vnode_setup.in 13 KB
Newer Older
1
#!/usr/bin/perl -wT
Leigh B. Stoller's avatar
Leigh B. Stoller committed
2 3 4

#
# EMULAB-COPYRIGHT
5
# Copyright (c) 2000-2004 University of Utah and the Flux Group.
Leigh B. Stoller's avatar
Leigh B. Stoller committed
6 7
# All rights reserved.
#
8 9 10 11 12 13 14 15 16 17 18 19 20
use English;
use Getopt::Std;

#
# Set up the vnode state on a virtual (multiplexed) node.
#
# XXX - This script should only be run from os_setup!
#
# The output is all jumbled together since the updates are issued in parallel.
# Might be a pain when debugging. 
# 
sub usage()
{
Kirk Webb's avatar
Kirk Webb committed
21
    print STDOUT "Usage: vnode_setup [-f] [-k] [-n <numbatch>] [-w <wait_time>] <pid> <eid> [node ...]\n";
22 23
    exit(-1);
}
Kirk Webb's avatar
Kirk Webb committed
24
my  $optlist = "fdkn:w:";
25 26 27 28 29 30 31 32 33 34

#
# We don't want to run this script unless its the real version.
# That is, it must be setuid root. 
#
if ($EUID != 0) {
    die("*** $0:\n".
	"    Must be root! Maybe its a development version?\n");
}

35 36 37 38 39 40 41
#
# Configure variables
#
my $TB		= "@prefix@";
my $TESTMODE    = @TESTMODE@;
my $TBOPS       = "@TBOPSEMAIL@";
my $TBLOGS      = "@TBLOGSEMAIL@";
42
my $CLIENT_BIN  = "@CLIENT_BINDIR@";
43 44 45 46 47 48

my $ssh		= "$TB/bin/sshtb -n";
my $debug       = 0;
my $force       = 0;
my $failed      = 0;
my $killmode    = 0;
Kirk Webb's avatar
Kirk Webb committed
49
my $numbatch    = 10;
Kirk Webb's avatar
Kirk Webb committed
50
my $childwait   = 120;
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
my $dbuid;

#
# Load the Testbed support stuff. 
#
use lib "@prefix@/lib";
use libdb;
use libtestbed;

# un-taint path
$ENV{'PATH'} = '/bin:/usr/bin:/usr/local/bin';
delete @ENV{'IFS', 'CDPATH', 'ENV', 'BASH_ENV'};

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

#
# Parse command arguments. Once we return from getopts, all that should be
# left are the required arguments.
#
%options = ();
if (! getopts($optlist, \%options)) {
    usage();
}
75
if (@ARGV < 2) {
76 77 78 79 80 81 82 83 84 85 86
    usage();
}
if (defined($options{"f"})) {
    $force = 1;
}
if (defined($options{"d"})) {
    $debug = 1;
}
if (defined($options{"k"})) {
    $killmode = 1;
}
Kirk Webb's avatar
Kirk Webb committed
87 88 89 90 91 92 93 94
if (defined($options{"n"})) {
    if ($options{"n"} =~ /^(\d+)$/) {
        $numbatch = $1;
    }
    else {
        die ("*** Bad data in numbatch: $options{'n'}");
    }
}
Kirk Webb's avatar
Kirk Webb committed
95 96 97 98 99 100 101 102
if (defined($options{"w"})) {
    if ($options{"w"} =~ /^(\d+)$/) {
        $childwait = $1;
    }
    else {
        die ("*** Bad data in wait_time: $options{'w'}");
    }
}
Kirk Webb's avatar
Kirk Webb committed
103

104 105
my $pid   = shift(@ARGV);
my $eid   = shift(@ARGV);
106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123

#
# Untaint the arguments.
#
if ($pid =~ /^([-\@\w]+)$/) {
    $pid = $1;
}
else {
    die("*** Bad data in pid: $pid\n");
}	
if ($eid =~ /^([-\@\w]+)$/) {
    $eid = $1;
}
else {
    die("*** Bad data in eid: $eid\n");
}

#
124 125
# Verify permission to muck with this experiment. Note that this script
# is run as root from the plab monitor daemon.
126
#
127
if ($UID && !TBAdmin($UID) &&
128 129 130
    !TBExptAccessCheck($UID, $pid, $eid, TB_EXPT_DESTROY)) {
    die("*** $0:\n".
	"    You do not have permission to mess with $pid/$eid!\n");
131 132 133
}

#
134
# Get the list of nodes in this experiment.
135
#
136 137
my @nodes = ExpNodes($pid, $eid);
if (! @nodes) {
138 139 140
    warn("*** $0:\n".
	 "    No allocated nodes in experiment $pid/$eid!\n");
    exit(0);
141
}
142 143

# Nodes on the command line. Operate only on this set.
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168
if (@ARGV) {
    my %fulllist = ();

    # Temporary hash list for searching.
    foreach my $node ( @nodes ) {
	$fulllist{$node} = 1;
    }
    @nodes = ();

    foreach my $node ( @ARGV ) {
	if ($node =~ /^([-\@\w]+)$/) {
	    $node = $1;

	    if (!defined($fulllist{$node})) {
		die("*** $0:\n".
		     "   Node $node is not allocated to $pid/$eid!\n");
	    }
	}
	else {
	    die("Bad node name: $node.");
	}
	push(@nodes, $node);
    }
}
my $exptstate = ExpState($pid, $eid);
169 170 171 172

# Just the vnodes mam.
foreach my $node (@nodes) {
    my $pnode;
173
    my $jailed;
174
    my $plab;
175
    my $allocstate;
176
    my $mode = ($killmode ? "teardown" : "setup");
177

178
    if (! TBIsNodeVirtual($node, \$jailed, \$plab)) {
179 180
	next;
    }
181 182 183 184
    if (! TBPhysNodeID($node, \$pnode)) {
	die("*** $0:\n".
	    "    No physical node for $node!\n");
    }
185

186 187
    TBGetNodeAllocState($node, \$allocstate);

188 189 190
    #
    # On remote nodes, or when forcemode is on, always do the deed.
    # Otherwise, look at experiment state.
191 192
    #
    if (!$force) {
193 194 195 196
	if ($exptstate eq EXPTSTATE_SWAPPING) {
	    #
	    # When swapping, local vnodes go down with the physnode.
	    #
197 198 199 200 201
	    if (! TBIsNodeRemote($node)) {
		print "$node will $mode with local node $pnode.\n";
		next;
	    }
	    elsif ($allocstate eq TBDB_ALLOCSTATE_DOWN) {
202 203 204 205 206 207 208
		if (!$plab) {
		    print "$node failed to boot; skipping $mode.\n";
		    next;
		}
		# Plab nodes need to be cleaned up.
		print "$node failed to boot; changing to cleanup.\n";
		$mode = "cleanup";
209
	    }
210
	}
211 212
	elsif ($exptstate eq EXPTSTATE_ACTIVATING ||
	       $exptstate eq EXPTSTATE_MODIFY_RESWAP) {
213
	    #
214 215 216 217 218 219 220 221 222 223 224
	    # The allocstate determines if the vnode actually needs to
	    # be setup or torndown. Note that a failed experiment will
	    # cause a bunch of vnodes to be torndown, while in the
	    # ACTIVATING state. See os_setup and assign_wrapper; the
	    # idea is to avoid doing setup/teardown up vnodes on
	    # machines that are rebooting anyway, or that failed.
	    # Complicated by modify which could add/subtract a vnode on
	    # an existing machine, but not reboot the machine. Note that
	    # free (now unused) vnodes will land in RES_TEARDOWN. It is
	    # assumed that these booted okay, and need to be torndown,
	    # even though they are not RES_READY.
225
	    #
226 227 228
	    if (! TBIsNodeRemote($node)) {
		if ($killmode) {
		    if ($allocstate eq TBDB_ALLOCSTATE_DOWN) {
229 230 231 232
			print "$node failed to boot; skipping $mode.\n";
			next;
		    }
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_CLEAN()) {
233 234 235
			print "$node never booted; skipping $mode.\n";
			next;
		    }
236 237 238 239
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
			print "$node will $mode with local node $pnode.\n";
			next;
		    }
240 241
		}
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
242
		    print "$node is already setting up on local node $pnode\n";
243 244 245
		    next;
		}
	    }
246 247
	    else {
		if ($killmode) {
248 249 250
		    if ($allocstate eq TBDB_ALLOCSTATE_DEAD) {
			# plab only. See below.
			print "$node failed to initialize; skipping $mode.\n";
251 252
			next;
		    }
253 254 255 256 257 258 259 260 261
		    elsif ($allocstate eq TBDB_ALLOCSTATE_DOWN) {
			if (!$plab) {
			    print "$node failed to boot; skipping $mode.\n";
			    next;
			}
			# Plab nodes need to be cleaned up.
			print "$node failed to boot; changing to cleanup.\n";
			$mode = "cleanup";
		    }
262 263 264 265
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_CLEAN()) {
			print "$node never booted; skipping $mode.\n";
			next;
		    }
266
		}
267 268
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
		    print "$node is already set up on remote node $pnode\n";
269 270
		    next;
		}
271
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_DIRTY()) {
272
		    print "$node needs a reboot on remote node $pnode\n";
273 274
		    $mode = "reboot";
		}
275 276 277 278 279 280 281
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_RECONFIG()) {
		    print "$node needs a reconfig on remote node $pnode\n";
		    # We do not actually reconfig virtual nodes; just
		    # reboot them. Might reconfig someday, in which case
		    # this would move up into os_setup.
		    $mode = "reboot";
		}
282 283
	    }
	}
284 285 286 287 288 289
    }

    #
    # When setting up a vnode, force its event state into SHUTDOWN since
    # no telling what its initial state is. 
    # 
Kirk Webb's avatar
Kirk Webb committed
290 291
    # XXX: Don't we always want to set this?
    #
292
    if ($mode eq "teardown" || $mode eq "reboot") {
293 294
	TBSetNodeEventState($node, TBDB_NODESTATE_SHUTDOWN);
    }
295 296

    #
297 298 299 300 301 302
    # Put this into the list of calls we have to make in the next loop
    #
    push @vnodes, [$node, $pnode, $mode, $jailed, $plab];

}

303

304 305
my $children = 0;
my %child_vnodes = ();
Kirk Webb's avatar
Kirk Webb committed
306 307
print "vnode_setup running at parallelization: $numbatch ". 
    "wait_time: $childwait\n";
308 309
while (1) {

Kirk Webb's avatar
Kirk Webb committed
310
    # Space out the invocation of child processes a little.
Kirk Webb's avatar
Kirk Webb committed
311
    sleep(1);
Kirk Webb's avatar
Kirk Webb committed
312

313 314 315 316 317 318 319
    #
    # We're done when we've hit the last vnode, and we've outlived all of our
    # children
    #
    if ((!@vnodes) && ($children == 0)) {
	last;
    }
320

321 322 323
    #
    # There are more free slots
    #
Kirk Webb's avatar
Kirk Webb committed
324
    if (($children < $numbatch) && @vnodes) {
325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
	#
	# Look for a vnode that is not on a pnode we're already working on
	# 
	# XXX - do this!
	
	my ($vnode, $pnode, $mode, $jailed, $plab) = @{pop @vnodes};
    
	print STDOUT "Doing $mode of vnode $vnode on $pnode ...\n";

	#
	# Run an ssh command in a child process, protected by an alarm to
	# ensure that the ssh is not hung up forever if the machine is in some
	# funky state.
	# 
	my $syspid = fork();

	if ($syspid) {
	    #
	    # Just keep track of it, we'll wait for it finish down below
	    #
	    $child_vnodes{$syspid} =
		[$vnode, $pnode, $mode, $jailed, $plab, time()];
	    $children++;
	} else {
349 350
	    TBdbfork();	# So we get the event system fork too ...
	    
351
            my $exval = 0;
352 353 354
	    # Must change our real UID to root so that ssh will work.
	    $UID = 0;

355
	    if ($plab && $mode eq "setup") {
Kirk Webb's avatar
Kirk Webb committed
356 357 358 359
                # Make sure vnode is in the proper state before trying to
                # bring it up.
                # XXX: do this for all vnodes (see above)?
                TBSetNodeEventState($vnode, TBDB_NODESTATE_SHUTDOWN);
Kirk Webb's avatar
Kirk Webb committed
360
		if (TBForkCmd("$TB/sbin/plabnode ". 
361
                              ($force ? "-f" : "").
Kirk Webb's avatar
Kirk Webb committed
362
                              " alloc $pid $eid $vnode", 1)) {
363 364
		    print STDERR "*** $0:\n" .
                        "    Plab node allocation failed\n";
365 366 367 368
		    # Should check DB state instead.
		    exit(99);
		}
	    }
369

370 371 372 373 374 375 376
	    # Cleanup is used only on plab nodes.
	    if ($mode ne "cleanup") {
		my $args = (($mode eq "teardown") ? "-k " :
			    ($mode eq "reboot" ? "-r " : " "));
		$args   .= ($jailed   ? "-j " : " ");
		$args   .= ($plab     ? "-p " : " ");
		$args   .= "$vnode ";
377 378 379 380 381 382 383 384 385 386 387 388
                
		$exval = TBForkCmd("$ssh -host $vnode ".
                                 "$CLIENT_BIN/vnodesetup $args",1);
            }

            # Free the plab node lease if necessary.
            if ($plab && ($mode eq "teardown" || $mode eq "cleanup")) {
                exec("$TB/sbin/plabnode free $pid $eid $vnode");
                die("*** $0:\n".
                    "    exec failed!\n");
            }
	    exit($exval);
389 390 391 392 393 394 395 396 397 398 399
	}
    } else {
	#
	# We have too many of the little rugrats, wait for one to die
	#

	#
	# Set up a timer - we want to kill processes after they hit 120 seconds
	# old (not much of a life, is it?), so we find the first one marked for
	# death.
	#
400
	my $oldest;
401
	my $oldestpid = 0;
Kirk Webb's avatar
Kirk Webb committed
402
        my $oldestvnode = "";
403 404
	while (my ($pid, $aref) = each %child_vnodes) {
	    my ($vnode, $pnode, $mode, $jailed, $plab, $birthtime) = @$aref;
405
	    if ((!$oldestpid) || ($birthtime < $oldest)) {
406 407
		$oldest = $birthtime;
		$oldestpid = $pid;
Kirk Webb's avatar
Kirk Webb committed
408
                $oldestvnode = $vnode;
409 410 411 412 413 414 415 416 417 418 419
	    }
	}

	#
	# Sanity check
	#
	if (!$oldest) {
	    die "*** $0\n".
	    	"Uh oh, I have no children left, something is wrong!\n";
	}

420
	#
Kirk Webb's avatar
Kirk Webb committed
421
	# If the oldest has already expired, just kill it off right now, and go
422 423 424
	# back around the loop
	#
	my $now = time();
Kirk Webb's avatar
Kirk Webb committed
425
	my $waittime = ($oldest + $childwait) - time();
426

427
	#
Kirk Webb's avatar
Kirk Webb committed
428
	# Kill of the oldest if it gets too old while we're waiting.
429
	#
Kirk Webb's avatar
Kirk Webb committed
430 431 432 433
	local $SIG{ALRM} = 
            sub { print "*** $0: timeout (alarm clock) waiting for ".
                        "vnode: $oldestvnode\n";
                  kill("TERM", $oldestpid); };
434
	if ($waittime <= 0) {
Kirk Webb's avatar
Kirk Webb committed
435
            print "*** $0: timeout waiting for vnode: $oldestvnode\n";
436 437 438 439
	    kill("TERM",$oldestpid);
	} else {
	    alarm($waittime);
	}
440 441

	my $childpid = wait();
442
	my $exitstatus = $?;
443

444 445
	alarm 0;

446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
	#
	# Another sanity check
	#
	if ($childpid < 0) {
	    die "*** $0\n".
	    	"wait() returned <0, something is wrong!\n";
	}

	#
	# Look up to see what vnode, etc. this was associated with - if we
	# don't know about this child, ignore it
	#
	my $aref = $child_vnodes{$childpid};
	next unless @$aref;
	my ($vnode, $pnode, $mode, $jailed, $plab, $birthtime) = @$aref;
	$children--;
	delete $child_vnodes{$childpid};

464
	#
465
	# Look for setup failure, reported back through ssh.
466
	# 
467
	if ($exitstatus) {
468 469
	    print STDERR "vnode $vnode $mode on $pnode returned $?.\n"
		if $debug;
470

471
	    if ($exitstatus == 256) {
472
		print STDERR "$vnode is not running sshd.\n" if $debug;
473 474
	    }
	    elsif ($exitstatus == 15) {
475
		print STDERR "$vnode is wedged.\n" if $debug;
476
	    }
477 478 479
	    elsif ($exitstatus >> 8 == 99) {
		print STDERR "$vnode did not allocate properly.\n" if $debug;
	    }
480

481 482
	    push @failed_nodes, [$vnode, $pnode, $mode, $exitstatus];
	    
483 484
	    warn("*** $0:\n".
		 "    Virtual node $vnode $mode failure!\n");
485
	}
486

487 488 489 490 491 492 493 494 495 496
	if ($plab) {
	    #
	    # If the node was in the setup process, then mark its allocstate
	    # as down so os_setup knows not to bother waiting for it. DEAD
	    # is a temp state, different then DOWN. It indicates the node
	    # was never instantiated (important distinction for plab). 
	    #
	    if ($exitstatus &&
		(($mode eq "setup") || ($mode eq "reboot"))) {
		TBSetNodeAllocState($vnode, TBDB_ALLOCSTATE_DEAD());
497 498
	    }
	}
499 500 501
    }
}

502 503 504 505 506 507
#
# In force node, do not bother with this other stuff. 
#
exit(scalar(@failed_nodes))
    if ($force);

508 509 510 511 512 513
#
# Send mail to testbed-ops about failed nodes
#
if (@failed_nodes) { 
    my $failed_lines = join("\n",map { join("\t",@{$_}) } @failed_nodes);
    SENDMAIL($TBOPS, "Virtual Node failure for $pid/$eid",
514 515 516
	     "The following virtual nodes failed: \n" .
	     "vnode\t\tpnode\tmode\texit status\n" .
	     $failed_lines);
517 518
}

519 520 521 522 523 524
if ($killmode) {
    print STDOUT "Vnode teardown finished.\n";
}
else {
    print STDOUT "Vnode setup initiated on all nodes ...\n";
}
525
exit(0);