vnode_setup.in 12.4 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 52
my $MAX_CHILDREN = 8;

53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74
#
# 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;
}
87 88
my $pid   = shift(@ARGV);
my $eid   = shift(@ARGV);
89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108

#
# 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");
}

#
# Verify permission to muck with this experiment.
#
109 110 111 112
if (!TBAdmin($UID) &&
    !TBExptAccessCheck($UID, $pid, $eid, TB_EXPT_DESTROY)) {
    die("*** $0:\n".
	"    You do not have permission to mess with $pid/$eid!\n");
113 114 115
}

#
116
# Get the list of nodes in this experiment.
117
#
118 119
my @nodes = ExpNodes($pid, $eid);
if (! @nodes) {
120 121 122
    warn("*** $0:\n".
	 "    No allocated nodes in experiment $pid/$eid!\n");
    exit(0);
123
}
124 125

# Nodes on the command line. Operate only on this set.
126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
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);
151 152 153 154

# Just the vnodes mam.
foreach my $node (@nodes) {
    my $pnode;
155
    my $jailed;
156
    my $plab;
157
    my $allocstate;
158
    my $mode = ($killmode ? "teardown" : "setup");
159

160
    if (! TBIsNodeVirtual($node, \$jailed, \$plab)) {
161 162
	next;
    }
163 164 165 166
    if (! TBPhysNodeID($node, \$pnode)) {
	die("*** $0:\n".
	    "    No physical node for $node!\n");
    }
167

168 169
    TBGetNodeAllocState($node, \$allocstate);

170 171 172
    #
    # On remote nodes, or when forcemode is on, always do the deed.
    # Otherwise, look at experiment state.
173 174
    #
    if (!$force) {
175 176 177 178
	if ($exptstate eq EXPTSTATE_SWAPPING) {
	    #
	    # When swapping, local vnodes go down with the physnode.
	    #
179 180 181 182 183
	    if (! TBIsNodeRemote($node)) {
		print "$node will $mode with local node $pnode.\n";
		next;
	    }
	    elsif ($allocstate eq TBDB_ALLOCSTATE_DOWN) {
184 185 186 187 188 189 190
		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";
		$mode = "cleanup";
191
	    }
192
	}
193 194
	elsif ($exptstate eq EXPTSTATE_ACTIVATING) {
	    #
195 196 197 198 199 200 201 202 203 204 205
	    # 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.
206
	    #
207 208 209
	    if (! TBIsNodeRemote($node)) {
		if ($killmode) {
		    if ($allocstate eq TBDB_ALLOCSTATE_DOWN) {
210 211 212 213
			print "$node failed to boot; skipping $mode.\n";
			next;
		    }
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_CLEAN()) {
214 215 216
			print "$node never booted; skipping $mode.\n";
			next;
		    }
217 218 219 220
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
			print "$node will $mode with local node $pnode.\n";
			next;
		    }
221 222
		}
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
223
		    print "$node is already setting up on local node $pnode\n";
224 225 226
		    next;
		}
	    }
227 228
	    else {
		if ($killmode) {
229 230 231
		    if ($allocstate eq TBDB_ALLOCSTATE_DEAD) {
			# plab only. See below.
			print "$node failed to initialize; skipping $mode.\n";
232 233
			next;
		    }
234 235 236 237 238 239 240 241 242
		    elsif ($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";
			$mode = "cleanup";
		    }
243 244 245 246
		    elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_CLEAN()) {
			print "$node never booted; skipping $mode.\n";
			next;
		    }
247
		}
248 249
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_READY()) {
		    print "$node is already set up on remote node $pnode\n";
250 251
		    next;
		}
252
		elsif ($allocstate eq TBDB_ALLOCSTATE_RES_INIT_DIRTY()) {
253
		    print "$node needs a reboot on remote node $pnode\n";
254 255
		    $mode = "reboot";
		}
256 257
	    }
	}
258 259 260 261 262 263
    }

    #
    # When setting up a vnode, force its event state into SHUTDOWN since
    # no telling what its initial state is. 
    # 
264
    if ($mode eq "teardown" || $mode eq "reboot") {
265 266
	TBSetNodeEventState($node, TBDB_NODESTATE_SHUTDOWN);
    }
267 268

    #
269 270 271 272 273 274
    # Put this into the list of calls we have to make in the next loop
    #
    push @vnodes, [$node, $pnode, $mode, $jailed, $plab];

}

275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296

sub plabnode($$$$) {
    my ($mode, $pid, $eid, $vnode) = @_;

    my $plabpid = 0;
    $plabpid = fork();
        
    if ($plabpid) {
        local $SIG{TERM} = 
            sub { 
                kill("TERM", $plabpid); 
                print STDERR "*** $0:\n".
                    "    Plab node allocation terminated.\n"; 
                exit(1);
            };
        
        my $childpid = wait();
        if ($childpid < 0) {
            die("*** $0:\n".
                "    Uh oh, wait() returned a negative number");
        }
        elsif ($childpid != $plabpid) {
297 298
            warn("*** $0:\n".
                "    Hmm, pid returned by wait() != pid ".
299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318
                "from fork().");
        }
        
        if ($?) {
            #print STDERR "*** $0:\n" .
            #    "    Plab node $mode failed\n";
            # Should check DB state instead.
            return($?);
        }
    }
    else {
        exec("$TB/sbin/plabnode $mode $pid $eid $vnode");    
        die("*** $0:\n".
            "    exec failed!\n");
    }

    return(0);
}


319 320 321 322 323 324 325 326 327 328 329
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;
    }
330

331 332 333
    #
    # There are more free slots
    #
334
    if (($children <= $MAX_CHILDREN) && @vnodes) {
335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361
	#
	# 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 {
	    # Must change our real UID to root so that ssh will work.
	    $UID = 0;

362 363 364
            # If this is a plab vnode, we need to allocate it first.
            # This will give us a vserver on the physical plab node to
            # work in.
365
	    if ($plab && $mode eq "setup") {
366 367 368 369 370 371 372
                if(plabnode("alloc", $pid, $eid, $vnode)) {
                    print STDERR "*** $0:\n" .
                                 "    Plab node allocation failed\n";
                    # Should check DB state instead.
                    exit(99);
                }
            }
373

374 375 376 377 378 379 380 381 382 383 384 385 386
	    # 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 ";
	
		exec("$ssh -host $vnode $CLIENT_BIN/vnodesetup $args");
		die("*** $0:\n".
		    "    exec failed!\n");
	    }
	    exit(0);
387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418
	}
    } 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.
	#
	my $oldest = 0;
	my $oldestpid = 0;
	while (my ($pid, $aref) = each %child_vnodes) {
	    my ($vnode, $pnode, $mode, $jailed, $plab, $birthtime) = @$aref;
	    if ($birthtime > $oldest) {
		$oldest = $birthtime;
		$oldestpid = $pid;
	    }
	}

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

	#
	# Kill of the oldest if he gets too old while we're waiting
	#
419 420
        local $SIG{ALRM} = sub { kill("TERM", $oldestpid); };
        alarm(($oldest + 120) - time());
421 422

	my $childpid = wait();
423
	my $exitstatus = $?;
424 425
	alarm 0;

426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443
	#
	# 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};

444
	#
445
	# Look for setup failure, reported back through ssh.
446
	# 
447
	if ($exitstatus) {
448 449
	    print STDERR "vnode $vnode $mode on $pnode returned $?.\n"
		if $debug;
450

451
	    if ($exitstatus == 256) {
452
		print STDERR "$vnode is not running sshd.\n" if $debug;
453 454
	    }
	    elsif ($exitstatus == 15) {
455
		print STDERR "$vnode is wedged.\n" if $debug;
456
	    }
457 458 459
	    elsif ($exitstatus >> 8 == 99) {
		print STDERR "$vnode did not allocate properly.\n" if $debug;
	    }
460

461 462
	    push @failed_nodes, [$vnode, $pnode, $mode, $exitstatus];
	    
463 464
	    warn("*** $0:\n".
		 "    Virtual node $vnode $mode failure!\n");
465
	}
466

467 468 469 470 471 472 473 474 475 476
	if ($plab) {
	    #
	    # I am totally unhappy with this mess.
	    #
	    if ((($mode eq "teardown") || ($mode eq "cleanup")) ||
		($exitstatus && (($exitstatus >> 8) != 99))) {
		#
		# Besides teardown/cleanup, we want to do this if the
		# vnode setup on the node failed, but only if the
		# plabnode alloc worked (it can fail too, exits with 99).
477 478
		#
                if (plabnode("free", $pid, $eid, $vnode)) {
479 480 481 482 483 484 485 486 487 488 489 490 491 492
		    warn("*** $0:\n".
			 "    Plab node free of $vnode failed");
		}
	    }
	    
	    #
	    # 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());
493 494
	    }
	}
495 496 497
    }
}

498 499 500 501 502 503 504
#
# 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",
	"The following virtual nodes failed: \n" .
Robert Ricci's avatar
Robert Ricci committed
505
	"vnode\t\tpnode\tmode\texit status\n" .
506 507 508
	$failed_lines);
}

509 510 511 512 513 514
if ($killmode) {
    print STDOUT "Vnode teardown finished.\n";
}
else {
    print STDOUT "Vnode setup initiated on all nodes ...\n";
}
515
exit 0;