batch_daemon.in 14.6 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
#!/usr/bin/perl -wT
use English;
use Getopt::Std;

#
# Create a batch experiment.
#
# BIG ASS WARNING: This works great as long as paper does not reboot!
# Needs some work if we want it to be stateless across reboots. 
# 
# usage: batch_daemon
#
sub usage()
{
15
16
    print STDOUT "Usage: batch_daemon [-d]\n" .
	"Use the -d option to prevent daemonization\n";
17
18
    exit(-1);
}
19
my  $optlist = "d";
20
21
22
23
24
25
26
27
28
29
30
31

#
# Configure variables
#
my $TB       = "@prefix@";
my $DBNAME   = "@TBDBNAME@";
my $TBOPS    = "@TBOPSEMAIL@";

my $tbbindir = "$TB/bin/";
my $batchdir = "$TB/batch";
my $startexp = "$TB/bin/startexp";
my $endexp   = "$TB/bin/endexp";
32
my $avail    = "$TB/sbin/avail";
33
34
my $batchlog = "$TB/log/batchlog";
my $projroot = "/proj";
35
my $debug    = 0;
36
37
38
39
40
41
42
43
44
my $dirname;

#
# These are valid in the children, not the parent. I suppose I could use
# dynamically scoped variables, but hardly worth it.
#
my $eid;
my $pid;
my $logname;
Leigh B. Stoller's avatar
Leigh B. Stoller committed
45
my $nsfile;
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
my $user_name  = "Batch Daemon";
my $user_email = "$TBOPS";

#
# Turn off line buffering on output
#
$| = 1;

#
# Untaint the path
# 
$ENV{'PATH'} = "/bin:/usr/bin:";
delete @ENV{'IFS', 'CDPATH', 'ENV', 'BASH_ENV'};

#
# Parse command arguments. Once we return from getopts, all that should be
# left are the required arguments.
#
%options = ();
if (! getopts($optlist, \%options)) {
    usage();
}
if (@ARGV != 0) {
    usage();
}
71
72
73
if (defined($options{"d"})) {
    $debug = $options{"d"};
}
74
75

# Go to ground.
76
77
78
if (! $debug) {
    daemonize();
}
79
80
81
82
83
84
85
86
87
88
89

#
# Set up for querying the database.
# 
use Mysql;
my $DB = Mysql->connect("localhost", $DBNAME, "script", "none");

#
# Loop, looking for batch experiments that want to run.
# 
while (1) {
90
91
92
    my($count);
    my(%row, %pending_row);
    
93
94
95
96
97
98
99
100
101
102
    #
    # Need to lock the table here because of cancelation in killbatchexp.
    # See the comments in there. We need to atomically grab the next
    # batch experiment we want to try, and then change its state from
    # new to configuring. We want to grab just one experiment, since
    # it takes a while to configure an experiment, and grabbing a bunch and
    # locking them up might result in having to wait a really long time
    # to cancel a batch experiment that hasn't really tried to start yet!
    # Thats would ne annoying to users, and we love our users, right?
    #
103
    # So, now you're wondering what my selection criteria is? Well, its
104
105
106
107
108
    # damn simplistic. I set the "started" datetime field each attempt,
    # and I pick the batch_experiment with the oldest time, thereby cycling
    # through in a "least recently attempted" manner. 
    #
    $query_result =
109
110
111
112
113
114
115
116
	DBquery("lock tables batch_experiments write");
    if (! $query_result) {
	print "DB Error locking tables. Waiting a bit ...\n";
	sleep(10);
	next;
    }
    
    $pending_result =
117
118
119
	DBquery("SELECT * FROM batch_experiments ".
		"WHERE status='new' and canceled=0 ORDER BY started LIMIT 1");

120
121
122
123
124
125
126
127
128
129
130
131
132
    $running_result =
	DBquery("SELECT * FROM batch_experiments ".
		"WHERE status='running' ORDER BY started");

    if (!$pending_result || !$running_result) {
	print "DB Error getting batch info. Waiting a bit ...\n";
	DBquery("unlock tables");
	sleep(10);
	next;

    }

    if (!$pending_result->numrows && !$running_result->numrows) {
133
134
135
136
137
138
	DBquery("unlock tables");
	sleep(10);
	next;
    }

    #
139
140
141
    # If we have a pending experiment to run, set its state to configuring
    # right away, while we have the tables locked. This prevents killbatchexp
    # from seeing it as something it can cancel.
142
    #
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
    if ($pending_result->numrows) {
	%pending_row = $pending_result->fetchhash();

	# Local vars!
	my $eid = $pending_row{'eid'};
	my $pid = $pending_row{'pid'};
	my $now = `date '+20%y-%m-%d %H:%M:%S'`;

	$query_result = 
	    DBquery("update batch_experiments set status='configuring', ".
		    "started='$now' where eid='$eid' and pid='$pid'");

	if (! $query_result) {
	    print "DB error setting batch $pid/$eid to configuring.\n";
	    DBquery("unlock tables");
	    sleep(10);
	    next;
	}
    }
162
163
    DBquery("unlock tables");

164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
    #
    # Okay, first we check the status of running batch mode experiments
    # since we want to end those before trying to start any new ones, cause
    # it would be nice to have as many nodes available as possible before
    # trying to add a new one. This can potentially delay startup, but thats
    # okay. Its a batch system.
    #
    # If you are wondering why I check for finished experiments in the main
    # loop instead of in the child that started the experiment, its so that
    # we fire up again and look for them in the event that paper goes down.
    #
    $count = $running_result->numrows;
    for ($i = 0; $i < $count; $i++) {
	%row = $running_result->fetchhash();

	my $canceled = $row{'canceled'};
	if ($canceled) {
	    dosomething("cancel", %row);
	    next;
	}
	if (isexpdone(%row)) {
	    dosomething("end", %row);
	    next;
	}
    }

    #
    # Finally start an actual experiment!
    #
    if ($pending_result->numrows) {
	dosomething("start", %pending_row);
    }
    sleep(60);
197
198
199
}

#
200
# Do something as the user. Either, start, end, or cancel an experiment.
201
#
202
sub dosomething($$)
203
{
204
205
    my($dowhat)   = shift;
    my(%exphash)  = @_;
206
207
208
209
210
211
    my($uid, $gid, $row);

    # Global vars
    $eid = $exphash{'eid'};
    $pid = $exphash{'pid'};

212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
    print "Doing a '$dowhat' to batch experiment $pid/$eid\n";

    #
    # Create a temporary name for a log file. We do this in the parent so
    # we can remove it when the child ends. The child could remove it, but
    # since it is open in the child, it has the tendency to stick around.
    #
    $logname = `mktemp /tmp/$dowhat-batch-$pid-$eid.XXXXXX`;

    # Note different taint check (allow /).
    if ($logname =~ /^([-\@\w.\/]+)$/) {
	$logname = $1;
    } else {
	die "Bad data in $logname";
    }

228
229
230
    #
    # Start up a child to run the guts. The parent waits. If the
    # experiment configures okay, the parent can return to try something
231
    # else.
232
233
234
    #
    $childpid = fork();
    if ($childpid) {
235
236
	print "Child PID $childpid started to $dowhat $pid/$eid\n";

237
	waitpid($childpid, 0);
Leigh B. Stoller's avatar
Leigh B. Stoller committed
238
	my $status = $? >> 8;
239

Leigh B. Stoller's avatar
Leigh B. Stoller committed
240
	print "Child PID $childpid exited with exit status $status\n";
241
242

	unlink($logname);
Leigh B. Stoller's avatar
Leigh B. Stoller committed
243
	return $status;
244
    }
245
246
247
248
249
250
251
252
    openlog($logname);

    #
    # Form a new connection to the DB since we are in the child. Not sure
    # if this happens as a result of the fork, but lets be sure.
    #
    undef($DB);
    $DB = Mysql->connect("localhost", $DBNAME, "script", "none");
253

254
255
256
257
258
    my $creator  = $exphash{'creator_uid'};
    my $longname = $exphash{'name'};
    
    # Global var
    $dirname  = "$batchdir/$pid-$eid";
Leigh B. Stoller's avatar
Leigh B. Stoller committed
259
    $nsfile  = "$dirname/$eid.ns";
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284

    #
    # Get some user information. 
    #
    $query_result =
	$DB->query("SELECT usr_name,usr_email from users ".
		   "WHERE uid='$creator'");

    if (! $query_result ||
	$query_result->numrows != 1) {
	fatal("DB Error getting user information for uid $creator\n");
    }
    @row = $query_result->fetchrow_array();
    $user_name  = $row[0];
    $user_email = $row[1];

    #
    # Figure out the unix uid/gid that the experiment configuration is
    # going to run as. 
    #
    (undef,undef,$uid) = getpwnam($creator) or
	fatal("No such user $creator");
    (undef,undef,$gid) = getgrnam($pid) or
	fatal("No such group $pid");

285
286
287
288
289
290
    #
    # Change the ownership of the log file before we flip.
    #
    chown($uid, $gid, $logname);

    # Flip to the user. We never flip back.
291
292
    $EGID = $GID = $gid;
    $EUID = $UID = $uid;
293
    $ENV{'USER'} = $creator;
294
    
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
    if ($dowhat eq "start") {
	startexp(%exphash);
    }
    elsif ($dowhat eq "end") {
	endexp(%exphash);
    }
    elsif ($dowhat eq "cancel") {
	cancelexp(1, %exphash);
    }
    exit(0);
}

#
# Try to start an experiment. Never returns.
# 
sub startexp($)
{
    my(%exphash)  = @_;
    my($exit_status, $running);

    my $creator   = $exphash{'creator_uid'};
    my $longname  = $exphash{'name'};
    my $numpcs    = $exphash{'numpcs'};
    my $numsharks = $exphash{'numsharks'};
    my $attempts  = $exphash{'attempts'};

321
    #
322
323
    # Lets see if there are enough nodes. This is a really hacky test,
    # especially for the sharks since they are allocated by shelf. 
324
    #
325
326
327
328
    system("$avail type=pc > /dev/null");
    my $availpcs = $? >> 8;
    system("$avail type=shark > /dev/null");
    my $availsharks = $? >> 8;
329

330
331
332
333
334
335
336
337
338
    if ($availpcs < $numpcs || $availsharks < $numsharks) {
	#
	# XXX - What if this update fails?
	# 
	$query_result = 
	    DBquery("update batch_experiments set status='new' ".
		    "where eid='$eid' and pid='$pid'");
	
	exit(69);
339
340
341
342
343
344
    }

    #
    # Insert an experiment record for startexp.
    #
    my $rightnow = `date '+20%y-%m-%d %H:%M:%S'`;
345
346
347
348
349
350
351
352
353
    $query_result =
	DBquery("insert into experiments ".
		"(eid, pid, expt_created, expt_name, ".
		"expt_head_uid, expt_start, expt_ready, batchmode) ".
		"VALUES ('$eid', '$pid', '$rightnow', '$longname', ".
		"'$creator', '$rightnow', 0, 1)");
    if (! $query_result) {
	fatal("DB error inserting experiment record. Quitting ...\n");
    }
354
355
356
357
358

    #
    # Try to start the experiment. If it fails, the experiment is gone.
    #
    system("$startexp -b $pid $eid $nsfile");
359
360
    $exit_status = $? >> 8;
    $running     = 1;
361
362
363
364
365
    if ($exit_status) {
	$running = 0;
    }
    
    #
366
367
    # Look for cancelation. If we get a DB error on this, just continue cause
    # we can pick up the cancelation later.
368
369
370
371
372
    #
    $query_result =
	DBquery("select canceled from batch_experiments ".
		"where eid='$eid' and pid='$pid'");

373
374
    if ($query_result) {
	@row = $query_result->fetchrow_array();
375

376
377
378
379
380
381
382
383
384
	if ($row[0]) {
	    cancelexp($running);
	    #
	    # Never returns, but just to be safe ...
	    #
	    exit(0);
	}
    }
			    
385
386
387
388
389
390
    #
    # If the configuration failed, then send email for now. This
    # part needs work. We have to reset the state to "new" so that
    # it will be retried again later. 
    #
    if (! $running) {
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
	#
	# XXX - What if this update fails?
	# 
	$query_result = 
	    DBquery("update batch_experiments set status='new', ".
		    "attempts=attempts+1 where eid='$eid' and pid='$pid'");

	if (($attempts % 5) == 0) {
	    $attempts++;
	    
	    fatal("Could not configure Batch Mode experiment $pid/$eid\n".
		  "There have been $attempts attempts made to start this ".
		  "batch\n");
	}
	exit(45);
406
407
408
409
410
    }

    #
    # Well, it configured! Lets set it state to running.
    #
411
    # XXX - What if this update fails?
Leigh B. Stoller's avatar
Leigh B. Stoller committed
412
    #
413
414
415
    $query_result = 
	DBquery("update batch_experiments set status='running' ".
		"where eid='$eid' and pid='$pid'");
Leigh B. Stoller's avatar
Leigh B. Stoller committed
416

417
418
    email_status("Batch Mode experiment $pid/$eid is now running!\n".
		 "Please consult the Web interface to see how it is doing\n");
Leigh B. Stoller's avatar
Leigh B. Stoller committed
419

420
    #
421
    # Done with this phase. Must exit.
422
    #
423
424
    exit(0);
}
425

426
427
428
429
430
431
#
# End an experiment. Never returns.
#
sub endexp($)
{
    my(%exphash)  = @_;
432
    
433
434
435
436
437
    system("$endexp -b $pid $eid");
    DBquery("DELETE from batch_experiments WHERE eid='$eid' and pid='$pid'");
    email_status("Batch Mode experiment $pid/$eid has finished!\n");
    system("rm -rf $dirname");
   
438
    #
439
    # Child must exit!
440
    #
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
    exit(0);
}

#
# Cancel an experiment. Never returns.
#
sub cancelexp($$)
{
    my($running) = shift;
    my(%exphash) = @_;
    
    if ($running) {
	system("$endexp -b $pid $eid");
    }

456
    DBquery("DELETE from batch_experiments WHERE eid='$eid' and pid='$pid'");
457
458
    notify_user("Your Batch Mode experiment has been canceled. You may now\n".
		"reuse the experiment name\n", "Canceled", 0);
459
460
461
462
463
464
465
466
    system("rm -rf $dirname");
   
    #
    # Child must exit!
    #
    exit(0);
}

467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
#
# Check experiment status. Looks to see if all of the nodes in an
# experiment have reported in.
#
sub isexpdone($)
{
    my(%exphash)  = @_;
    my($row, $done);
    
    # Global vars
    $eid = $exphash{'eid'};
    $pid = $exphash{'pid'};

    print "Checking to see if $pid/$eid has finished up yet\n";

    #
    # Look to see if any nodes yet to report status. If so, spin again.
    #
    $query_result =
	DBquery("SELECT startstatus FROM nodes LEFT JOIN reserved ".
		"ON nodes.node_id=reserved.node_id ".
		"WHERE reserved.eid='$eid' and reserved.pid='$pid'");

    if (! $query_result) {
	return 0;
    }

    $done = 1;
    for ($i = 0; $i < $query_result->numrows; $i++) {
	@row = $query_result->fetchrow_array();
	
	if ($row[0] eq "none") {
	    $done = 0;
	}
    }
    return $done;
}

505
506
507
sub DBquery($)
{
    my($query) = $_[0];
508
    my($result);
509

510
    $result = $DB->query($query);
511

512
513
    if (! $result) {
	print "DB Query failed: $query\n";
514
515
    }

516
    return $result;
517
518
519
520
}

#
# Start up a child, and set its descriptors talking to a log file.
Leigh B. Stoller's avatar
Leigh B. Stoller committed
521
# The log file already exists, created with mktemp above.
522
523
524
525
526
527
528
# 
sub openlog($)
{
    my($logname) = $_[0];
	
    #
    # We have to disconnect from the caller by redirecting both STDIN and
Leigh B. Stoller's avatar
Leigh B. Stoller committed
529
530
    # STDOUT away from the pipe. Otherwise the caller will continue to wait
    # even though the parent has exited. 
531
532
    #
    open(STDIN, "< /dev/null") or
533
	fatal("opening /dev/null for STDIN: $!");
534
535
536
537
538
539
540
541
542

    open(STDERR, ">> $logname") or
	fatal("opening $logname for STDERR: $!");
    open(STDOUT, ">> $logname") or
	fatal("opening $logname for STDOUT: $!");

    return 0;
}

543
sub fatal($)
544
545
546
{
    my($mesg) = $_[0];

547
    notify_user($mesg, "Failure", 1);
548
549
550
551

    exit(-1);
}

552
sub email_status($)
553
{
554
    my($mesg) = $_[0];
555

556
    notify_user($mesg, "Status", 0);
557
558
}

559
sub notify_user($$$)
560
{
561
562
    my($mesg, $subline, $iserr) = @_;
    my($to,$cc);
563
564
565

    print STDOUT "$mesg\n";

566
567
568
569
570
571
572
573
574
575
576
577
578
    #
    # If an error, goes to TBOPS and is CC'ed to user.
    # Otherwise goes to user and is CC'ed to TBOPS
    #
    if ($iserr) {
	$to = $TBOPS;
	$cc = "$user_name <$user_email>";
    }
    else {
	$to = "$user_name <$user_email>";
	$cc = $TBOPS;
    }

579
    open(MAIL, "| /usr/bin/mail ".
580
581
	 "-s \"TESTBED: Batch Mode Experiment $subline $pid/$eid\" ".
	 "-c $cc \"$to\" >/dev/null 2>&1")
582
583
584
585
586
587
588
589
590
591
592
593
	or die "Cannot start mail program: $!";

    print MAIL $mesg;

    if (defined($logname) && open(IN, "$logname")) {
	print MAIL "\n\n---------\n\n";
	
	while (<IN>) {
	    print MAIL "$_";
	}
	close(IN);
    }
Leigh B. Stoller's avatar
Leigh B. Stoller committed
594
595
596
597
598
599
600
601
602
603

    if (defined($nsfile) && open(IN, "$nsfile")) {
	print MAIL "\n\n---------\n\n";
	
	while (<IN>) {
	    print MAIL "$_";
	}
	close(IN);
    }
    
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
    close(MAIL);
}

#
# Become a daemon.
# 
sub daemonize()
{
    my $mypid = fork();
    if ($mypid) {
	exit(0);
    }

    #
    # We have to disconnect from the caller by redirecting both STDIN and
    # STDOUT away from the pipe. Otherwise the caller will continue to wait
    # even though the parent has exited. 
    #
    open(STDIN, "< /dev/null") or
	die("opening /dev/null for STDIN: $!");

    #
    # Open the batch log and start writing to it. 
    #
    open(STDERR, ">> $batchlog") or die("opening $batchlog for STDERR: $!");
    open(STDOUT, ">> $batchlog") or die("opening $batchlog for STDOUT: $!");

    return 0;
}