tbswap.in 51.2 KB
Newer Older
1 2 3 4
#!/usr/bin/perl -w

#
# EMULAB-COPYRIGHT
5
# Copyright (c) 2000-2006 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
    print STDERR "Usage: $0 -force { in | out | update [-reboot] [-eventsys_restart] } pid eid\n";
28 29 30 31 32 33 34
    exit(-1);
}

#
# Configure variables
#
my $TBROOT         = "@prefix@";
35
my $TBOPS          = "@TBOPSEMAIL@";
Timothy Stack's avatar
Timothy Stack committed
36
my $TBLOGS         = "@TBLOGSEMAIL@";
37
my $MAINSITE	   = @TBMAINSITE@;
38
my $THISHOMEBASE   = "@THISHOMEBASE@";
39 40
my $TESTMODE       = @TESTMODE@;
my $DISABLE_EVENTS = "@DISABLE_EVENT_SCHED@";
41
my $piper          = "$TBROOT/sbin/locpiper";
Timothy Stack's avatar
Timothy Stack committed
42
my $NFSTRACESUPPORT= @NFSTRACESUPPORT@;
43 44 45 46 47 48 49 50 51 52 53

# 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;
54
use libadminctrl;
55
use libadminmfs;
56
use libtblog;
Leigh B. Stoller's avatar
Leigh B. Stoller committed
57
use libArchive;
58

59
#require exitonwarn; # exitonwarn isn't really a module, so just require it
60 61 62 63

#
# Actual swap-in and swap-out functions, defined below.
#
64 65 66
sub doSwapout($);
sub doSwapin($);

67 68 69 70
#
# Firewall stuff
# XXX maybe should be elsewhere
#
71 72 73 74 75
sub FWSETUP()     { return 1; }
sub FWADDNODES()  { return 2; }
sub FWDELNODES()  { return 3; }
sub FWTEARDOWN()  { return 4; }
sub doFW($$$$);
76 77

# XXX fixme: should not be hardwired!
78
my $cnetstack = "-S Control";
79 80 81
my $cnetvlanname = "Control";


82 83 84 85
sub REAL()    { return 4; }
sub CLEANUP() { return 3; }
sub RETRY()   { return 2; }
sub UPDATE()  { return 1; }
Chad Barb's avatar
Chad Barb committed
86
sub UPDATE_RECOVER()  { return 0; }
87

88 89 90
#
# Grab global enable of linkdelays.
#
91
my $enablelinkdelays = TBGetSiteVar("general/linux_endnodeshaping");
92

93 94 95 96 97 98
#
# Turn off line buffering on output
#

$| = 1;

99 100
my $updateReboot   = 0;
my $updateReconfig = 1;
101 102
my $update_Eventsys_restart = 0;
my $elabinelab     = 0;
103
my $plabinelab     = 0;
104 105
my $force  = 0;
my $errors = 0;
106
my $updatehosed = 0;
107
my $state;
108
my $canceled;
109 110
my $os_setup_pid;
my $cleanvlans;
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;	

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

#
# Get other arguments.
#

while ($#ARGV > 1) {
    $arg = shift;
    if ($arg eq "-force") {
	$force = 1;
Chad Barb's avatar
Chad Barb committed
135 136
    } elsif ($arg eq "-reboot") {
	$updateReboot = 1;
137
	$updateReconfig = 0;
138 139 140
    } elsif ($arg eq "-noreconfig") {
	$updateReboot   = 0;
	$updateReconfig = 0;
141 142
    } elsif ($arg eq "-eventsys_restart" && $swapop eq "update") {
	$update_Eventsys_restart = 1;
143 144 145 146 147 148 149 150 151
    } else {
	usage();
    }
}
if ($#ARGV < 1) {
    usage();
}
my ($pid,$eid) = @ARGV;

152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167
#
# 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");
}

168 169 170 171 172 173
#
# Set Error reporting info
# 
tblog_set_info($pid,$eid,$UID);

#
174
# Turn on timestamps
175
#
176 177
TBDebugTimeStampsOn();

178 179 180 181 182 183
# Need this below.
my ($dbuid, $uname, $umail);
if (UNIX2DBUID($UID, \$dbuid)) {
    UserDBInfo($dbuid, \$uname, \$umail);
}

184 185 186
#
# Print starting message.
#
187 188
my $exptidx;
TBExptIDX($pid, $eid, \$exptidx);
189

190 191
print "Beginning swap-$swapop for $pid/$eid ($exptidx). " .
    TBTimeStampWithDate() . "\n";
192 193 194 195 196 197
TBDebugTimeStamp("tbswap $swapop started");

#
# Get experiment state; verify that experiment exists.
#
if (! ($state = ExpState($pid, $eid))) {
198
    tbdie "No such experiment $pid/$eid";
199
}
200 201 202
# Sanity check the current state. 
if (!$force) {
    if ($swapop eq "in") {
203
	tbdie("Experiment should be ACTIVATING. Currently $state.")
204
	    if ($state ne EXPTSTATE_ACTIVATING);
205
    }
206
    elsif ($swapop eq "out") {
207
	tbdie("Experiment should be SWAPPING. Currently $state.")
208
	    if ($state ne EXPTSTATE_SWAPPING);
209
    }
210
    elsif ($swapop eq "update") {
211
	tbdie("Experiment should be MODIFY_RESWAP. Currently $state.")
212
	    if ($state ne EXPTSTATE_MODIFY_RESWAP);
213 214
    }
}
215 216
# Get elabinelab status. See below.
if (! TBExptIsElabInElab($pid, $eid, \$elabinelab)) {
217
    tbdie("Could not get elabinelab status for experiment $pid/$eid");
218
}
219 220 221 222 223 224 225 226 227 228 229
# 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);
}
230

231 232 233 234 235
#
# See if the experiment is firewalled
#
my $firewalled = TBExptFirewall($pid, $eid);

236 237 238
#
# Do actual swapping
#
239 240 241 242
if ($swapop eq "out") {
    #
    # Swap out
    #
243
    tblog_set_attempt(1);
244
    $errors = doSwapout(REAL);
245
}
246
elsif ($swapop eq "update") {
247 248 249 250 251 252 253
    #
    # 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";
    if (! TBAdmissionControlCheck($UID, $pid, $eid, undef)) {
254
	tberror "Admission control failure!\n";
255 256 257 258 259 260
	print "Failingly finished swap-$swapop for $pid/$eid. " .
	    TBTimeStamp() . "\n";
	TBDebugTimeStamp("tbswap $swapop finished (failed)");
	exit(1);
    }
    
261 262
    #
    # Update.
263 264 265
    #
    tblog_set_attempt(1);

266 267 268 269 270
    #
    # Phase One -- swap experiment partially out.
    #
    print STDERR "Backing up physical state...\n";
    TBExptBackupPhysicalState($pid,$eid);
271

272
    $errors = doSwapout(UPDATE);
Chad Barb's avatar
Chad Barb committed
273

Leigh B. Stoller's avatar
Leigh B. Stoller committed
274 275 276 277 278 279
    print STDERR "Doing a swapmodswapout on the experiment archive ...\n";    
    if (libArchive::TBExperimentArchiveSwapModSwapOut($pid, $eid) < 0) {
	tberror("Failed to do a swapmodswapout on the experiment archive!");
	$errors = 1;
    }

280
    if ($errors) {
281
	#
282 283 284 285 286
	# Clean up the mess, leaving the experiment in the SWAPPED state,
	# 
	print STDERR "Cleaning up after errors.\n";
	doSwapout(CLEANUP);
	$updatehosed = 1;
287
    }
288
    else {
289
	#
290
	# Phase Two -- swap experiment back in.
291
	#
Chad Barb's avatar
Chad Barb committed
292 293
	$errors = doSwapin(UPDATE);

294
	if ($errors) {
Chad Barb's avatar
Chad Barb committed
295 296 297
	    #
	    # There were errors; see if we can recover.
	    #
298
	    my $CanRecover = 1;
Chad Barb's avatar
Chad Barb committed
299 300 301

	    if ($errors != 7) {
		print STDERR "Update failure occurred _after_ assign phase; ";
302
		$CanRecover = 0;
Chad Barb's avatar
Chad Barb committed
303 304
	    }

305 306 307 308 309 310 311 312
	    if ($CanRecover) {
		print STDERR "Recovering virtual and physical state.\n";

		if (TBExptRemoveVirtualState($pid, $eid) ||
		    TBExptRestoreVirtualState($pid, $eid) ||
		    TBExptRestorePhysicalState($pid,$eid)) {
		    print STDERR "Could not restore backed-up state; ";
		    $CanRecover = 0;
Chad Barb's avatar
Chad Barb committed
313
		}
314 315 316 317 318 319 320
		else {
		    print STDERR "Doing a recovery swap-in of old state.\n";

		    if (doSwapin(UPDATE_RECOVER)) {
			print STDERR "Could not swap in old physical state; ";
			$CanRecover = 0;
		    }
Chad Barb's avatar
Chad Barb committed
321 322
		}
	    }
323 324 325 326 327 328 329

	    #
	    # 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.
	    # 
	    if (! $CanRecover) {
330
		print STDERR "Recovery aborted! Swapping experiment out.\n";
Chad Barb's avatar
Chad Barb committed
331
		doSwapout(CLEANUP);
332 333 334 335
		$updatehosed = 1;
	    }
	    else {
		print STDERR "Update recovery successful.\n";
Chad Barb's avatar
Chad Barb committed
336
	    }
337
	}
Chad Barb's avatar
Chad Barb committed
338 339
    }
}
340 341 342 343 344
elsif ($swapop eq "in") {
    #
    # Swap in
    #
    my $retries = 2;
345 346 347 348 349 350 351

    #
    # Before real swapin, do cursory admission control. assign_wrapper does
    # a more stringent check using assign.
    #
    print STDERR "Checking with Admission Control ...\n";
    if (! TBAdmissionControlCheck($UID, $pid, $eid, undef)) {
352
	tberror "Admission control failure!\n";
353 354 355 356 357
	print "Failingly finished swap-$swapop for $pid/$eid. " .
	    TBTimeStamp() . "\n";
	TBDebugTimeStamp("tbswap $swapop finished (failed)");
	exit(1);
    }
358 359

    tblog_set_attempt(1);
360
    $errors = doSwapin(REAL);
Chad Barb's avatar
Chad Barb committed
361

362 363 364 365 366 367 368 369 370 371
    #
    # 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--;
372
	tblog_inc_attempt();
373

374 375
	print STDERR "Cleaning up after errors; will try again.\n";
	doSwapout(RETRY);
376

377 378
	print STDERR "Trying again...\n";
	$errors = doSwapin(RETRY);
379
    }
380 381
    if ($errors || $canceled) {
	print STDERR "Cleaning up after " .
382
	    ($canceled ? "cancelation" : "errors") . ".\n";
383
	doSwapout(CLEANUP);
384 385 386
    }
}

387 388 389 390 391 392 393 394
tblog_set_attempt(0);

TBGetCancelFlag($pid, $eid, \$canceled);
if ($canceled) {
    tberror ({type=>'summary', cause=>'canceled'},
	     "Experiment swap-$swapop canceled by user.");
}

395 396 397 398
#
# Write appropriate message and exit.
#
if ($errors) {
399
    print "Failingly finished swap-$swapop for $pid/$eid. ".TBTimeStamp()."\n";
400
    TBDebugTimeStamp("tbswap $swapop finished (failed)");
Chad Barb's avatar
Chad Barb committed
401

402 403
    # Pass out magic value to indicate that update failed!
    exit(1 | ($updatehosed ? 0x40 : 0));
404
}
405
print "Successfully finished swap-$swapop for $pid/$eid. " .TBTimeStamp()."\n";
406 407
TBDebugTimeStamp("tbswap $swapop finished (succeeded)");
exit(0);
408 409 410 411

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

##
Chad Barb's avatar
Chad Barb committed
412
#
413 414
# doSwapout - Swaps experiment out.
#
Chad Barb's avatar
Chad Barb committed
415
#             If in REAL or CLEANUP,
416 417 418
#             this function will free all nodes for the 
#             experiment.
#
Chad Barb's avatar
Chad Barb committed
419
#             If in RETRY or UDPATE,
420 421 422 423 424 425
#             only nodes not in RES_READY will be freed.
#
#             Returns 0 on success, >0 on failure.
#
##

426 427
sub doSwapout($) {
    my $type = shift; # REAL==4, CLEANUP==3, RETRY==2, UPDATE==1.
428 429
    my $swapout_errors = 0;

430 431
    tblog_set_cleanup(1) if $type == CLEANUP;

432 433 434 435 436 437 438 439 440
    #
    # 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
441

442
    if (0 && $NFSTRACESUPPORT && $type == REAL) {
Timothy Stack's avatar
Timothy Stack committed
443 444
	print "Getting files accessed via NFS.\n";
	TBDebugTimeStamp("nfstrace started");
445
	system("nfstrace transfer $pid $eid");
Timothy Stack's avatar
Timothy Stack committed
446
	TBDebugTimeStamp("nfstrace finished");
447
    }
448

449
    if (! $TESTMODE) { 
450 451 452
	if (! ($DISABLE_EVENTS || $elabinelab)) {
	    if ($type >= RETRY ||
		($update_Eventsys_restart && $type == UPDATE) ) {
453 454
		print "Stopping the event system\n";
		if (system("eventsys_control stop $pid $eid")) {
455
		    tberror "Failed to stop the event system.";
456 457
		    $swapout_errors = 1;
		}
458 459 460 461 462 463 464

		#
		# Stop the location piper.
		#
		if (-x $piper) {
		    print "Stopping the location piper\n";
		    if (system("$piper -k $pid $eid")) {
465
			tberror "Failed to stop location piper.";
466 467 468
			$swapout_errors = 1;
		    }
		}
469 470
	    }
	}
471 472 473 474 475 476
	
	#
	# 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).
	#
477
	if ($elabinelab && $type >= CLEANUP) {
478 479
	    print "Tearing down elabinelab. This could take a while.\n";
	    if (system("elabinelab -k $pid $eid")) {
480
		tberror "Failed to teardown elabinelab!";
481 482 483 484
		$swapout_errors = 1;
	    }
	}

485
	#
Chad Barb's avatar
Chad Barb committed
486
	# Clean up any VLANs in experiment.
487
	#
488 489 490 491 492 493 494
	# When modifying an elabinelab experiment, leave the vlans intact
	# so that the inner networks are not suddenly disconnected!
	#
	if (! ($elabinelab && $type == UPDATE)) {
	    TBDebugTimeStamp("snmpit started");
	    print STDERR "Removing VLANs.\n";
	    if (system("snmpit -r $pid $eid")) {
495
		tberror "Failed to reset VLANs";
496 497 498 499 500
		$swapout_errors = 1;
	    } else {
		$cleanvlans = 0;
	    }
	    TBDebugTimeStamp("snmpit finished");
501
	}
502
    }
Chad Barb's avatar
Chad Barb committed
503

504
    if ($type >= CLEANUP) {
505 506 507 508 509 510 511 512 513 514 515 516 517
	#
	# 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")) {
		tberror "Failed to teardown plabinelab!";
		$swapout_errors = 1;
	    }
	}

518 519
	#
	# We're not attempting a retry;
520
	#
521
	# Stop all of the vnodes.
522
	#
523
	if (! $TESTMODE) { 	
524 525 526
	    print "Tearing down virtual nodes.\n";
	    TBDebugTimeStamp("vnode_setup -k started");
	    if (system("vnode_setup -d -k $pid $eid")) {
527
		tberror "Failed to tear down vnodes.";
528 529 530
		$swapout_errors = 1;
	    }
	    TBDebugTimeStamp("vnode_setup finished");
531 532
	}

533 534 535 536 537
	#
	# Nodes behind a firewall are treated special.
	# See undoFWNodes for details.
	#
	if ($firewalled && undoFWNodes($pid, $eid)) {
538
	    tblog_set_cleanup(0);
539 540
	    return 1;
	}
Leigh B. Stoller's avatar
Leigh B. Stoller committed
541

542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561
	#
	# 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, $eid, \%soaction);
	}
562
	if ($soaction{'command'} && doSwapoutAction($pid, $eid, %soaction)) {
563
	    tblog_set_cleanup(0);
564
	    return 1;
565 566
	}

567 568
	#
	# remove all nodes from the experiment.
Chad Barb's avatar
Chad Barb committed
569
	# (nfree will send them to RES_FREE_DIRTY)
570 571 572 573
	#
	print STDERR "Freeing nodes.\n";
	TBDebugTimeStamp("nfree started");
	if (system("nfree $pid $eid")) {
574
	    tberror "Could not free nodes.";
575 576 577
	    $swapout_errors = 1;
	}
	TBDebugTimeStamp("nfree finished");
578 579 580 581 582

	#
	# Since this is an actual swapout, 
	# reset our count of swap out nag emails sent.
	#
583
	DBQueryWarn("update experiments set swap_requests='',sim_reswap_count='0' ".
584
		    "where eid='$eid' and pid='$pid'");
585 586
    } else {
	#
587
	# $type == RETRY or $type == UPDATE.
588 589
	# 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
590
	# (nfree will send deallocated nodes to RES_FREE_DIRTY)
591
	#
592 593 594 595 596 597 598 599
	my @failedpnodes = ();
	my @failedvnodes = ();
	
	my $db_result =
	    DBQueryFatal("select rv.node_id,n.allocstate,nt.isvirtnode ".
                         "  from reserved as rv ".
			 "left join nodes as n on n.node_id = rv.node_id ".
			 "left join node_types as nt on nt.type=n.type ".
600 601
			 "where rv.pid='$pid' and rv.eid='$eid'");

602
	while (my ($node,$allocstate,$isvirt) = $db_result->fetchrow_array) {
603
	    if ($allocstate ne TBDB_ALLOCSTATE_RES_READY()) {
604 605 606 607 608 609
		if ($isvirt) {
		    push(@failedvnodes, $node);
		}
		else {
		    push(@failedpnodes, $node);
		}
610 611 612
	    }
	}

613 614 615 616 617 618 619
	#
	# 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")) {
620
		tberror "Failed to tear down vnodes.";
621 622 623 624
		$swapout_errors = 1;
	    }
	    TBDebugTimeStamp("vnode_setup -k finished");
	}
625

626 627 628 629 630 631 632 633 634 635 636
	#
	# 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;
	    }
	}

637 638 639 640
	#
	# Release all failed nodes.
	# 
	if (@failedpnodes > 0 || @failedvnodes > 0) {
641 642
	    print STDERR "Freeing failed nodes.\n";
	    
643
	    TBDebugTimeStamp("nfree started");
Chad Barb's avatar
Chad Barb committed
644 645 646 647 648
	    #
	    # 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.
	    #
649 650
	    if (system("nfree -x $pid $eid " .
		       join(" ", (@failedpnodes, @failedvnodes)))) {
651
		tberror "Could not free nodes.";
652 653 654 655 656 657
		$swapout_errors = 1;
	    }
	    TBDebugTimeStamp("nfree finished");
	}
    }

658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681
    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) {
	    # Are there any dslice nodes left?
	    $db_result =
		DBQueryFatal("select n.node_id from nodes as n ".
			     "left join node_types as nt on n.type = nt.type ".
			     "left join reserved as r ".
			     " on r.node_id = n.node_id ".
			     "where r.pid='$pid' and r.eid='$eid' ".
			     " and nt.isplabdslice = 1");

	    if (!$db_result->numrows) {
		print "Destroying Planetlab slice.\n";
		TBDebugTimeStamp("plabslice destroy started");
682
		if (system("plabslice destroy $pid $eid")) {
683
		    tberror "Failed to destroy Plab dslice.";
684 685 686 687 688 689 690
		    $swapout_errors = 1;
		}
		TBDebugTimeStamp("plabslice destroy finished");
	    }
	}
    }

691 692 693 694 695 696 697 698
    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
699
	# not a failed swapin(cleanup), update, or retry.
700
	#
701
	if ($type == REAL) {
702 703 704
	    print "Resetting mountpoints.\n";
	    TBDebugTimeStamp("exports started");
	    if (system("exports_setup")) {
705
		tberror "Failed to reset mountpoints.";
706 707 708
	    }
	    TBDebugTimeStamp("exports finished");
	}
Chad Barb's avatar
Chad Barb committed
709

710 711 712 713
	#
	# Resetting named maps and email lists is fast and idempotent,
	# so whatever.
	#
714 715 716
	print "Resetting named maps.\n";
	TBDebugTimeStamp("named started");
	if (system("named_setup")) {
717
	    tbwarn "Failed to reset named map.";
718 719
	}
	TBDebugTimeStamp("named finished");
Chad Barb's avatar
Chad Barb committed
720

721 722
	print "Resetting email lists.\n";
	TBDebugTimeStamp("genelists started");
723
	if (system("genelists -t")) {
724
	    tbwarn "Failed to reset email lists.";
725 726 727 728
	}
	TBDebugTimeStamp("genelists finished");
    }

729
    #
730 731 732
    # Wipe the DB clean except during UPDATE or RETRY. In those
    #    cases, assign_wrapper will reset the DB after reading
    #    the info.
733
    #
734 735 736 737
    if ( $type >= CLEANUP ) {
	print STDERR "Resetting DB.\n";
	TBExptRemovePhysicalState( $pid, $eid );
    }
738

739
    tblog_set_cleanup(0);
740 741 742 743
    return $swapout_errors;
}

##
Chad Barb's avatar
Chad Barb committed
744
#
745 746
# doSwapin - Swaps experiment in.
#
Chad Barb's avatar
Chad Barb committed
747 748 749 750 751
#            Returns:
#              0 - successful swapin
#              1 - failed swapin; cleanup required.
#              3 - failed swapin; cleanup required; can retry.
#              7 - failed swapin; assign failed; no cleanup.
752 753
##

754
sub doSwapin($) {
755 756 757
    my $type = shift; # REAL==4, RETRY==2, UPDATE==1, UPDATE_RECOVER=0.
    # Just the physnodes ...
    my @deleted_pnodes = ();
758

759 760 761 762
    #
    # assign_wrapper does all the virtual to physical mapping 
    # and updating the DB state.
    #
763
    
Chad Barb's avatar
Chad Barb committed
764
    if ($type > UPDATE_RECOVER) {
765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780
        #
        # Hacky test to allow disabling of linkdelays if the node is going
        # to run Linux. See sitevar above.
        #
	if (! $enablelinkdelays) {
	    $db_result =
		DBQueryFatal("select distinct e.pid,e.eid,vl.vnode,vn.osname ".
			 "  from experiments as e ".
			 "left join virt_lans as vl on vl.pid=e.pid and ".
			 "     vl.eid=e.eid ".
			 "left join virt_nodes as vn on vn.pid=e.pid and ".
			 "     vn.eid=e.eid and vn.vname=vl.vnode ".
			 "left join os_info as o on o.osname=vn.osname and ".
			 "  (o.pid=vl.pid or o.pid='" . TBOPSPID() . "') ".
			 "where (vl.uselinkdelay!=0 or e.uselinkdelays!=0 or ".
			 "       e.forcelinkdelays!=0) and ".
781 782
			 "     (o.os is NULL or o.os='Linux' or ".
			 "      o.os='Fedora') and ".
783 784 785
			 "     e.pid='$pid' and e.eid='$eid'");

	    if ($db_result->numrows) {
786 787
		tberror "Endnodeshaping is disabled on Linux Images!";
		tberror "You must modify your experiment to swap it in.";
788 789 790 791
		return 1;
	    }
	}
	
Chad Barb's avatar
Chad Barb committed
792 793 794 795
	print "Mapping to physical reality ...\n";
	TBDebugTimeStamp("assign_wrapper started");

	#
796 797 798 799
	# 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
800 801
	#
	my $exitcode;
802 803 804
	my $wrapper = "assign_wrapper -u";
	$wrapper .= " -f"
	    if ($type == RETRY);
805
	
806
	if (system("$wrapper $pid $eid")) {
807 808
	    $exitcode = $? >> 8;

809
	    tberror "Failed ($exitcode) to map to reality.";
810

811 812
	    # Wrapper sets this bit when recovery is possible.
	    if ($exitcode & 64) {
813
		# We can recover. 
Chad Barb's avatar
Chad Barb committed
814
		return 7;
815 816
	    }
	    else {
817
		# No recovery, no retry.
Chad Barb's avatar
Chad Barb committed
818 819 820 821
		return 1;
	    }
	}
	TBDebugTimeStamp("assign_wrapper finished");
822

Chad Barb's avatar
Chad Barb committed
823 824
	print "Mapped to physical reality!\n";
    }
825

826
    # Check cancel flag before continuing. No retry, 
827
    TBGetCancelFlag($pid, $eid, \$canceled);
828 829 830
    return 1
	if ($canceled);

831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865
    #
    # 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.
    #
    if ($type == UPDATE) {
	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);
		}
	    }
866 867 868
	    # See below.
	    @deleted_pnodes = @physnodes;
	    
869 870 871
	    if (@virtnodes) {
		TBDebugTimeStamp("vnode_setup started");
		
872
		if (system("vnode_setup -k $pid $eid @virtnodes")) {
873
		    tberror "Failed to tear down unused virtnodes!\n";
874 875 876 877 878
		    return 1;
		}
		TBDebugTimeStamp("vnode_setup finished");
		
		if (system("nfree $pid $eid @virtnodes")) {
879
		    tberror "Failed to nfree unused virtnodes!\n";
880 881 882 883
		    return 1;
		}
	    }
	    if (@physnodes) {
884 885 886
		if ($elabinelab) {
		    print "Removing nodes from inner elab.\n";
		    if (system("elabinelab -r $pid $eid @physnodes")) {
887
			tberror "Failed to remove inner nodes!";
888 889
			return 1;
		    }
890 891 892 893 894 895
		} elsif ($plabinelab) {
		    print "Removing nodes from inner plab.\n";
		    if (system("plabinelab -r $pid $eid @physnodes")) {
			tberror "Failed to remove inner nodes!";
			return 1;
		    }
896 897 898
		}

		#
899 900
		# If the experiment is firewalled, cleanup the nodes
		# we are releasing.
901
		# 
902
		if ($firewalled && undoFWNodes($pid, $eid, @deleted_pnodes)) {
903 904 905
		    return 1;
		}
		
906
		if (system("nfree $pid $eid @physnodes")) {
907
		    tberror "Failed to nfree unused physnodes!\n";
908 909 910 911 912 913
		    return 1;
		}
	    }
	}
    }

914 915 916 917 918 919
    # Exit here if we are testing.
    if ($TESTMODE) {
	print "Testing run - Stopping here.\n";
	return 0;
    }

920 921 922 923 924 925 926 927 928 929 930
    #
    # 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?
	#
931
	tberror "Failed to set up tarballs.";
932 933 934 935
	return 1;
    }
    TBDebugTimeStamp("tarfiles_setup finished");

936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963
    #
    # If there are any Plab dslice nodes in the experiment, create the
    # dslice now
    #
    if ($type > UPDATE_RECOVER) {
	# Are there any Plab nodes?
	$db_result =
	    DBQueryFatal("select n.node_id from nodes as n ".
			 "left join node_types as nt on n.type = nt.type ".
			 "left join reserved as r on r.node_id = n.node_id ".
			 "where r.pid='$pid' and r.eid='$eid' ".
			 " and nt.isplabdslice = 1");

	if ($db_result->numrows) {
	    # Does slice already exist?
	    $db_result =
		DBQueryFatal("select slicename from plab_slices ".
			     "where pid='$pid' and eid='$eid'");

	    if (! $db_result->numrows) {
		my @plabnodes = ();
		
		while (my ($node) = $db_result->fetchrow_array()) {
		    push(@plabnodes, $node);
		}
		
		print "Creating Planetlab slice.\n";
		TBDebugTimeStamp("plabslice create started");
964
		if (system("plabslice create $pid $eid")) {
965
		    tberror "Failed to create Plab dslice";
966 967 968 969 970 971 972
		    return 3;
		}
		TBDebugTimeStamp("plabslice alloc finished");
	    }
	}
    }

973
    # Check cancel flag before continuing. No retry, 
974
    TBGetCancelFlag($pid, $eid, \$canceled);
975 976 977
    return 1
	if ($canceled);

978 979 980 981 982 983 984 985 986
    #
    # 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")) {
987
	tberror "Failed to setup mountpoints.";
988 989 990
	return 1;
    }
    TBDebugTimeStamp("mountpoints finished");
Chad Barb's avatar
Chad Barb committed
991

992 993 994
    TBDebugTimeStamp("named started");
    print "Setting up named maps.\n";
    if (system("named_setup")) {
995
	tbwarn "Failed to add node names to named map.";
996 997
	#
	# This is a non-fatal error.
Chad Barb's avatar
Chad Barb committed
998
	#
999 1000
    }
    TBDebugTimeStamp("named finished");
Chad Barb's avatar
Chad Barb committed
1001

Timothy Stack's avatar
Timothy Stack committed
1002 1003 1004 1005 1006 1007 1008 1009
    if ($NFSTRACESUPPORT) {
	print "Cleaning NFS traces.\n";
	TBDebugTimeStamp("nfstrace gc started");
	if (system("nfstrace gc $pid $eid")) {
	    tberror "Failed to setup nfstrace.";
	    return 1;
	}
	TBDebugTimeStamp("nfstrace gc finished");
1010 1011
    }

1012
    # Check cancel flag before continuing. No retry, 
1013
    TBGetCancelFlag($pid, $eid, \$canceled);
1014 1015
    return 1
	if ($canceled);
Chad Barb's avatar
Chad Barb committed
1016

1017 1018 1019 1020
    #
    # Setup any control-net firewall.
    # This must be done before reloading and rebooting nodes.
    #
1021 1022
    if ($firewalled && ($type == REAL || $type == UPDATE) &&
	doFW($pid, $eid, (($type == UPDATE) ? FWADDNODES : FWSETUP), undef)) {
1023 1024 1025
	return 1;
    }

1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043
    #
    # 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).
    #
    if ($plabinelab && !$TESTMODE && $type > UPDATE_RECOVER) {
	# for UPDATE and RETRY we pass in the -u to pick up new nodes
	my $optarg = ($type == REAL ? "" : "-u");
	
	print "Setting up plabinelab.\n";
	TBDebugTimeStamp("plabinelab setup started");
	if (system("plabinelab $optarg $pid $eid")) {
	    tberror "Failed to setup plabinelab!";
	    return 1;
	}
	TBDebugTimeStamp("plabinelab setup finished");
    }

Chad Barb's avatar
Chad Barb committed
1044 1045
    #
    # If user specified -reboot to update,
1046
    # and we are successfully performing the update,
1047
    # then mark all nodes in experiment so os_setup will reboot them.
1048 1049
    # We must reboot nodes on a RETRY as well, since assign has been rerun
    # and may have remapped interfaces on the nodes.
Chad Barb's avatar
Chad Barb committed
1050
    #
1051 1052
    if ($type == RETRY ||
	($type == UPDATE && ($updateReboot || $updateReconfig))) {
Mike Hibler's avatar
Mike Hibler committed
1053
	my $needreboot = ($type == RETRY || $updateReboot) ? 1 : 0;
1054

Mike Hibler's avatar
Mike Hibler committed
1055 1056
	print STDERR "Marking nodes for ",
		     $needreboot ? "reboot" : "reconfig", ".\n";
Chad Barb's avatar
Chad Barb committed
1057
	$db_result =
1058 1059 1060
	    DBQueryFatal("select r.node_id,n.allocstate from reserved as r ".
			 "left join nodes as n on n.node_id=r.node_id ".
			 "where r.pid='$pid' and r.eid='$eid'");
Chad Barb's avatar
Chad Barb committed
1061

1062 1063 1064 1065 1066 1067 1068 1069
	while (my ($node,$allocstate) = $db_result->fetchrow_array) {
	    #
	    # If the node is INIT_CLEAN, leave it alone. It will still get
	    # rebooted, but will not falsely be tagged as dirty. This is
	    # important for vnodes too, where INIT_CLEAN indicated the vnode
	    # does not even exist yet (plab nodes).
	    #
	    if ($allocstate ne TBDB_ALLOCSTATE_RES_INIT_CLEAN()) {
1070
		TBSetNodeAllocState($node,
1071
				    ($needreboot ?
1072 1073
				     TBDB_ALLOCSTATE_RES_INIT_DIRTY() :
				     TBDB_ALLOCSTATE_RES_RECONFIG()));
1074
	    }
Chad Barb's avatar
Chad Barb committed
1075 1076 1077
	}
    }

1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090
    #
    # Lets run gentopofile again, so we get ltmap right. This will come out
    # later, most likely.
    #
    TBDebugTimeStamp("gentopofile started");
    print "Generating ltmap (again) ...\n";

    if (system("gentopofile $pid $eid")) {
        tberror("gentopofile failed!");
	return 1;
    }
    TBDebugTimeStamp("gentopofile finished");

1091
    # XXX fer now hack
1092
    if (0 && !$firewalled && !$elabinelab && !$plabinelab &&
1093
	($pid eq "testbed" || $pid eq "tbres")) {
1094 1095 1096 1097 1098
	DBQueryWarn("update experiments set ".
		    "    savedisk=1 where pid='$pid' and eid='$eid'");
    }


1099 1100 1101 1102 1103
    #
    # Since it'll take a while for the nodes to reboot, we'll start now, and
    # wait for the os_setup to finish, down below
    #
    print "Resetting OS and rebooting.\n";
1104
    TBDebugTimeStamp("launching os_setup");
1105 1106 1107
    if (!($os_setup_pid = fork())) { 
	exec("os_setup $pid $eid") or return 1;
    } elsif ($os_setup_pid == -1) {
1108
	tberror "Fork failed.";
1109 1110
	return 1;
    }
Chad Barb's avatar
Chad Barb committed
1111

1112
    #
1113 1114 1115 1116 1117
    # XXX
    # Don't add any steps between here and the waitpid() call below
    # without verifying that 1) It's OK for nodes to come up before
    # the step has completed and 2) It's OK for the command to run in
    # parallel with os_setup (no DB dependencies, etc.)
1118 1119 1120 1121 1122
    #

    print "Setting up VLANs.\n";
    TBDebugTimeStamp("snmpit started");
    if (system("snmpit -t $pid $eid")) {
1123
	tberror {type=>'summary'}, "Failed to set up VLANs.";
1124 1125 1126
	return 1;
    }
    TBDebugTimeStamp("snmpit finished");
Chad Barb's avatar
Chad Barb committed
1127

1128 1129 1130 1131
    #
    # An error now means that the VLANS need to be cleaned up.
    #
    $cleanvlans = 1;
Chad Barb's avatar
Chad Barb committed
1132

1133 1134
    print "Setting up email lists.\n";
    TBDebugTimeStamp("genelists started");
1135
    if (system("genelists -t")) {
1136
	tbwarn "Failed to update email lists.";
1137 1138 1139 1140 1141
	#
	# This is a non-fatal error.
	# 
    }
    TBDebugTimeStamp("genelists finished");
Chad Barb's avatar
Chad Barb committed
1142

1143 1144 1145 1146 1147 1148 1149 1150
    #
    # Don't clear port counters on UPDATE.
    # (XXX should clear new nodes' port counters.)

    if ($type >= RETRY) {
	print "Clearing port counters.\n";
	TBDebugTimeStamp("portstats started");
	if (system("portstats -z -a -q $pid $eid")) {
1151
	    tbwarn "Failed to clear port counters.";
1152 1153 1154 1155 1156
	    #
	    # This is a non-fatal error.
	    # 
	}
	TBDebugTimeStamp("portstats finished");
1157
    }
Chad Barb's avatar
Chad Barb committed
1158

1159 1160 1161 1162 1163 1164 1165
    #
    # OK, let's see how that os_setup did
    #
    $kid = waitpid($os_setup_pid,0);
    if ($kid == $os_setup_pid) {
	undef $os_setup_pid; # Make sure doswapout() doesn't wait for it.
	if ($CHILD_ERROR) {
1166
	    tberror "Failed to reset OS and reboot nodes.";
1167 1168 1169 1170 1171 1172 1173 1174 1175
	    #
	    # If there is a firewall involved, it could be that the
	    # firewall rules are preventing essential communication,
	    # so don't retry.
	    #
	    # XXX should only do this if the user has specified additional
	    # rules.  But right now, I may screw up too!
	    #
	    if ($firewalled) {
1176
		tberror "Not retrying, ".
1177
		    "possibly an error in firewall setup or configuration.";
1178 1179 1180
		return 1;
	    }

1181
	    #
1182 1183
	    # Use returncode from os_setup process to
	    # set global $retry flag, indicating to caller
1184 1185 1186
	    # that it may be beneficial to attempt
	    # a doSwapin() again.
	    #
1187
            if (($CHILD_ERROR >> 8) == 1) {
Chad Barb's avatar
Chad Barb committed
1188
		return 3;
1189 1190
	    } else {
		print STDERR "Not retrying due to error type.\n";
Chad Barb's avatar
Chad Barb committed
1191
		return 1;
1192
	    }
1193 1194 1195
	}
    } else {
	undef $os_setup_pid;
1196
	tberror "Error waiting for os_setup to finish.";
1197 1198
	return 1;
    }
Chad Barb's avatar
Chad Barb committed
1199

1200 1201 1202 1203
    #
    # Okay, start the event system now that we know all the nodes have
    # rebooted (os_setup is done). This only takes a moment (puts itself
    # in the background), so its not enough of a delay to worry about.
1204 1205
    # Don't do this during an update, since we didn't kill the 
    # event system previously, so starting it again will fail!
1206
    # 
1207
    if (! ($DISABLE_EVENTS || $elabinelab)) {
1208 1209 1210 1211 1212 1213 1214
	#
	# For the robot testbed, start the location piper *before* the event
	# system.
	#
	if (-x $piper && ($type != UPDATE && $type != UPDATE_RECOVER)) {
	    print "Starting the location piper.\n";
	    if (system("$piper $pid $eid")) {
1215
		tberror "Failed to start the location piper.";
1216 1217 1218 1219
		return 1;
	    }
	}
	
1220 1221
	if ( $update_Eventsys_restart || 
	    ($type != UPDATE && $type != UPDATE_RECOVER) ) {
1222 1223 1224
	    print "Starting the event system.\n";
	    TBDebugTimeStamp("eventsys_control started");
	    if (system("eventsys_control start $pid $eid")) {
1225
		tberror "Failed to start the event system.";