vnode_setup.in 11.6 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) 2000-2003 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()
{
21
    print STDOUT "Usage: vnode_setup [-f] [-k] <pid> <eid> [node ...]\n";
22 23 24
    exit(-1);
}
my  $optlist = "fdk";
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 49 50

my $ssh		= "$TB/bin/sshtb -n";
my $debug       = 0;
my $force       = 0;
my $failed      = 0;
my $killmode    = 0;
my $dbuid;

51
my $MAX_CHILDREN = 8;
52
my $CHILD_TIMEOUT = 960;  # 15 + 1 mins slop - plab needs a LOOOONNNGG TIME....
53

54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
#
# 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($optlist, \%options)) {
    usage();
}
76
if (@ARGV < 2) {
77 78 79 80 81 82 83 84 85 86 87
    usage();
}
if (defined($options{"f"})) {
    $force = 1;
}
if (defined($options{"d"})) {
    $debug = 1;
}
if (defined($options{"k"})) {
    $killmode = 1;
}
88 89
my $pid   = shift(@ARGV);
my $eid   = shift(@ARGV);
90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107

#
# 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");
}

#
108 109
# Verify permission to muck with this experiment. Note that this script
# is run as root from the plab monitor daemon.
110
#
111
if ($UID && !TBAdmin($UID) &&
112 113 114
    !TBExptAccessCheck($UID, $pid, $eid, TB_EXPT_DESTROY)) {
    die("*** $0:\n".
	"    You do not have permission to mess with $pid/$eid!\n");
115 116 117
}

#
118
# Get the list of nodes in this experiment.
119
#
120 121
my @nodes = ExpNodes($pid, $eid);
if (! @nodes) {
122 123 124
    warn("*** $0:\n".
	 "    No allocated nodes in experiment $pid/$eid!\n");
    exit(0);
125
}
126 127

# Nodes on the command line. Operate only on this set.
128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152
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);
153 154 155 156

# Just the vnodes mam.
foreach my $node (@nodes) {
    my $pnode;
157
    my $jailed;
158
    my $plab;
159
    my $allocstate;
160
    my $mode = ($killmode ? "teardown" : "setup");
161

162
    if (! TBIsNodeVirtual($node, \$jailed, \$plab)) {
163 164
	next;
    }
165 166 167 168
    if (! TBPhysNodeID($node, \$pnode)) {
	die("*** $0:\n".
	    "    No physical node for $node!\n");
    }
169

170 171
    TBGetNodeAllocState($node, \$allocstate);

172 173 174
    #
    # On remote nodes, or when forcemode is on, always do the deed.
    # Otherwise, look at experiment state.
175 176
    #
    if (!$force) {
177 178 179 180
	if ($exptstate eq EXPTSTATE_SWAPPING) {
	    #
	    # When swapping, local vnodes go down with the physnode.
	    #
181 182 183 184 185
	    if (! TBIsNodeRemote($node)) {
		print "$node will $mode with local node $pnode.\n";
		next;
	    }
	    elsif ($allocstate eq TBDB_ALLOCSTATE_DOWN) {
186 187 188 189 190 191 192
		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";
		$mode = "cleanup";
193
	    }
194
	}
195 196
	elsif ($exptstate eq EXPTSTATE_ACTIVATING ||
	       $exptstate eq EXPTSTATE_MODIFY_RESWAP) {
197
	    #
198 199 200 201 202 203 204 205 206 207 208
	    # The allocstate determines if the vnode actually needs to
	    # be setup or torndown. Note that a failed experiment will
	    # cause a bunch of vnodes to be torndown, while in the
	    # ACTIVATING state. See os_setup and assign_wrapper; the
	    # idea is to avoid doing setup/teardown up vnodes on
	    # machines that are rebooting anyway, or that failed.
	    # Complicated by modify which could add/subtract a vnode on
	    # an existing machine, but not reboot the machine. Note that
	    # free (now unused) vnodes will land in RES_TEARDOWN. It is
	    # assumed that these booted okay, and need to be torndown,
	    # even though they are not RES_READY.
209
	    #
210 211 212
	    if (! TBIsNodeRemote($node)) {
		if ($killmode) {
		    if ($allocstate eq TBDB_ALLOCSTATE_DOWN) {
213 214 215 216
			print "$node failed to boot; skipping $mode.\n";
			next;
		    }
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_CLEAN()) {
217 218 219
			print "$node never booted; skipping $mode.\n";
			next;
		    }
220 221 222 223
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
			print "$node will $mode with local node $pnode.\n";
			next;
		    }
224 225
		}
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
226
		    print "$node is already setting up on local node $pnode\n";
227 228 229
		    next;
		}
	    }
230 231
	    else {
		if ($killmode) {
232 233 234
		    if ($allocstate eq TBDB_ALLOCSTATE_DEAD) {
			# plab only. See below.
			print "$node failed to initialize; skipping $mode.\n";
235 236
			next;
		    }
237 238 239 240 241 242 243 244 245
		    elsif ($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";
			$mode = "cleanup";
		    }
246 247 248 249
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_CLEAN()) {
			print "$node never booted; skipping $mode.\n";
			next;
		    }
250
		}
251 252
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
		    print "$node is already set up on remote node $pnode\n";
253 254
		    next;
		}
255
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_DIRTY()) {
256
		    print "$node needs a reboot on remote node $pnode\n";
257 258
		    $mode = "reboot";
		}
259 260
	    }
	}
261 262 263 264 265 266
    }

    #
    # When setting up a vnode, force its event state into SHUTDOWN since
    # no telling what its initial state is. 
    # 
267
    if ($mode eq "teardown" || $mode eq "reboot") {
268 269
	TBSetNodeEventState($node, TBDB_NODESTATE_SHUTDOWN);
    }
270 271

    #
272 273 274 275 276 277
    # Put this into the list of calls we have to make in the next loop
    #
    push @vnodes, [$node, $pnode, $mode, $jailed, $plab];

}

278

279 280 281 282 283 284 285 286 287 288 289
my $children = 0;
my %child_vnodes = ();
while (1) {

    #
    # We're done when we've hit the last vnode, and we've outlived all of our
    # children
    #
    if ((!@vnodes) && ($children == 0)) {
	last;
    }
290

291 292 293
    #
    # There are more free slots
    #
294
    if (($children <= $MAX_CHILDREN) && @vnodes) {
295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
	#
	# 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 {
319
            my $exval = 0;
320 321 322
	    # Must change our real UID to root so that ssh will work.
	    $UID = 0;

323
	    if ($plab && $mode eq "setup") {
324 325 326
		if (TBForkCmd("$TB/sbin/plabnode alloc ". 
                              ($force ? "-f" : "").
                              " $pid $eid $vnode", 1)) {
327 328
		    print STDERR "*** $0:\n" .
                        "    Plab node allocation failed\n";
329 330 331 332
		    # Should check DB state instead.
		    exit(99);
		}
	    }
333

334 335 336 337 338 339 340
	    # 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 ";
341 342 343 344 345 346 347 348 349 350 351 352
                
		$exval = TBForkCmd("$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);
353 354 355 356 357 358 359 360 361 362 363
	}
    } 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.
	#
364
	my $oldest;
365 366 367
	my $oldestpid = 0;
	while (my ($pid, $aref) = each %child_vnodes) {
	    my ($vnode, $pnode, $mode, $jailed, $plab, $birthtime) = @$aref;
368
	    if ((!$oldestpid) || ($birthtime < $oldest)) {
369 370 371 372 373 374 375 376 377 378 379 380 381
		$oldest = $birthtime;
		$oldestpid = $pid;
	    }
	}

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

382 383 384 385 386
	#
	# If the oldest has already expried, just kill it off right now, and go
	# back around the loop
	#
	my $now = time();
387
	my $waittime = ($oldest + $CHILD_TIMEOUT) - time();
388

389 390 391
	#
	# Kill of the oldest if he gets too old while we're waiting
	#
392
	local $SIG{ALRM} = sub { kill("TERM", $oldestpid); };
393 394 395 396 397
	if ($waittime <= 0) {
	    kill("TERM",$oldestpid);
	} else {
	    alarm($waittime);
	}
398 399

	my $childpid = wait();
400
	my $exitstatus = $?;
401

402 403
	alarm 0;

404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421
	#
	# 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};

422
	#
423
	# Look for setup failure, reported back through ssh.
424
	# 
425
	if ($exitstatus) {
426 427
	    print STDERR "vnode $vnode $mode on $pnode returned $?.\n"
		if $debug;
428

429
	    if ($exitstatus == 256) {
430
		print STDERR "$vnode is not running sshd.\n" if $debug;
431 432
	    }
	    elsif ($exitstatus == 15) {
433
		print STDERR "$vnode is wedged.\n" if $debug;
434
	    }
435 436 437
	    elsif ($exitstatus >> 8 == 99) {
		print STDERR "$vnode did not allocate properly.\n" if $debug;
	    }
438

439 440
	    push @failed_nodes, [$vnode, $pnode, $mode, $exitstatus];
	    
441 442
	    warn("*** $0:\n".
		 "    Virtual node $vnode $mode failure!\n");
443
	}
444

445 446 447 448 449 450 451 452 453 454
	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());
455 456
	    }
	}
457 458 459
    }
}

460 461 462 463 464 465
#
# In force node, do not bother with this other stuff. 
#
exit(scalar(@failed_nodes))
    if ($force);

466 467 468 469 470 471
#
# 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",
472 473 474
	     "The following virtual nodes failed: \n" .
	     "vnode\t\tpnode\tmode\texit status\n" .
	     $failed_lines);
475 476
}

477 478 479 480 481 482
if ($killmode) {
    print STDOUT "Vnode teardown finished.\n";
}
else {
    print STDOUT "Vnode setup initiated on all nodes ...\n";
}
483
exit(0);