Megatest

(define (common:get-num-cpus remote-host)
  (let* ((actual-host (or remote-host (get-host-name))))
    (or (common:get-cached-info actual-host "num-cpus" age: (+ 2592000 (random 3600))) ;; hosts had better not be changing the number of cpus too often!
	(let* ((proc   (lambda ()
			 (let loop ((numcpu 0)
				    (inl    (read-line)))
			   (if (eof-object? inl)

			       (if (> numcpu 0)
				   numcpu
				   #f) ;; if zero return #f so caller knows that things are not working
			       (loop (if (string-match "^processor\\s+:\\s+\\d+$" inl)
					 (+ numcpu 1)
					 numcpu)
				     (read-line))))))
	       (result (if remote-host
			   (with-input-from-pipe 
			       (conc "ssh " remote-host " cat /proc/cpuinfo")
			     proc)
			   (with-input-from-file "/proc/cpuinfo" proc))))
	  (if (> result 0)(common:write-cached-info actual-host "num-cpus" result))
	  result))))

;; wait for normalized cpu load to drop below maxload
;;
(define (common:wait-for-normalized-load maxload msg remote-host #!optional (rem-tries 5))
  (let ((num-cpus (common:get-num-cpus remote-host)))
    (if num-cpus
	(common:wait-for-cpuload maxload num-cpus 15 msg: msg remote-host: remote-host)
	(begin
	  (thread-sleep! (random 60)) ;; we failed to get num cpus. wait a bit and try again
	  (if (> rem-tries 0)
	      (common:wait-for-normalized-load maxload msg remote-host (- rem-tries 1))
	      #f)))))

;; DO NOT CALL THIS DIRECTLY. It is called from common:wait-for-normalized-load
;;
(define (common:wait-for-cpuload maxload-in numcpus-in waitdelay #!key (count 1000) (msg #f)(remote-host #f)(force-maxload #f))
  (let* ((loadavg (common:get-cpu-load remote-host))
	 (numcpus (if (<= 1 numcpus-in) ;; not possible to have zero.  If we get 1, it's possible that we got the previous default, and we should check again
		      (common:get-num-cpus remote-host)

           #f
           #t)))))

;; given a launch delay (minimum time from last launch) return amount of time to wait
;;
;; (define (db:launch-delay-left dbstruct run-id launch-delay)
  


(define (db:get-status-from-final-status-file run-dir)
  (let (
       (infile (conc run-dir "/.final-status")))

       ;; first verify we are able to write the output file
       (if (not (file-read-access? infile))
          (begin 
	        (debug:print 0 *default-log-port* "ERROR: cannot read " infile)
          (debug:print 0 *default-log-port* "ERROR: run-dir is " run-dir)
          #f
          )
          (with-input-from-file infile read-lines)
       )
  )
)




;;  select end_time-now from
;;      (select testname,item_path,event_time+run_duration as
;;                          end_time,strftime('%s','now') as now from tests where state in
;;      ('RUNNING','REMOTEHOSTSTART','LAUNCHED'));

(define (db:find-and-mark-incomplete dbstruct run-id ovr-deadtime)
  (let* ((incompleted '())
	 (oldlaunched '())
	 (toplevels   '())
          ;; The default running-deadtime is 720 seconds = 12 minutes.
          ;; "(running-deadtime-default (+ server-start-allowance (* 2 launch-monitor-period)))" = 200 + (2 * (200 + 30 + 30))
         (deadtime-trim (or ovr-deadtime (configf:lookup-number *configdat* "setup" "deadtime")))
         (server-start-allowance 200)
         (server-overloaded-budget 200)
         (launch-monitor-off-time (or (configf:lookup-number *configdat* "setup" "test-stats-update-period") 30))
         (launch-monitor-on-time-budget 30)
         (launch-monitor-period (+ launch-monitor-off-time launch-monitor-on-time-budget server-overloaded-budget))
         (remotehoststart-deadtime-default (+ server-start-allowance server-overloaded-budget 30))
         (remotehoststart-deadtime (or deadtime-trim remotehoststart-deadtime-default))
         (running-deadtime-default (+ server-start-allowance (* 2 launch-monitor-period)))
         (running-deadtime (or deadtime-trim running-deadtime-default)) ;; two minutes (30 seconds between updates, this leaves 3x grace period)
         )
    (debug:print-info 4  *default-log-port* "running-deadtime = " running-deadtime)
    (debug:print-info 4  *default-log-port* "deadtime-trim = " deadtime-trim)

    (db:with-db 
     dbstruct #f #f
     (lambda (db)
       ;; in RUNNING or REMOTEHOSTSTART for more than 10 minutes
       ;;
       ;; HOWEVER: this code in run:test seems to work fine
       ;;              (> (- (current-seconds)(+ (db:test-get-event_time testdat)

              (begin
                (set! toplevels   (cons (list test-id run-dir uname testname item-path run-id) toplevels))
                (debug:print-info 0 *default-log-port* "Found old toplevel test in RUNNING state, test-id=" test-id))
              (begin
                (set! incompleted (cons (list test-id run-dir uname testname item-path run-id) incompleted))
                (debug:print-info 0 *default-log-port* "Found old test in RUNNING state, test-id=" test-id" exceeded running-deadtime "running-deadtime" now="(current-seconds)" event-time="event-time" run-duration="run-duration))))
        db
        
        "SELECT id,rundir,uname,testname,item_path,event_time,run_duration FROM tests WHERE run_id=? AND (strftime('%s','now') - event_time) > (run_duration + ?) AND state IN ('RUNNING');"
        run-id running-deadtime) ;; default time 720 seconds

       
       (sqlite3:for-each-row 
        (lambda (test-id run-dir uname testname item-path event-time run-duration)
          (if (and (equal? uname "n/a")
                   (equal? item-path "")) ;; this is a toplevel test
              ;; what to do with toplevel? call rollup?
              (begin
                (set! toplevels   (cons (list test-id run-dir uname testname item-path run-id) toplevels))
                (debug:print-info 0 *default-log-port* "Found old toplevel test in RUNNING state, test-id=" test-id))
              (begin
                (debug:print-info 0 *default-log-port* "Found old test in REMOTEHOSTSTART state, test-id=" test-id" exceeded running-deadtime "running-deadtime" now="(current-seconds)" event-time="event-time" run-duration="run-duration)
                (set! incompleted (cons (list test-id run-dir uname testname item-path run-id) incompleted)))))
        db
        "SELECT id,rundir,uname,testname,item_path,event_time,run_duration FROM tests WHERE run_id=? AND (strftime('%s','now') - event_time) > (run_duration + ?) AND state IN ('REMOTEHOSTSTART');"
        run-id remotehoststart-deadtime) ;; default time 230 seconds.

       ;; in LAUNCHED for more than one day. Could be long due to job queues TODO/BUG: Need override for this in config
       ;;
       ;; (db:delay-if-busy dbdat)
       (sqlite3:for-each-row
        (lambda (test-id run-dir uname testname item-path)
          (if (and (equal? uname "n/a")

              ;;      		    incompleted))
              (min-incompleted-ids (map car incompleted)) ;; do 'em all
              (all-ids             (append min-incompleted-ids (map car oldlaunched))))
         (if (> (length all-ids) 0)
             (begin
               (debug:print 0 *default-log-port* "WARNING: Marking test(s); " (string-intersperse (map conc all-ids) ", ") " as DEAD")
               (for-each
                  (lambda (test-id)
                    (let* (
                      (run-dir (db:test-get-rundir-from-test-id dbstruct run-id test-id))
                      (result (db:get-status-from-final-status-file run-dir)))
                    (if (and (list? result) (> (length result) 1) (equal? "PASS" (cadr result)) (equal? "COMPLETED" (car result))) 
                      (begin
                        (debug:print 0 *default-log-port* "INFO: test " test-id " actually passed, so marking PASS not DEAD")
                        (db:set-state-status-and-roll-up-items dbstruct run-id test-id 'foo "COMPLETED" "PASS" "Test stopped responding but it has PASSED; marking it PASS in the DB.")
                      )
                      (begin
                      (debug:print 0 *default-log-port* "INFO: test " test-id " final state/status is not COMPLETED/PASS. It is " result)
                      (db:set-state-status-and-roll-up-items dbstruct run-id test-id 'foo "COMPLETED" "DEAD" "Test stopped responding while in RUNNING or REMOTEHOSTSTART; presumed dead.")
                      )
                     )
                    )
                  )
                  all-ids)
             )
         )
       )
     )
   )
 )
)

;; ALL REPLACED BY THE BLOCK ABOVE
;;
;; 	    (sqlite3:execute 
;; 	     db
;; 	     (conc "UPDATE tests SET state='INCOMPLETE' WHERE run_id=? AND id IN (" 
;; 		   (string-intersperse (map conc all-ids) ",")

  (let* ((stepname       (car ezstep))  ;; do stuff to run the step
	 (stepinfo       (cadr ezstep))
	;; (let ((info (cadr ezstep)))
	;; 		   (if (proc? info) "" info)))
	;; (stepproc       (let ((info (cadr ezstep)))
	;; 		   (if (proc? info) info #f)))
	 (stepparts      (string-match (regexp "^(\\{([^\\}\\{]*)\\}\\s*|)(.*)$") stepinfo))
	 (stepparams     (if (and (list? stepparts)
				  (> (length stepparts) 1))
			     (list-ref stepparts 2)
			     #f)) ;; for future use, {VAR=1,2,3}, run step for each
	 (paramparts     (if (string? stepparams)
			     (map (lambda (x)(string-split x "=")) (string-split-fields "[^;]*=[^;]*" stepparams))
			     '()))
	 (subrun         (alist-ref "subrun" paramparts equal?))
	 (stepcmd        (if (and (list? stepparts)
				  (> (length stepparts) 2))
			     (list-ref stepparts 3)
			     (conc "# error, no command for step "stepname)))
	 (script         "") ; "#!/bin/bash\n") ;; yep, we depend on bin/bash FIXME!!!\
	 (logpro-file    (conc stepname ".logpro"))
	 (html-file      (conc stepname ".html"))
	 (dat-file       (conc stepname ".dat"))
	 (tconfig-logpro (configf:lookup testconfig "logpro" stepname))
	 (logpro-used    (common:file-exists? logpro-file)))

	  ;; We are about to actually kick off the test
	  ;; so this is a good place to remove the records for 
	  ;; any previous runs
	  ;; (db:test-remove-steps db run-id testname itemdat)
	  ;; now is also a good time to write the .testconfig file
	  (let* ((tconfig-fname   (conc work-area "/.testconfig"))
		 (tconfig-tmpfile (conc tconfig-fname ".tmp"))
		 (tconfig         (tests:get-testconfig test-name item-path tconfigreg #t force-create: #t)) ;; 'return-procs)))
		 (scripts (configf:get-section tconfig "scripts")))
	    ;; create .testconfig file
	    (configf:write-alist tconfig tconfig-tmpfile)
	    (file-move tconfig-tmpfile tconfig-fname #t)
	    (delete-file* ".final-status")

	    ;; extract scripts from testconfig and write them to files in test run dir
	    (for-each
	     (lambda (scriptdat)
	       (match scriptdat
		      ((name content)
		       (with-output-to-file name
			 (lambda ()
			   (print content)
			   (change-file-mode name (bitwise-ior perm/irwxg perm/irwxu)))))
		      (else
		       (debug:print-info 0 "Invalid script definiton found in [scripts] section of testconfig. \"" scriptdat "\""))))
	     scripts))
	  ;; 
	  (let* ((m            (make-mutex))
		 (kill-job?    #f)
		 (exit-info    (make-launch:einf pid: #t exit-status: #t exit-code: #t rollup-status: 0)) ;; pid exit-status exit-code (i.e. process was successfully run) rollup-status
		 (job-thread   #f)
		 ;; (keep-going   #t)
		 (misc-flags   (let ((ht (make-hash-table)))

				      (if (equal? (db:test-get-status testinfo) "AUTO") "AUTO-WARN" "WARN"))
				     ((eq? (launch:einf-rollup-status exit-info) 3) "CHECK")
				     ((eq? (launch:einf-rollup-status exit-info) 4) "WAIVED")
				     ((eq? (launch:einf-rollup-status exit-info) 5) "ABORT")
				     ((eq? (launch:einf-rollup-status exit-info) 6) "SKIP")
				     (else "FAIL")))) ;; (db:test-get-status testinfo)))
		    (debug:print-info 1 *default-log-port* "Test exited in state=" (db:test-get-state testinfo) ", setting state/status based on exit code of " (launch:einf-exit-status exit-info) " and rollup-status of " (launch:einf-rollup-status exit-info))
   
        ;; Leave a .final-status file for each sub-test
        (tests:save-final-status run-id test-id)

		    (tests:test-set-status! run-id 
					    test-id 
					    new-state
					    new-status
					    (args:get-arg "-m") #f)
		    ;; need to update the top test record if PASS or FAIL and this is a subtest
		    ;; NO NEED TO CALL set-state-status-and-roll-up-items HERE, THIS IS DONE IN set-state-status-and-roll-up-items called by tests:test-set-status!
		    ))
	      ;; for automated creation of the rollup html file this is a good place...
	      (if (not (equal? item-path ""))
		      (tests:summarize-items run-id test-id test-name #f))
	      (tests:summarize-test run-id test-id)  ;; don't force - just update if no
        ;; Leave a .final-status file for the top level test
        (tests:save-final-status run-id test-id)
	      (rmt:update-run-stats run-id (rmt:get-raw-run-stats run-id)))
	    (mutex-unlock! m)
            (launch:end-of-run-check run-id )
	    (debug:print 2 *default-log-port* "Output from running " fullrunscript ", pid " (launch:einf-pid exit-info) " in work area " 
			 work-area ":\n====\n exit code " (launch:einf-exit-code exit-info) "\n" "====\n")
	    (if (not (launch:einf-exit-status exit-info))
		(exit 4))))

  ;; then files other than *testdat.db*
  (directory-fold 
   (lambda (f x)
     (let ((fullname (conc real-dir "/" f)))
       (if (not (string-search (regexp "testdat.db") f))
	   (runs:recursive-delete-with-error-msg fullname)))
     (+ 1 x))
   0 real-dir #t)
  ;; then the entire directory
  (runs:recursive-delete-with-error-msg real-dir))

;; cleanup often needs to remove all but the last N runs per target
;;
;; target-patts a1/b1/c1,a2/b2/c2 ...
;;

			   (if (eq? time-a time-b)
			       (< id-a id-b)
			       ;; (string<? (conc (vector-ref a 2))
			       ;;	    (conc (vector-ref b 2)))
			       #f))
		       (string<? (conc time-a)(conc time-b)))))))))


;; Save test state and status in to a file .final-status in the test directory
;;
(define (tests:save-final-status run-id test-id)
  (let* ((test-dat  (rmt:get-test-info-by-id run-id test-id))
	 (out-dir   (db:test-get-rundir test-dat))
	 (status-file  (conc out-dir "/.final-status"))
   )
    ;; first verify we are able to write the output file
    (if (not (file-write-access? out-dir))
	    (debug:print 0 *default-log-port* "ERROR: cannot write .final-status to " out-dir)
	    (let* 
         ((outp      (open-output-file status-file))
	       (status    (db:test-get-status   test-dat))
         (state     (db:test-get-state    test-dat)))
        (fprintf outp "~S\n" state) 
        (fprintf outp "~S\n" status) 
        (close-output-port outp)))))


;; summarize test in to a file test-summary.html in the test directory
;;
(define (tests:summarize-test run-id test-id)
  (let* ((test-dat  (rmt:get-test-info-by-id run-id test-id))
	 (out-dir   (db:test-get-rundir test-dat))
	 (out-file  (conc out-dir "/test-summary.html")))

	       (oup       (open-output-file out-file))
	       (status    (db:test-get-status   test-dat))
	       (color     (common:get-color-from-status status))
	       (logf      (db:test-get-final_logf test-dat))
	       (steps-dat (tests:get-compressed-steps run-id test-id)))
	  ;; (dcommon:get-compressed-steps #f 1 30045)
	  ;; (#("wasting_time" "23:36:13" "23:36:21" "0" "8.0s" "wasting_time.log"))
	
	  (s:output-new
	   oup
	   (s:html
	    (s:title "Summary for " full-name)
	    (s:body 
	     (s:h2 "Summary for " full-name)
	     (s:table 'cellspacing "0" 'border "1"