
(define (rmt:print-db-stats)
  (let ((fmtstr "~40a~7-d~9-d~20,2-f")) ;; "~20,2-f"
    (debug:print 18 *default-log-port* "DB Stats\n========")
    (debug:print 18 *default-log-port* (format #f "~40a~8a~10a~10a" "Cmd" "Count" "TotTime" "Avg"))
    (for-each (lambda (cmd)
		(let ((cmd-dat (hash-table-ref *db-stats* cmd)))

				 (cons 'none 0))
			     (loop (car tal)(cdr tal) newmax-cmd currmax)))))))
    (mutex-unlock! *db-stats-mutex*)
    res))

(define (rmt:open-qry-close-locally cmd run-id params #!key (remretries 5))
  (let* ((qry-is-write    (not (member cmd api:read-only-queries)))
	 (db-file-path    (common:make-tmpdir-name *toppath* "")) ;;  0))
	 (dbstructs-local (db:setup #t))
	 (read-only       (not (file-write-access? db-file-path)))
	 (start           (current-milliseconds))
	 (resdat          (if (not (and read-only qry-is-write))
			      (let ((v (api:execute-requests dbstructs-local (vector (symbol->string cmd) params))))
			;;	(handle-exceptions ;; there has been a long history of receiving strange errors from values returned by the client when things go wrong..
			;;	 exn               ;;  This is an attempt to detect that situation and recover gracefully

	      #f))
	(begin
	  ;; (rmt:update-db-stats run-id cmd params duration)
	  ;; mark this run as dirty if this was a write, the watchdog is responsible for syncing it
	  (if qry-is-write
	      (let ((start-time (current-seconds)))
		(mutex-lock! *db-multi-sync-mutex*)
		(set! *db-last-access* start-time)  ;; THIS IS PROBABLY USELESS? (we are on a client)
                (mutex-unlock! *db-multi-sync-mutex*)))))
    res))

;;======================================================================
;;
;; A C T U A L   A P I   C A L L S  
;;

	     (debug:print 5 *default-log-port* " exn=" (condition->list exn))
	     (thread-sleep! 1)
	     (tasks:open-db numretries (- numretries 1)))
	   (begin
	     (print-call-chain (current-error-port))
	     (debug:print 0 *default-log-port* " message: " ((condition-property-accessor 'exn 'message) exn))
	     (debug:print 5 *default-log-port* " exn=" (condition->list exn))))
       (let* ((dbpath        (common:make-tmpdir-name *toppath* "")) ;; (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       (common: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))

;; (max-connections 4096) 

;; do all the busy work of finding and setting up conn for
;; connecting to a server
;; 
(define (tt:client-connect-to-server ttdat dbfname run-id testsuite)
  (assert (tt:valid-run-id run-id dbfname) "FATAL: invalid run-id "run-id)
  (debug:print-info 2 *default-log-port* "tt:client-connect-to-server " dbfname " " run-id)
  (let* ((conn (hash-table-ref/default (tt-conns ttdat) dbfname #f))
	 (server-start-proc (lambda ()
			      (print "dbfname: "dbfname)
			      (assert (equal? dbfname "main.db") ;; only main.db is started here
				      "FATAL: called server-start-proc for db other than main.db")
			      (tt:server-process-run
			       (tt-areapath ttdat)
			       testsuite ;; (dbfile:testsuite-name)
			       (common:find-local-megatest)
			       run-id))
			    ))
    (if conn
	(begin 
          (debug:print-info 2 *default-log-port* "already connected to a server")
           conn) ;; we are already connected to the server
	(let* ((sdat (tt:get-current-server-info ttdat dbfname)))
	  (match sdat
	    ((host port start-time server-id pid dbfname2 servinffile)
	     (assert (equal? dbfname dbfname2) "FATAL: read server info from wrong file.")
             (debug:print-info 2 *default-log-port* "no conn - in match servinffile:" servinffile)
	     (let* ((host-port (conc host":"port))
		    (conn (make-tt-conn
			   host: host
			   port: port
			   host-port: host-port
			   dbfname: dbfname
			   servinf-file: servinffile
			   server-id: server-id
			   server-start: start-time
			   pid: pid)))
	       ;; verify we can talk to this server
	       (let* ((result   (tt:timed-ping host port server-id))
		      (ping-res (car result))
		      (ping     (cdr result)))
                 (debug:print-info 2 *default-log-port* "host " host " port " port " ping time: " ping " result " ping-res)
		 (case ping-res
		   ((running)
                    (debug:print-info 2 *default-log-port* "Setting conn = " conn " in hash table")
		    (hash-table-set! (tt-conns ttdat) dbfname conn) ;;; is this ok to save before validating that the connection is good?
		    conn)
		   ((starting)
		    (thread-sleep! 0.5)
                    (debug:print-info 0 *default-log-port* "server for " dbfname " is in starting state, retrying connect")
		    (tt:client-connect-to-server ttdat dbfname run-id testsuite))
		   (else
		    (let* ((curr-secs (current-seconds)))
		      ;; rm the (last server) would go here
		      (if (> (- curr-secs (tt-last-serv-start ttdat)) 10)
			  (begin
			    (tt-last-serv-start-set! ttdat curr-secs)
			    (server-start-proc))) ;; start server if 10 sec since last attempt
		      (thread-sleep! 1)
                      (debug:print-info 2 *default-log-port* "server ping result was neither running nor starting. Retrying connect")
		      (tt:client-connect-to-server ttdat dbfname run-id testsuite)))))))
	    (else ;; no good server found, if haven't started server in > 5 secs, start another
	     (if (> (- (current-seconds) (tt-last-serv-start ttdat)) 3) ;; BUG - grow this number really do not want to swamp the machine with servers
		 (begin
		   (debug:print-info 0 *default-log-port* "Starting server for "dbfname)
		   (server-start-proc)
		   (tt-last-serv-start-set! ttdat (current-seconds))
                   (thread-sleep! 3)
                   ))
	     (thread-sleep! 1)
             (debug:print-info 0 *default-log-port* "Connect to server for " dbfname)
	     (tt:client-connect-to-server ttdat dbfname run-id testsuite)))))))

(define (tt:timed-ping host port server-id)
  (let* ((start-time (current-milliseconds))
	 (result     (tt:ping host port server-id)))
    (cons result (- (current-milliseconds) start-time))))
    

       ;; (debug:print 0 *default-log-port* "res not in form (status errmsg result meta), got: "res)
       (try-again)))))

;; client side handler
;;
;;(tt:handler #<tt> get-keys #f () 2 #f "/home/matt/data/megatest/ext-tests" #f "main.db" "ext-tests" "/home/matt/data/megatest/bin/.22.04/../megatest")
;;
(define (tt:handler ttdat cmd run-id params attemptnum readonly-mode dbfname testsuite mtexe)
  ;; connect-to-server will start a server if needed.
  (let* ((areapath (tt-areapath ttdat))
	 (conn     (tt:client-connect-to-server ttdat dbfname run-id testsuite))) ;; looks up conn keyed by dbfname
    (if conn
	;; have connection, call the server
	(let* ((res (tt:send-receive ttdat conn cmd run-id params)))
	  ;; res is (status errmsg result meta)

	  (match res
	    ((status errmsg result meta)
	     (if (list? meta)
		 (let* ((delay-wait (alist-ref 'delay-wait meta)))
		   (if (and (number? delay-wait)
			    (> delay-wait 0))
		       (begin
			 (debug:print 0 *default-log-port* "Server is loaded, delaying "delay-wait" seconds")
			 (thread-sleep! delay-wait)))))
	     (case status
	       ((busy) ;; result will be how long the server wants you to delay
		(let* ((dly  (if (number? result) result 0.1)))
		  (debug:print 0 *default-log-port* "WARNING: server for "dbfname" is busy, will try again in "dly" seconds.")
		  (thread-sleep! dly)
		  (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe)))
	       ((loaded)
		(debug:print 0 *default-log-port* "WARNING: server for "dbfname" is loaded, slowing queries.")
		(tt:backoff-incr (tt-conn-host conn)(tt-conn-port conn))
		result) ;; (tt:handler  ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))
	       (else
		result)))
	    (else ;; did not receive properly formated result
	     (if (not res) ;; tt:send-receive telling us that communication failed
		 (let* ((host    (tt-conn-host conn))
			(port    (tt-conn-port conn))
			;; (dbfname (tt-conn-port conn)) ;; 192.168.0.127:4242-726924:4.db
			(pid     (tt-conn-pid  conn))
                        ;;(servinf (tt-conn-servinf-file conn))) 
			(servinf (tt-servinf-file ttdat))) ;; (conc areapath"/.servinfo/"host":"port"-"pid":"dbfname))) ;; TODO, use (server:get-servinfo-dir areapath)
		   (hash-table-set! (tt-conns ttdat) dbfname #f)
		   (if (and servinf (file-exists? servinf))
		       (begin
			 (if (< attemptnum 10)
			     (begin
			       (thread-sleep! 0.5)
			       (tt:handler ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))
			     (begin
			       (debug:print 0 *default-log-port* "INFO: no response from server "host":"port" for "dbfname)
			       (if (and (file-exists? servinf)
					(> (- (current-seconds)(file-modification-time servinf)) 60))
				   (begin
				     (debug:print 0 *default-log-port* "INFO: "servinf" file seems old and no ping response, removing it.")
				     (handle-exceptions
					 exn
				       #f
				       (delete-file* servinf))
				     (tt:handler ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))
				   (begin
				     ;; start server - addressed in client-connect-to-server
				     ;; delay        - addressed in client-connect-to-server
				     ;; try again
				     (thread-sleep! 0.25) ;; dunno, I think this needs to be here
				     (tt:handler ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))
				   ))))
		       (begin ;; no server file, delay and try again
			 (debug:print 2 *default-log-port* "INFO: connection to server "host":"port" broken for "dbfname", no servinf file. Server exited? ")
			 (thread-sleep! 0.5)
			 (tt:handler ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe))))
		 (begin ;; this case is where res is malformed. Probably should abort
		   (assert #f "FATAL: tt:handler received bad data "res)
		   ;; (debug:print 0 *default-log-port* "INFO: got corrupt data from server "host":"port", "res", for "dbfname", will try again.")
		   ;; (tt:handler ttdat cmd run-id params (+ attemptnum 1) readonly-mode dbfname testsuite mtexe)
		   )))))
	(begin
	  (thread-sleep! 1) ;; no conn yet set up, give it a rest and try again
	  (tt:handler ttdat cmd run-id params attemptnum readonly-mode dbfname testsuite mtexe)))))

(define (tt:bid-for-servership run-id)
  #f)

;; gets server info and appends path to server file
;; sorts by age, oldest first
;;

;; This is the routine called in megatest.scm to start a server
;;
;; Server viability is checked in keep-running. Blindly start and run here.
;;
(define (tt:start-server areapath run-id dbfname-in handler keys)
  (assert areapath "FATAL: areapath not provided for tt:start-server")
  ;; is there already a server for this dbfile? Then exit.
  (debug:print 2 *default-log-port* "tt:start-server: " dbfname-in)
  (let* ((ttdat   (make-tt areapath: areapath))
	 (dbfname (or dbfname-in (dbmod:run-id->dbfname run-id)))
	 (servers (tt:find-server areapath dbfname))) ;; should use tt:get-current-server-info instead
         (debug:print 0 *default-log-port* "Found " (length servers) " already running for " dbfname)
    (if (> (length servers) 0)
	(begin
	  (debug:print 0 *default-log-port* "INFO: found server(s) already running for db "dbfname", "(string-intersperse servers ",")" Exiting.")
	  (exit))
	(let* ((dbstruct   (dbmod:open-dbmoddb areapath run-id dbfname (dbfile:db-init-proc) keys)))
	  (tt-handler-set! ttdat (handler dbstruct))
	  (let* ((tcp-thread (make-thread
			      (lambda ()

		    (if (tt-cleanup-proc ttdat)
			((tt-cleanup-proc ttdat)))
		    (dbfile:with-no-sync-db nosyncdbpath
					    (lambda (db)
					      (let* ((dbtmpname (dbr:dbstruct-dbtmpname dbstruct)))
						(debug:print-info 0 *default-log-port* "Running clean up, including removing db file "dbtmpname)
						(db:no-sync-del! db dbfname)

                                                ))))))
    (set! *server-info* ttdat)
    (let loop ((count 0))
      (if (> count 240)
	  (begin
	    (debug:print 0 *default-log-port* "FATAL: Could not start a tcp server, giving up.")
	    (exit 1))
	  (if (not (tt-port ttdat)) ;; no connection yet

						(debug:print 0 *default-log-port* "Failed to get server lock for "dbfname)
						#f))))))
			  (if (and res (common:low-noise-print 120 "top server message"))
			      (debug:print-info 0 *default-log-port* "Keep running, I'm the top server for "
						dbfname" on "(tt-host ttdat)":"(tt-port ttdat)))
			  res))
		       (else
                        ;; wrong servinfo file
			(debug:print-info 0 *default-log-port* "I'm not the lead server: "servers)
			(let* ((leadsrv (car servers)))
			  (match leadsrv
			    ((host port startseconds server-id pid dbfname servinfofile)
			     (let* ((result  (tt:timed-ping host port server-id))
				    (res     (car result))
				    (ping    (cdr result)))

;; get servers listed, last part of name must match :<dbfname>
;; if more than one, wait one second and look again
;; future: ping oldest, if alive remove other :<dbfname> files
;;
(define (tt:find-server areapath dbfname)
  (let* ((servdir  (tt:get-servinfo-dir areapath))
	 (sfiles   (glob (conc servdir"/*:"dbfname))))
    
    ;; filter the files here by looking in processes table (if we are not main.db)
    ;; and or look at the time stamp on the servinfo file, a running server will
    ;; touch the file every minute (again, this will only apply for main.db)
    
    sfiles))

;; given a path to a server info file return: host port startseconds server-id pid dbfname logf
;; example of what it's looking for in the log file:
;;     SERVER STARTED: 10.38.175.67:50216 AT 1616502350.0 server-id: 4907e90fc55c7a09694e3f658c639cf4 
;;
(define (tt:server-get-info logf)
  (let ((server-rx    (regexp "^SERVER STARTED: (\\S+):(\\d+) AT ([\\d\\.]+) server-id: (\\S+) pid: (\\d+) dbfname: (\\S+)")) ;; SERVER STARTED: host:port AT timesecs server id
        (dbprep-rx    (regexp "^SERVER: dbprep"))
        (dbprep-found 0)
	(bad-dat      (list #f #f #f #f #f #f logf)))
     (let ((fdat     (handle-exceptions
			 exn
		       (begin
			 ;; BUG, TODO: add err checking, for now blanket ignore the errors?
			 (debug:print-info 0 *default-log-port* "Unable to get server info from "logf
					   ", exn="(condition->list exn))
			 '()) ;; no idea what went wrong, call it a bad server, return empty list
		       (with-input-from-file logf read-lines))))
       (if (null? fdat) ;; bad data, return bad-dat
	   bad-dat
	   (let loop ((inl  (car fdat))
		      (tail (cdr fdat))
		      (lnum 0))

(define  (tt:server-process-run areapath testsuite mtexe run-id #!key (profile-mode "")) ;; areapath is *toppath* for a given testsuite area
  (assert areapath  "FATAL: tt:server-process-run called without areapath defined.")
  (assert testsuite "FATAL: tt:server-process-run called without testsuite defined.")
  (assert mtexe     "FATAL: tt:server-process-run called without mtexe defined.")
  ;; mtest -server - -m testsuite:ext-tests -db 6.db
  (let* ((dbfname  (dbmod:run-id->dbfname run-id))
	 (load     (get-normalized-cpu-load))
	 (srvrs    (tt:find-server areapath dbfname))
	 (trying   (length srvrs))
	 (nrun     (number-of-processes-running (conc "mtest.*server.*"testsuite".*"dbfname))))
    (cond
     ((> load 2.0)
      (debug:print 0 *default-log-port* "Normalized load "load" on " (get-host-name) " is over the limit of 2.0. Not starting a server. Please reduce the load on "(get-host-name)" by killing some processes")
      (thread-sleep! 1)
      #f)
     ((> nrun 100)
      (debug:print 0 *default-log-port* nrun" servers running on " (get-host-name) ", not starting another.")
      (thread-sleep! 1)
      #f)
     ((> trying 2)
      (debug:print 0 *default-log-port* trying" servers registered in .servinfo dir. not starting another.")
      (thread-sleep! 1)
      #f)
     (else
      (if (not (file-exists? (conc areapath"/logs")))
	      (create-directory (conc areapath"/logs") #t))
	  (let* ((logfile   (conc areapath "/logs/server-"dbfname"-"(current-process-id)".log")) ;; -" curr-pid "-" target-host ".log"))
		 (cmdln     (conc
			     mtexe
			     " -startdir "areapath
			     " -server - ";; (or target-host "-")
			     " -m testsuite:"testsuite
			     " -db "dbfname ;; (dbmod:run-id->dbfname run-id)
			     " " profile-mode
			     (conc " >> " logfile " 2>&1 &"))))
	    ;; we want the remote server to start in *toppath* so push there
	    ;; (push-directory areapath) ;; use cd in the command line instead
	    (debug:print 2 *default-log-port* "INFO: Trying to start server in tcp mode (" cmdln ") at "(common:human-time)" for "areapath)
	    ;; (debug:print 0 *default-log-port* "INFO: starting server at " (common:human-time))

	    (system cmdln)
	    ;; ;; use below to go back to nbfake - nbfake does cause trouble ...
	    ;; (setenv "NBFAKE_QUIET" "yes") ;; BUG: change to with-environment-variable ...
	    ;; (setenv "NBFAKE_LOG" logfile)
	    ;; (system (conc "cd "areapath" ; nbfake " cmdln))
	    ;; (unsetenv "NBFAKE_QUIET")
	    ;; (unsetenv "NBFAKE_LOG")
	    
	    ;;(pop-directory)
	    #t)))))

;;======================================================================
;; tcp connection stuff
;;======================================================================

;; find a port and start tcp-server. This only starts the tcp portion of
;; the server, look at (tt:start-server ...) above for the entry point

#     along with Megatest.  If not, see <http://www.gnu.org/licenses/>.

MT_TMPDISPLAY=$DISPLAY
MT_TMPUSER=$USER
MT_HOME=$HOME

tmpfile=`mktemp`
if [[ -e megatest.sh ]]; then
  grep -v "export USER=" megatest.sh | grep -v "export HOME=" > $tmpfile
  source $tmpfile
  rm $tmpfile
fi




export DISPLAY=$MT_TMPDISPLAY
export USER=$USER
export HOME=$MT_HOME

if [ x"$MT_XTERM_CMD" == "x" ];then
  exec xterm "$@"
else