tbswap.in 43.7 KB
Newer Older
1 2 3 4
#!/usr/bin/perl -w

#
# EMULAB-COPYRIGHT
5
# Copyright (c) 2000-2009 University of Utah and the Flux Group.
6 7 8 9
# All rights reserved.
#
use English;

10 11 12
# Returns 0 on success.
# Returns 1 on non-assign_wrapper failure.
# Returns (1 | assign_wrapper's errorcode) on assign_wrapper failure.
13 14
# Returns |0x40 if update caused a swapout. Icky.
# Returns -1 on uncontrolled error (die called).
15

16 17 18 19 20
# XXX: handle error cases for update? (backup the db?)
# XXX: Shouldn't do idempotent stuff twice for update.
# XXX: repush/calc routing for update??? (tbprerun)
# XXX: previz for update???              (tbprerun)
# XXX: make snmpit faster for update.
Chad Barb's avatar
Chad Barb committed
21 22 23
#
# XXX: for update, expt is swapped out on os_setup fail.
#      (we only recover if assign fails)
24 25 26

sub usage()
{
27 28 29
    print STDERR
	"Usage: $0 {in|out|modify [-reboot] [-eventsys_restart]} ".
	"[-noswapout] [-genimode] pid eid\n";
30 31 32 33 34 35 36
    exit(-1);
}

#
# Configure variables
#
my $TBROOT         = "@prefix@";
37
my $TBOPS          = "@TBOPSEMAIL@";
Timothy Stack's avatar
Timothy Stack committed
38
my $TBLOGS         = "@TBLOGSEMAIL@";
39
my $MAINSITE	   = @TBMAINSITE@;
40
my $THISHOMEBASE   = "@THISHOMEBASE@";
41 42
my $TESTMODE       = @TESTMODE@;
my $DISABLE_EVENTS = "@DISABLE_EVENT_SCHED@";
43
my $piper          = "$TBROOT/sbin/locpiper";
44 45
my $mapper         = "$TBROOT/bin/mapper";
my $wrapper        = "$TBROOT/libexec/assign_wrapper";
Timothy Stack's avatar
Timothy Stack committed
46
my $NFSTRACESUPPORT= @NFSTRACESUPPORT@;
47 48 49 50 51 52 53 54 55 56 57

# Untaint the path
$ENV{'PATH'} = "/usr/bin:$TBROOT/libexec:$TBROOT/libexec/ns2ir" . 
    ":$TBROOT/sbin:$TBROOT/bin";

#
# Testbed Support libraries
#
use lib "@prefix@/lib";
use libdb;
use libtestbed;
58
use libadminctrl;
59
use libadminmfs;
60
use libtblog;
Leigh B. Stoller's avatar
Leigh B. Stoller committed
61
use libArchive;
62
use Experiment;
63
use User;
64
use Lan;
65
use Firewall;
66

67
#require exitonwarn; # exitonwarn isn't really a module, so just require it
68 69 70 71

#
# Actual swap-in and swap-out functions, defined below.
#
72 73 74
sub doSwapout($);
sub doSwapin($);

75
# XXX fixme: should not be hardwired!
76
my $cnetstack = "-S Control";
77 78 79
my $cnetvlanname = "Control";


80 81 82 83
sub REAL()    { return 5; }
sub CLEANUP() { return 4; }
sub RETRY()   { return 3; }
sub MODIFY()  { return 2; }
84
sub UPDATE()  { return 1; }
85
sub MODIFY_RECOVER()  { return 0; }
86

87 88 89
#
# Grab global enable of linkdelays.
#
90
my $enablelinkdelays = TBGetSiteVar("general/linux_endnodeshaping");
91

92 93 94 95 96 97
#
# Turn off line buffering on output
#

$| = 1;

98 99
my $updateReboot   = 0;
my $updateReconfig = 1;
100 101
my $update_Eventsys_restart = 0;
my $elabinelab     = 0;
102
my $plabinelab     = 0;
103 104 105 106 107
my $force          = 0;
my $noswapout      = 0;
my $genimode       = 0;
my $errors         = 0;
my $updatehosed    = 0;
108
my $state;
109
my $canceled;
110
my $os_setup_pid;
111
my $nextState;
112 113

#
114
# First argument is either "in", "out", or "update";
115 116 117 118 119
# this value goes into $swapop.
#

my $swapop = shift;	

120 121 122
if (!$swapop || 
    (($swapop ne "in") && 
     ($swapop ne "out") &&
123
     ($swapop ne "modify") &&
124
     ($swapop ne "update"))) {
125 126 127 128 129 130 131 132 133 134 135
    usage();
}

#
# Get other arguments.
#

while ($#ARGV > 1) {
    $arg = shift;
    if ($arg eq "-force") {
	$force = 1;
Chad Barb's avatar
Chad Barb committed
136 137
    } elsif ($arg eq "-reboot") {
	$updateReboot = 1;
138
	$updateReconfig = 0;
139 140 141
    } elsif ($arg eq "-noreconfig") {
	$updateReboot   = 0;
	$updateReconfig = 0;
142
    } elsif ($arg eq "-eventsys_restart" && $swapop eq "modify") {
143
	$update_Eventsys_restart = 1;
144
    } elsif ($arg eq "-noswapout") {
145
	$noswapout = 0;
146 147 148 149 150 151 152 153 154
    } else {
	usage();
    }
}
if ($#ARGV < 1) {
    usage();
}
my ($pid,$eid) = @ARGV;

155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
#
# Untaint the arguments.
#
if ($pid =~ /^([-\@\w.]+)$/) {
    $pid = $1;
}
else {
    die("Tainted argument $pid!\n");
}
if ($eid =~ /^([-\@\w.]+)$/) {
    $eid = $1;
}
else {
    die("Tainted argument $eid!\n");
}

171 172 173 174 175 176
#
# Set Error reporting info
# 
tblog_set_info($pid,$eid,$UID);

#
177
# Turn on timestamps
178
#
179 180
TBDebugTimeStampsOn();

181 182 183 184 185 186
#
# Verify user and get his DB uid and other info for later.
#
my $this_user = User->ThisUser();
if (! defined($this_user)) {
    tbdie("You ($UID) do not exist!");
187
}
188 189 190
my $user_uid   = $this_user->uid();
my $user_name  = $this_user->name();
my $user_email = $this_user->email();
191

192 193 194 195 196 197
# Slowly convert to using Experiment module.
my $experiment = Experiment->Lookup($pid, $eid);
if (!defined($experiment)) {
    tbdie("Could not lookup experiment object!")
}

198 199 200
#
# Print starting message.
#
201 202
my $exptidx;
TBExptIDX($pid, $eid, \$exptidx);
203

204 205
print "Beginning swap-$swapop for $pid/$eid ($exptidx). " .
    TBTimeStampWithDate() . "\n";
206 207 208 209 210 211
TBDebugTimeStamp("tbswap $swapop started");

#
# Get experiment state; verify that experiment exists.
#
if (! ($state = ExpState($pid, $eid))) {
212
    tbdie "No such experiment $pid/$eid";
213
}
214 215 216
# Sanity check the current state. 
if (!$force) {
    if ($swapop eq "in") {
217
	tbdie("Experiment should be ACTIVATING. Currently $state.")
218
	    if ($state ne EXPTSTATE_ACTIVATING);
219
    }
220
    elsif ($swapop eq "out") {
221
	tbdie("Experiment should be SWAPPING. Currently $state.")
222
	    if ($state ne EXPTSTATE_SWAPPING);
223
    }
224
    elsif ($swapop eq "modify" || $swapop eq "update") {
225
	tbdie("Experiment should be MODIFY_RESWAP. Currently $state.")
226
	    if ($state ne EXPTSTATE_MODIFY_RESWAP);
227 228
    }
}
229 230
# Get elabinelab status. See below.
if (! TBExptIsElabInElab($pid, $eid, \$elabinelab)) {
231
    tbdie("Could not get elabinelab status for experiment $pid/$eid");
232
}
233 234 235 236 237 238 239 240 241 242 243
# and plabinelab status.
if (! TBExptIsPlabInElab($pid, $eid, \$plabinelab)) {
    tbdie("Could not get plabinelab status for experiment $pid/$eid");
}
if ($elabinelab && $plabinelab) {
    tberror "Cannot get my head around Plab in Elab in Elab!\n";
    print "Failingly finished swap-$swapop for $pid/$eid. " .
	TBTimeStamp() . "\n";
    TBDebugTimeStamp("tbswap $swapop finished (failed)");
    exit(1);
}
244

245 246 247 248 249
#
# See if the experiment is firewalled
#
my $firewalled = TBExptFirewall($pid, $eid);

250 251 252
#
# Do actual swapping
#
253 254 255 256
if ($swapop eq "out") {
    #
    # Swap out
    #
257
    tblog_set_attempt(1);
258
    $errors = doSwapout(REAL);
259
}
260
elsif ($swapop eq "update" || $swapop eq "modify") {
261 262 263 264 265 266
    #
    # Before swapout, do cursory admission control to see if the
    # modified experiment will be swappable. assign_wrapper does a
    # more stringent check using assign.
    #
    print STDERR "Checking with Admission Control ...\n";
267
    if (! TBAdmissionControlCheck(undef, $experiment, undef)) {
268 269 270
	tberror({type => 'secondary', severity => SEV_SECONDARY,
		 error => ['admission_control_failure']},
		"Admission control failure!\n");
271 272 273 274 275 276
	print "Failingly finished swap-$swapop for $pid/$eid. " .
	    TBTimeStamp() . "\n";
	TBDebugTimeStamp("tbswap $swapop finished (failed)");
	exit(1);
    }
    
277 278
    #
    # Update.
279 280 281
    #
    tblog_set_attempt(1);

282 283 284 285
    #
    # Phase One -- swap experiment partially out.
    #
    print STDERR "Backing up physical state...\n";
286
    $experiment->BackupPhysicalState();
287

288 289 290 291 292
    #
    # Actually, in update mode this is not done cause we are anticipating
    # adding nodes only.
    #
    $errors = ($swapop eq "modify" ? doSwapout(MODIFY) : 0);
Chad Barb's avatar
Chad Barb committed
293

294 295 296 297 298 299
    if (0) {
	print STDERR "Doing a swapmodswapout on the experiment archive ...\n";
	if (libArchive::TBExperimentArchiveSwapModSwapOut($pid, $eid) < 0) {
	    tberror("Failed swapmodswapout on the experiment archive!");
	    $errors = 1;
	}
Leigh B. Stoller's avatar
Leigh B. Stoller committed
300 301
    }

302
    if ($errors) {
303
	#
304
	# Clean up the mess, leaving the experiment in the SWAPPED state,
305 306 307
	# 
	print STDERR "Cleaning up after errors.\n";
	doSwapout(CLEANUP);
308
	$updatehosed = 1;
309
    }
310
    else {
311
	#
312
	# Phase Two -- swap experiment back in.
313
	#
314
	$errors = doSwapin(MODIFY);
Chad Barb's avatar
Chad Barb committed
315

316
	if ($errors) {
Chad Barb's avatar
Chad Barb committed
317 318 319
	    #
	    # There were errors; see if we can recover.
	    #
320
	    my $CanRecover = 1;
Chad Barb's avatar
Chad Barb committed
321 322 323

	    if ($errors != 7) {
		print STDERR "Update failure occurred _after_ assign phase; ";
324
		$CanRecover = 0;
Chad Barb's avatar
Chad Barb committed
325 326
	    }

327 328 329
	    if ($CanRecover) {
		print STDERR "Recovering virtual and physical state.\n";

330 331 332
		# It is safe to remove the phystate since we know it was
		# backed up above, and cause we do not know if assign_wrapper
		# made it to that point before it failed. 
333 334 335 336
		if ($experiment->RemoveVirtualState($pid, $eid) ||
		    $experiment->RestoreVirtualState($pid, $eid) ||
		    $experiment->RemovePhysicalState($pid, $eid) ||
		    $experiment->RestorePhysicalState($pid,$eid)) {
337 338
		    print STDERR "Could not restore backed-up state; ";
		    $CanRecover = 0;
Chad Barb's avatar
Chad Barb committed
339
		}
340 341 342
		else {
		    print STDERR "Doing a recovery swap-in of old state.\n";

343
		    if (doSwapin(MODIFY_RECOVER)) {
344 345 346
			print STDERR "Could not swap in old physical state; ";
			$CanRecover = 0;
		    }
Chad Barb's avatar
Chad Barb committed
347 348
		}
	    }
349 350 351 352 353 354

	    #
	    # Some part of the recovery failed; must swap it out. swapexp
	    # (caller) will then have to do more clean up, hence the special
	    # exit status indicated by $updatehosed.
	    # 
355 356 357 358 359 360 361 362 363 364 365
	    if (!$CanRecover) {
		if ($noswapout) {
		    print STDERR
			"No Recovery, but leaving experiment swapped in.\n";
		}
		else {
		    print STDERR
			"Recovery aborted! Swapping experiment out.\n";
		    doSwapout(CLEANUP);
		    $updatehosed = 1;
		}
366 367 368
	    }
	    else {
		print STDERR "Update recovery successful.\n";
Chad Barb's avatar
Chad Barb committed
369
	    }
370
	}
Chad Barb's avatar
Chad Barb committed
371 372
    }
}
373 374 375 376 377
elsif ($swapop eq "in") {
    #
    # Swap in
    #
    my $retries = 2;
378 379 380 381 382 383

    #
    # Before real swapin, do cursory admission control. assign_wrapper does
    # a more stringent check using assign.
    #
    print STDERR "Checking with Admission Control ...\n";
384
    if (! TBAdmissionControlCheck(undef, $experiment, undef)) {
385 386 387
	tberror({type => 'secondary', severity => SEV_SECONDARY,
		 error => ['admission_control_failure']},
		"Admission control failure!\n");
388 389 390 391 392
	print "Failingly finished swap-$swapop for $pid/$eid. " .
	    TBTimeStamp() . "\n";
	TBDebugTimeStamp("tbswap $swapop finished (failed)");
	exit(1);
    }
393 394

    tblog_set_attempt(1);
395
    $errors = doSwapin(REAL);
Chad Barb's avatar
Chad Barb committed
396

397 398 399 400 401 402 403 404 405 406
    #
    # Attempt a retry if: 
    #   a) there were errors, 
    #   b) doswapin() indicated (via return code 3) a retry is appropriate,
    #   c) we haven't tried too many times already.
    #   d) The cancelflag has not been set.
    #   e) $TESTMODE == 0.
    #
    while ($errors == 3 && $retries && !$canceled && !$TESTMODE) {
	$retries--;
407
	tblog_inc_attempt();
408

409 410
	print STDERR "Cleaning up after errors; will try again.\n";
	doSwapout(RETRY);
411

412 413
	print STDERR "Trying again...\n";
	$errors = doSwapin(RETRY);
414
    }
415 416
    if ($errors || $canceled) {
	print STDERR "Cleaning up after " .
417
	    ($canceled ? "cancelation" : "errors") . ".\n";
418
	doSwapout(CLEANUP);
419 420 421
    }
}

422 423 424 425
tblog_set_attempt(0);

TBGetCancelFlag($pid, $eid, \$canceled);
if ($canceled) {
426 427
    tberror ({type=>'summary', cause=>'canceled', 
	      severity=>SEV_IMMEDIATE, error=>['cancel_flag']},
428 429 430
	     "Experiment swap-$swapop canceled by user.");
}

431 432 433 434
#
# Write appropriate message and exit.
#
if ($errors) {
435
    print "Failingly finished swap-$swapop for $pid/$eid. ".TBTimeStamp()."\n";
436
    TBDebugTimeStamp("tbswap $swapop finished (failed)");
Chad Barb's avatar
Chad Barb committed
437

438 439
    # Pass out magic value to indicate that update failed!
    exit(1 | ($updatehosed ? 0x40 : 0));
440
}
441
print "Successfully finished swap-$swapop for $pid/$eid. " .TBTimeStamp()."\n";
442 443
TBDebugTimeStamp("tbswap $swapop finished (succeeded)");
exit(0);
444 445 446 447

#################################

##
Chad Barb's avatar
Chad Barb committed
448
#
449 450
# doSwapout - Swaps experiment out.
#
Chad Barb's avatar
Chad Barb committed
451
#             If in REAL or CLEANUP,
452 453 454
#             this function will free all nodes for the 
#             experiment.
#
Chad Barb's avatar
Chad Barb committed
455
#             If in RETRY or UDPATE,
456 457 458 459 460 461
#             only nodes not in RES_READY will be freed.
#
#             Returns 0 on success, >0 on failure.
#
##

462
sub doSwapout($) {
463
    my $type = shift;
464 465
    my $swapout_errors = 0;

466 467
    tblog_set_cleanup(1) if $type == CLEANUP;

468 469 470 471 472 473 474 475 476
    #
    # wait for os_setup;
    # this only applies if called after a failed doswapin.
    #
    if ($os_setup_pid) {
	print "Waiting for os_setup to finish\n";
	waitpid($os_setup_pid, 0);
	undef $os_setup_pid;
    }
Chad Barb's avatar
Chad Barb committed
477

478
    if (0 && $NFSTRACESUPPORT && $type == REAL) {
Timothy Stack's avatar
Timothy Stack committed
479 480
	print "Getting files accessed via NFS.\n";
	TBDebugTimeStamp("nfstrace started");
481
	system("nfstrace transfer $pid $eid");
Timothy Stack's avatar
Timothy Stack committed
482
	TBDebugTimeStamp("nfstrace finished");
483
    }
484

485
    if (! $TESTMODE) { 
486 487
	if (! ($DISABLE_EVENTS || $elabinelab)) {
	    if ($type >= RETRY ||
488
		($update_Eventsys_restart && $type == MODIFY) ) {
489
		print "Stopping the event system\n";
490
		if (system("eventsys_control stop $pid,$eid")) {
491 492 493
		    tberror({type => 'secondary', severity => SEV_SECONDARY,
			     error => ['eventsys_stop_failed']},
			    "Failed to stop the event system.");
494 495
		    $swapout_errors = 1;
		}
496 497 498 499 500 501 502

		#
		# Stop the location piper.
		#
		if (-x $piper) {
		    print "Stopping the location piper\n";
		    if (system("$piper -k $pid $eid")) {
503 504 505
			tberror({type => 'secondary', severity => SEV_SECONDARY,
				 error => ['piper_stop_failed']},
				"Failed to stop location piper.");
506 507 508
			$swapout_errors = 1;
		    }
		}
509 510
	    }
	}
511 512 513 514 515 516
	
	#
	# Do teardown of inner elab. We must do this before we teardown the
	# vlans since the inner control network is a vlan, and we want that
	# active so inner boss can reboot the inner nodes (avoid power cycle).
	#
517
	if ($elabinelab && $type >= CLEANUP) {
518 519
	    print "Tearing down elabinelab. This could take a while.\n";
	    if (system("elabinelab -k $pid $eid")) {
520 521 522
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['elabinelab_tear_down_failed']},
			"Failed to teardown elabinelab!");
523 524 525 526
		$swapout_errors = 1;
	    }
	}

527
	#
Chad Barb's avatar
Chad Barb committed
528
	# Clean up any VLANs in experiment.
529
	#
530
	# When modifying an elabinelab experiment, leave the vlans intact
531
	# so that the inner networks are not suddenly disconnected.
532
	#
533
	if ($type != MODIFY || ($type == MODIFY && Lan->GotTrunks($experiment))) {
534 535 536
	    TBDebugTimeStamp("snmpit started");
	    print STDERR "Removing VLANs.\n";
	    if (system("snmpit -r $pid $eid")) {
537 538 539
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['vlan_reset_failed']},
			"Failed to reset VLANs");
540 541 542
		$swapout_errors = 1;
	    }
	    TBDebugTimeStamp("snmpit finished");
543
	}
544 545 546 547 548 549 550 551 552 553 554 555 556 557
	#
	# Must check for stale vlans that we kept around in the above clause
	# since they will not be in the lans table anymore.
	#
	if ($type == CLEANUP) {
	    my @stale;
	    if (VLan->StaleVlanList($experiment, \@stale) != 0) {
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['vlan_reset_failed']},
			"Failed to get stale VLANs");
		$swapout_errors = 1;
	    }
	    if (@stale) {
		print "Removing stale vlans @stale\n";
558
		system("snmpit -f ". join(" ", map("-o $_", @stale)));
559 560 561 562 563 564 565 566
		if ($?) {
		    tberror({type => 'summary', severity => SEV_SECONDARY,
			     error => ['vlan_reset_failed']},
			    "Failed to remove stale vlans");
		    $swapout_errors = 1;
		}
	    }
	}
567
    }
568
	
569
    if ($type >= CLEANUP) {
570 571 572 573 574 575 576 577
	#
	# Undo plab in elab specialness.
	# No need to worry about VLANs here, as all the special abilities
	# involve the control network.
	#
	if (! $TESTMODE && $plabinelab) {
	    print "Tearing down plabinelab.\n";
	    if (system("plabinelab -k $pid $eid")) {
578 579 580
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['plabinelab_tear_down_failed']},
			"Failed to teardown plabinelab!");
581 582 583 584
		$swapout_errors = 1;
	    }
	}

585 586
	#
	# We're not attempting a retry;
587
	#
588
	# Stop all of the vnodes.
589
	#
590
	if (! $TESTMODE) { 	
591 592 593
	    print "Tearing down virtual nodes.\n";
	    TBDebugTimeStamp("vnode_setup -k started");
	    if (system("vnode_setup -d -k $pid $eid")) {
594 595 596
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['vnode_tear_down_failed']},
			"Failed to tear down vnodes.");
597 598 599
		$swapout_errors = 1;
	    }
	    TBDebugTimeStamp("vnode_setup finished");
600 601
	}

602 603 604 605
	#
	# Nodes behind a firewall are treated special.
	# See undoFWNodes for details.
	#
606
	if ($firewalled && undoFWNodes($experiment)) {
607
	    tblog_set_cleanup(0);
608 609
	    return 1;
	}
Leigh B. Stoller's avatar
Leigh B. Stoller committed
610

611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630
	#
	# Perform swapout time admin actions.  Right now there is at most
	# one of these.  It isn't really a general mechanism, just a hook
	# for state saving or data collection during swapout.
	# A couple of important "fer now" notes:
	#
	#	We don't do this for firewalled experiments.  We need a way
	#	to "tag" the saved disk state to ensure it doesn't get
	#	instantiated outside of a firewall.
	#
	#	We only do this on REAL swapouts, and not on CLEANUPs.
	#	There are some types of CLEANUPs where we may want to
	#	do this, in particular an invocation caused by a failed
	#	modify operation, where the admin action is to save the
	#	experiment state.  So we will need to revisit this.
	#
	my %soaction = ();
	if ($type == REAL && !$firewalled) {
	    TBExptGetSwapoutAction($pid, $eid, \%soaction);
	}
631
	if ($soaction{'command'} && doSwapoutAction($pid, $eid, %soaction)) {
632
	    tblog_set_cleanup(0);
633
	    return 1;
634 635
	}

636 637
	#
	# remove all nodes from the experiment.
Chad Barb's avatar
Chad Barb committed
638
	# (nfree will send them to RES_FREE_DIRTY)
639 640 641
	#
	print STDERR "Freeing nodes.\n";
	TBDebugTimeStamp("nfree started");
642
	if (system("nfree -a $pid $eid")) {
643 644 645
	    tberror({type => 'secondary', severity => SEV_SECONDARY,
		     error => ['nfree_failed']},
		    "Could not free nodes.");
646 647 648
	    $swapout_errors = 1;
	}
	TBDebugTimeStamp("nfree finished");
649 650 651 652 653

	#
	# Since this is an actual swapout, 
	# reset our count of swap out nag emails sent.
	#
654 655
	DBQueryWarn("update experiments set swap_requests='', ".
		    "   sim_reswap_count='0' ".
656
		    "where eid='$eid' and pid='$pid'");
657 658
    } else {
	#
659
	# $type == RETRY or $type == MODIFY.
660 661
	# Therefore, don't deallocate nodes which have been successfully
	# incorporated into the experiment (i.e., are RES_READY).
Chad Barb's avatar
Chad Barb committed
662
	# (nfree will send deallocated nodes to RES_FREE_DIRTY)
663
	#
664 665 666 667 668 669 670 671
	my @failedpnodes = ();
	my @failedvnodes = ();
	
	my $db_result =
	    DBQueryFatal("select rv.node_id,n.allocstate,nt.isvirtnode ".
                         "  from reserved as rv ".
			 "left join nodes as n on n.node_id = rv.node_id ".
			 "left join node_types as nt on nt.type=n.type ".
672 673
			 "where rv.pid='$pid' and rv.eid='$eid' and ".
			 "      rv.genisliver_idx is null");
674

675
	while (my ($node,$allocstate,$isvirt) = $db_result->fetchrow_array) {
676
	    if ($allocstate ne TBDB_ALLOCSTATE_RES_READY()) {
677 678 679 680 681 682
		if ($isvirt) {
		    push(@failedvnodes, $node);
		}
		else {
		    push(@failedpnodes, $node);
		}
683 684 685
	    }
	}

686 687 688 689 690 691 692
	#
	# Tear down failed vnodes. Perhaps not needed?
	# 
	if (!$TESTMODE && @failedvnodes > 0) {
	    print "Tearing down failed virtual nodes.\n";
	    TBDebugTimeStamp("vnode_setup -k started");
	    if (system("vnode_setup -d -k $pid $eid @failedvnodes")) {
693 694 695
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['vnode_tear_down_failed']},
			"Failed to tear down vnodes.");
696 697 698 699
		$swapout_errors = 1;
	    }
	    TBDebugTimeStamp("vnode_setup -k finished");
	}
700

701 702 703 704 705 706 707 708 709 710 711
	#
	# Undo plabinelab setup
	#
	if (!$TESTMODE && $plabinelab && @failedpnodes > 0) {
	    print "Removing failed nodes from inner plab.\n";
	    if (system("plabinelab -r $pid $eid @failedpnodes")) {
		tberror "Failed to remove inner nodes!";
		$swapout_errors = 1;
	    }
	}

712 713 714 715
	#
	# Release all failed nodes.
	# 
	if (@failedpnodes > 0 || @failedvnodes > 0) {
716 717
	    print STDERR "Freeing failed nodes.\n";
	    
718
	    TBDebugTimeStamp("nfree started");
Chad Barb's avatar
Chad Barb committed
719 720 721 722 723
	    #
	    # Specify -x switch so when a physical node gets freed,
	    # any virtual nodes (owned by this experiment)
	    # sitting on top of it are freed as well.
	    #
724 725
	    if (system("nfree -x $pid $eid " .
		       join(" ", (@failedpnodes, @failedvnodes)))) {
726 727 728
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['nfree_failed']},
			"Could not free nodes.");
729 730 731 732 733 734
		$swapout_errors = 1;
	    }
	    TBDebugTimeStamp("nfree finished");
	}
    }

735 736 737 738 739 740 741 742 743 744 745 746
    if (! $TESTMODE) {
	#
	# If the experiment has no Plab dslice nodes left, but still has
	# a Plab slice, destroy the slice
	#

	# Does the slice exist?
	$db_result =
	    DBQueryFatal("select slicename from plab_slices ".
			 "where pid='$pid' and eid='$eid'");

	if ($db_result->numrows) {
747
	    # Are there any nodes left in the slice?
748
	    $db_result =
749 750 751 752 753 754 755 756 757
		DBQueryFatal("select r.node_id ".
			     " from plab_slices as ps".
			     " left join plab_slice_nodes as psn ".
			     "   on (ps.slicename=psn.slicename ".
			     "       and ps.plc_idx=psn.plc_idx) ".
			     " left join reserved as r ".
			     "   on psn.node_id=r.node_id ".
			     " where ps.pid='$pid' and ps.eid='$eid'".
			     "   and r.node_id is not NULL");
758 759

	    if (!$db_result->numrows) {
760
		print "Tearing down Slices.\n";
761
		TBDebugTimeStamp("plabslice destroy started");
762
		if (system("plabslice destroy $pid $eid")) {
763
		    tberror "Failed to tear down Slices.";
764 765 766 767 768 769 770
		    $swapout_errors = 1;
		}
		TBDebugTimeStamp("plabslice destroy finished");
	    }
	}
    }

771 772 773 774 775 776 777 778
    if (! $TESTMODE) {
	#
	# All of these errors are non-fatal on swapout. We find out about them
	# via email sent from the individual scripts.
	#

	#
	# Only reset mountpoints if this is an actual swapout, and
779
	# not a failed swapin(cleanup), update, or retry.
780
	#
781
	if ($type == REAL) {
782 783 784
	    print "Resetting mountpoints.\n";
	    TBDebugTimeStamp("exports started");
	    if (system("exports_setup")) {
785 786 787
		tberror({severity => SEV_WARNING,
			 error => ['mountpoint_reset_failed']},
			"Failed to reset mountpoints.");
788 789
	    }
	    TBDebugTimeStamp("exports finished");
Chad Barb's avatar
Chad Barb committed
790

791 792 793 794 795 796 797 798 799
	    #
	    # Ditto these two. 
	    #
	    print "Resetting named maps.\n";
	    TBDebugTimeStamp("named started");
	    if (system("named_setup")) {
		tbwarn "Failed to reset named map.";
	    }
	    TBDebugTimeStamp("named finished");
Chad Barb's avatar
Chad Barb committed
800

801 802 803 804 805 806
	    print "Resetting email lists.\n";
	    TBDebugTimeStamp("genelists started");
	    if (system("genelists -t")) {
		tbwarn "Failed to reset email lists.";
	    }
	    TBDebugTimeStamp("genelists finished");
807 808 809
	}
    }

810
    #
811
    # Wipe the DB clean except during MODIFY or RETRY. In those
812 813
    #    cases, assign_wrapper will reset the DB after reading
    #    the info.
814
    #
815 816
    if ( $type >= CLEANUP ) {
	print STDERR "Resetting DB.\n";
817 818 819
	$experiment->RemovePhysicalState();
	# Special. 
	$experiment->ClearPortRegistration();
820
    }
821

822
    tblog_set_cleanup(0);
823 824 825 826
    return $swapout_errors;
}

##
Chad Barb's avatar
Chad Barb committed
827
#
828 829
# doSwapin - Swaps experiment in.
#
Chad Barb's avatar
Chad Barb committed
830 831 832 833 834
#            Returns:
#              0 - successful swapin
#              1 - failed swapin; cleanup required.
#              3 - failed swapin; cleanup required; can retry.
#              7 - failed swapin; assign failed; no cleanup.
835 836
##

837
sub doSwapin($) {
838
    my $type = shift; 
839 840
    # Just the physnodes ...
    my @deleted_pnodes = ();
841

842 843 844 845
    #
    # assign_wrapper does all the virtual to physical mapping 
    # and updating the DB state.
    #
846
    
847
    if ($type > MODIFY_RECOVER) {
848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863
        #
        # Hacky test to allow disabling of linkdelays if the node is going
        # to run Linux. See sitevar above.
        #
	if (! $enablelinkdelays) {
	    $db_result =
		DBQueryFatal("select distinct e.pid,e.eid,vl.vnode,vn.osname ".
			 "  from experiments as e ".
			 "left join virt_lans as vl on vl.pid=e.pid and ".
			 "     vl.eid=e.eid ".
			 "left join virt_nodes as vn on vn.pid=e.pid and ".
			 "     vn.eid=e.eid and vn.vname=vl.vnode ".
			 "left join os_info as o on o.osname=vn.osname and ".
			 "  (o.pid=vl.pid or o.pid='" . TBOPSPID() . "') ".
			 "where (vl.uselinkdelay!=0 or e.uselinkdelays!=0 or ".
			 "       e.forcelinkdelays!=0) and ".
864 865
			 "     (o.os is NULL or o.os='Linux' or ".
			 "      o.os='Fedora') and ".
866 867 868
			 "     e.pid='$pid' and e.eid='$eid'");

	    if ($db_result->numrows) {
869 870
		tberror "Endnodeshaping is disabled on Linux Images!";
		tberror "You must modify your experiment to swap it in.";
871 872 873 874
		return 1;
	    }
	}
	
Chad Barb's avatar
Chad Barb committed
875
	print "Mapping to physical reality ...\n";
876
	TBDebugTimeStamp("mapper wrapper started");
Chad Barb's avatar
Chad Barb committed
877 878

	#
879 880 881 882
	# Pass the -u (update) switch into assign_wrapper, which turns on
	# update mode. When doing a retry, must also fix the current nodes
	# to avoid stuff jumping around when simply trying to replace a node
	# that did not boot.
Chad Barb's avatar
Chad Barb committed
883 884
	#
	my $exitcode;
885 886 887 888 889 890
	my $cmd = "$wrapper";
	if ($pid eq "testbed" || $pid eq "tbres" || $pid eq "emulab-ops" ||
	    $this_user->stud()) {
	    $cmd = "$mapper";
	}
	$cmd .= " -f"
891
	    if ($type == RETRY);
892
	
893
	if (system("$cmd -u $pid $eid")) {
894 895
	    $exitcode = $? >> 8;

896
	    tberror "Failed ($exitcode) to map to reality.";
897

898 899
	    # Wrapper sets this bit when recovery is possible.
	    if ($exitcode & 64) {
900
		# We can recover. 
901
		tbreport(SEV_SECONDARY, 'assign_wrapper_failed', $exitcode);
Chad Barb's avatar
Chad Barb committed
902
		return 7;
903 904
	    }
	    else {
905
		# No recovery, no retry.
906
		tbreport(SEV_SECONDARY, 'assign_wrapper_failed', $exitcode);
Chad Barb's avatar
Chad Barb committed
907 908 909
		return 1;
	    }
	}
910
	TBDebugTimeStamp("mapper wrapper finished");
911

Chad Barb's avatar
Chad Barb committed
912 913
	print "Mapped to physical reality!\n";
    }
914

915
    # Check cancel flag before continuing. No retry, 
916
    TBGetCancelFlag($pid, $eid, \$canceled);
917 918 919 920
    if ($canceled) {
	tbreport(SEV_IMMEDIATE, 'cancel_flag');
	return 1
    }
921

922 923 924 925 926 927 928
    #
    # When doing a modify, we have to compare vlans to determine which
    # vlans actually changed and need to be deleted, before processing
    # the new vlans for the experiment. Note that vlans that already
    # exist on the switches will be left as is by snmpit.
    #
    # We must do this before the nfree of deleted nodes in the next section
929
    # because the new test in nfree the prevents nodes from accidentally
930 931 932 933 934
    # getting released when they are in a vlan.
    #
    if ($type == MODIFY) {
	my @diff = ();
	my @same = ();
935
	if (Lan->CompareVlansWithSwitches($experiment, \@diff, \@same) != 0) {
936 937 938 939 940
	    tberror({type => 'summary', severity => SEV_SECONDARY,
		     error => ['vlan_setup_failed']},
		    "Failed to compare old vlans");
	    return 1;
	}
941 942 943
	if (@diff) {
	    print "Removing obsolete vlans @diff\n";
	    system("snmpit -f ". join(" ", map("-o $_", @diff)));
944 945 946
	    if ($?) {
		tberror({type => 'summary', severity => SEV_SECONDARY,
			 error => ['vlan_setup_failed']},
947
			"Failed to remove old vlans");
948 949 950 951 952
		return 1;
	    }
	}
    }

953 954 955 956 957 958 959 960 961
    #
    # Look for any nodes in RES_TEARDOWN. These need to be released,
    # and if a virtnode, they need to be torn down. We cannot wait for
    # the virtnodes to go down with the physnode they are hosted on,
    # so teardown and release the virtnodes first, and then do the
    # physnodes.
    #
    # Errors are fatal; no recovery or retry.
    #
962
    if ($type == MODIFY || $type == UPDATE) {
963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987
	my $allocstate = TBDB_ALLOCSTATE_RES_TEARDOWN();
	
	$db_result =
	    DBQueryFatal("select r.node_id,nt.isvirtnode,nt.isremotenode ".
			 "  from reserved as r ".
			 "left join nodes as n on n.node_id=r.node_id ".
			 "left join node_types as nt on nt.type=n.type ".
			 "where r.pid='$pid' and r.eid='$eid' and ".
			 "      n.allocstate='$allocstate'");

	if ($db_result->numrows) {
	    my @virtnodes = ();
	    my @physnodes = ();
	    
	    print "Tearing down and releasing unused nodes\n";

	    # First teardown/release virtnodes. 
	    while (my ($node,$isvirt,$isrem) = $db_result->fetchrow_array()) {
		if ($isvirt) {
		    push(@virtnodes, $node);
		}
		elsif (!$isrem) {
		    push(@physnodes, $node);
		}
	    }
988 989 990
	    # See below.
	    @deleted_pnodes = @physnodes;
	    
991 992 993
	    if (@virtnodes) {
		TBDebugTimeStamp("vnode_setup started");
		
994
		if (system("vnode_setup -k $pid $eid @virtnodes")) {
995
		    tberror "Failed to tear down unused virtnodes!\n";
996 997 998 999 1000
		    return 1;
		}
		TBDebugTimeStamp("vnode_setup finished");
		
		if (system("nfree $pid $eid @virtnodes")) {
1001
		    tberror "Failed to nfree unused virtnodes!\n";
1002 1003 1004 1005
		    return 1;
		}
	    }
	    if (@physnodes) {
1006 1007 1008
		if ($elabinelab) {
		    print "Removing nodes from inner elab.\n";
		    if (system("elabinelab -r $pid $eid @physnodes")) {
1009
			tberror "Failed to remove inner nodes!";
1010 1011
			return 1;
		    }
1012 1013 1014 1015 1016 1017
		} elsif ($plabinelab) {
		    print "Removing nodes from inner plab.\n";
		    if (system("plabinelab -r $pid $eid @physnodes")) {
			tberror "Failed to remove inner nodes!";
			return 1;
		    }
1018 1019 1020
		}

		#
1021 1022
		# If the experiment is firewalled, cleanup the nodes
		# we are releasing.
1023
		# 
1024 1025
		if ($firewalled && undoFWNodes($experiment, 1,
					       @deleted_pnodes)) {
1026 1027 1028
		    return 1;
		}
		
1029
		if (system("nfree $pid $eid @physnodes")) {
1030
		    tberror "Failed to nfree unused physnodes!\n";
1031 1032 1033 1034 1035 1036
		    return 1;
		}
	    }
	}
    }

1037 1038 1039 1040 1041 1042
    # Exit here if we are testing.
    if ($TESTMODE) {
	print "Testing run - Stopping here.\n";
	return 0;
    }

1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053
    #
    # Handle tarballs - we might need to fetch some from URLs if the user
    # asked for that.
    #
    print "Fetching tarballs and RPMs (if any) ...\n";
    TBDebugTimeStamp("tarfiles_setup started");

    if (system("tarfiles_setup $pid $eid")) {
	#
	# No recovery for now - what would we do?
	#
1054 1055 1056
	tberror({type => 'secondary', severity => SEV_SECONDARY,
		 error => ['tarfiles_setup_failed']},
		"Failed to set up tarballs.");
1057 1058 1059 1060
	return 1;
    }
    TBDebugTimeStamp("tarfiles_setup finished");

1061 1062 1063 1064
    #
    # If there are any Plab dslice nodes in the experiment, create the
    # dslice now
    #
1065
    if ($type > MODIFY_RECOVER) {
1066 1067 1068
	# Are there any Plab nodes?  First get a list of node types in the exp;
	# if any are types hosted by any of the PLCs we know about, create
	# all slices necessary for the experiment in a single plabslice call.
1069
	$db_result =
1070 1071 1072 1073 1074 1075 1076 1077
	    DBQueryFatal("select nt.type,ppi.plc_name from nodes as n ".
			 "  left join node_types as nt on n.type = nt.type ".
			 "  left join reserved as r on r.node_id = n.node_id ".
			 "  left join plab_plc_info as ppi ".
			 "    on nt.type = ppi.node_type".
			 " where r.pid='$pid' and r.eid='$eid'".
			 "    and ppi.node_type is not NULL".
			 " group by nt.type");
1078 1079

	if ($db_result->numrows) {
1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
	    my $info_str = "";
	    while (my ($nt,$plc) = $db_result->fetchrow_array()) {
		$info_str .= " $plc($nt)";
	    }
	    print "Configuring Slices for $info_str.\n";
	    TBDebugTimeStamp("plabslice started");
	    if (system("plabslice create $pid $eid")) {
		tberror({type => 'secondary', severity => SEV_SECONDARY,
			 error => ['plabslice_create_failed']},
			"Failed to configure Slices.");
		return 3;
1091
	    }
1092
	    TBDebugTimeStamp("plabslice finished");
1093 1094 1095
	}
    }

1096
    # Check cancel flag before continuing. No retry, 
1097
    TBGetCancelFlag($pid, $eid, \$canceled);
1098 1099 1100 1101
    if ($canceled) {
	tbreport(SEV_IMMEDIATE, 'cancel_flag');
	return 1
    }
1102

1103 1104 1105 1106 1107 1108 1109 1110 1111
    #
    # These things need to get started before the nodes come up, so we'll
    # do them before the os_setup. Everything else can done in parallel with
    # os_setup. (Actually, these probably can too, since they should finish
    # long before the nodes reboot, but better safe than sorry)
    #
    print "Setting up mountpoints.\n";
    TBDebugTimeStamp("mountpoints started");
    if (system("exports_setup")) {
1112 1113 1114
	tberror({type => 'secondary', severity => SEV_SECONDARY,
		 error => ['mountpoints_setup_failed']},
		"Failed to setup mountpoints.");
1115 1116 1117
	return 1;
    }
    TBDebugTimeStamp("mountpoints finished");
Chad Barb's avatar
Chad Barb committed
1118

1119 1120 1121
    TBDebugTimeStamp("named started");
    print "Setting up named maps.\n";
    if (system("named_setup")) {
1122
	tbwarn "Failed to add node names to named map.";
1123 1124
	#
	# This is a non-fatal error.
Chad Barb's avatar
Chad Barb committed
1125
	#
1126 1127
    }
    TBDebugTimeStamp("named finished");
Chad Barb's avatar
Chad Barb committed
1128

Timothy Stack's avatar
Timothy Stack committed
1129 1130 1131 1132
    if ($NFSTRACESUPPORT) {
	print "Cleaning NFS traces.\n";
	TBDebugTimeStamp("nfstrace gc started");
	if (system("nfstrace gc $pid $eid")) {
1133 1134 1135
	    tberror({type => 'secondary', severity => SEV_SECONDARY,
		     error => ['nfstrace_setup_failed']},
		    "Failed to setup nfstrace.");
Timothy Stack's avatar
Timothy Stack committed
1136 1137 1138
	    return 1;
	}
	TBDebugTimeStamp("nfstrace gc finished");
1139 1140
    }

1141
    # Check cancel flag before continuing. No retry, 
1142
    TBGetCancelFlag($pid, $eid, \$canceled);
1143 1144 1145 1146
    if ($canceled) {
	tbreport(SEV_IMMEDIATE, 'cancel_flag');
	return 1
    }
Chad Barb's avatar
Chad Barb committed
1147

1148 1149 1150 1151
    #
    # Setup any control-net firewall.
    # This must be done before reloading and rebooting nodes.
    #
1152
    if ($firewalled && ($type == REAL || $type == MODIFY) &&
1153 1154
	doFWlans($experiment,
		 (($type == MODIFY) ? FWADDNODES : FWSETUP), undef)) {
1155 1156 1157
	return 1;
    }

1158 1159 1160 1161 1162
    #
    # PlabinElab setup.  This is currently just tweaking out the dhcpd.conf
    # file and that must be done before os_setup (i.e., before nodes are
    # rebooted).
    #
1163 1164
    if ($plabinelab && !$TESTMODE && $type > MODIFY_RECOVER) {
	# for MODIFY and RETRY we pass in the -u to pick up new nodes
1165 1166 1167 1168 1169
	my $optarg = ($type == REAL ? "" : "-u");
	
	print "Setting up plabinelab.\n";
	TBDebugTimeStamp("plabinelab setup started");
	if (system("plabinelab $optarg $pid $eid")) {
1170 1171 1172
	    tberror({type => 'secondary', severity => SEV_SECONDARY,
		     error => ['plabinelab_setup_failed']},
		    "Failed to setup plabinelab!");
1173 1174 1175 1176 1177
	    return 1;
	}
	TBDebugTimeStamp("plabinelab setup finished");
    }

Chad Barb's avatar
Chad Barb committed
1178 1179
    #
    # If user specified -reboot to update,
1180