2 # -*- mode: perl; perl-indent-level: 2; indent-tabs-mode: nil; -*-
6 crunch-job: Execute job steps, save snapshots as requested, collate output.
10 Obtain job details from Arvados, run tasks on compute nodes (typically
11 invoked by scheduler on controller):
13 crunch-job --job x-y-z --git-dir /path/to/repo/.git
15 Obtain job details from command line, run tasks on local machine
16 (typically invoked by application or developer on VM):
18 crunch-job --job '{"script_version":"/path/to/working/tree","script":"scriptname",...}'
20 crunch-job --job '{"repository":"https://github.com/curoverse/arvados.git","script_version":"master","script":"scriptname",...}'
28 If the job is already locked, steal the lock and run it anyway.
32 Path to a .git directory (or a git URL) where the commit given in the
33 job's C<script_version> attribute is to be found. If this is I<not>
34 given, the job's C<repository> attribute will be used.
38 Arvados API authorization token to use during the course of the job.
42 Do not clear per-job/task temporary directories during initial job
43 setup. This can speed up development and debugging when running jobs
48 UUID of the job to run, or a JSON-encoded job resource without a
49 UUID. If the latter is given, a new job object will be created.
53 =head1 RUNNING JOBS LOCALLY
55 crunch-job's log messages appear on stderr along with the job tasks'
56 stderr streams. The log is saved in Keep at each checkpoint and when
59 If the job succeeds, the job's output locator is printed on stdout.
61 While the job is running, the following signals are accepted:
65 =item control-C, SIGINT, SIGQUIT
67 Save a checkpoint, terminate any job tasks that are running, and stop.
71 Save a checkpoint and continue.
75 Refresh node allocation (i.e., check whether any nodes have been added
76 or unallocated) and attributes of the Job record that should affect
77 behavior (e.g., cancel job if cancelled_at becomes non-nil).
85 use POSIX ':sys_wait_h';
86 use POSIX qw(strftime);
87 use Fcntl qw(F_GETFL F_SETFL O_NONBLOCK);
91 use Digest::MD5 qw(md5_hex);
97 use File::Path qw( make_path remove_tree );
99 use constant TASK_TEMPFAIL => 111;
100 use constant EX_TEMPFAIL => 75;
101 use constant EX_RETRY_UNLOCKED => 93;
103 $ENV{"TMPDIR"} ||= "/tmp";
104 unless (defined $ENV{"CRUNCH_TMP"}) {
105 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
106 if ($ENV{"USER"} ne "crunch" && $< != 0) {
107 # use a tmp dir unique for my uid
108 $ENV{"CRUNCH_TMP"} .= "-$<";
112 # Create the tmp directory if it does not exist
113 if ( ! -d $ENV{"CRUNCH_TMP"} ) {
114 make_path $ENV{"CRUNCH_TMP"} or die "Failed to create temporary working directory: " . $ENV{"CRUNCH_TMP"};
117 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
118 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
119 $ENV{"CRUNCH_WORK"} = $ENV{"JOB_WORK"}; # deprecated
120 mkdir ($ENV{"JOB_WORK"});
129 my $cgroup_root = "/sys/fs/cgroup";
130 my $docker_bin = "docker.io";
131 my $docker_run_args = "";
132 GetOptions('force-unlock' => \$force_unlock,
133 'git-dir=s' => \$git_dir,
134 'job=s' => \$jobspec,
135 'job-api-token=s' => \$job_api_token,
136 'no-clear-tmp' => \$no_clear_tmp,
137 'resume-stash=s' => \$resume_stash,
138 'cgroup-root=s' => \$cgroup_root,
139 'docker-bin=s' => \$docker_bin,
140 'docker-run-args=s' => \$docker_run_args,
143 if (defined $job_api_token) {
144 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
147 my $have_slurm = exists $ENV{SLURM_JOB_ID} && exists $ENV{SLURM_NODELIST};
152 $main::ENV{CRUNCH_DEBUG} = 1;
156 $main::ENV{CRUNCH_DEBUG} = 0;
159 my $arv = Arvados->new('apiVersion' => 'v1');
168 if ($jobspec =~ /^[-a-z\d]+$/)
170 # $jobspec is an Arvados UUID, not a JSON job specification
171 $Job = api_call("jobs/get", uuid => $jobspec);
176 $local_job = JSON::decode_json($jobspec);
180 # Make sure our workers (our slurm nodes, localhost, or whatever) are
181 # at least able to run basic commands: they aren't down or severely
184 if (($Job || $local_job)->{docker_image_locator}) {
185 $cmd = [$docker_bin, 'ps', '-q'];
187 Log(undef, "Sanity check is `@$cmd`");
188 my ($exited, $stdout, $stderr) = srun_sync(
189 ["srun", "--nodes=\Q$ENV{SLURM_NNODES}\E", "--ntasks-per-node=1"],
191 {label => "sanity check"});
193 Log(undef, "Sanity check failed: ".exit_status_s($exited));
196 Log(undef, "Sanity check OK");
199 my $User = api_call("users/current");
202 if (!$force_unlock) {
203 # Claim this job, and make sure nobody else does
204 eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
206 Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
215 map { croak ("No $_ specified") unless $local_job->{$_} }
216 qw(script script_version script_parameters);
219 $local_job->{'is_locked_by_uuid'} = $User->{'uuid'};
220 $local_job->{'started_at'} = gmtime;
221 $local_job->{'state'} = 'Running';
223 $Job = api_call("jobs/create", job => $local_job);
225 $job_id = $Job->{'uuid'};
227 my $keep_logfile = $job_id . '.log.txt';
228 log_writer_start($keep_logfile);
230 $Job->{'runtime_constraints'} ||= {};
231 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
232 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
234 my $gem_versions = `gem list --quiet arvados-cli 2>/dev/null`;
236 $gem_versions =~ s/^arvados-cli \(/ with arvados-cli Gem version(s) /;
237 chomp($gem_versions);
238 chop($gem_versions); # Closing parentheses
243 "running from " . ((-e $0) ? realpath($0) : "stdin") . $gem_versions);
245 Log (undef, "check slurm allocation");
248 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
252 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
253 push @sinfo, "$localcpus localhost";
255 if (exists $ENV{SLURM_NODELIST})
257 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
261 my ($ncpus, $slurm_nodelist) = split;
262 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
265 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
268 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
271 foreach (split (",", $ranges))
284 push @nodelist, map {
286 $n =~ s/\[[-,\d]+\]/$_/;
293 push @nodelist, $nodelist;
296 foreach my $nodename (@nodelist)
298 Log (undef, "node $nodename - $ncpus slots");
299 my $node = { name => $nodename,
301 # The number of consecutive times a task has been dispatched
302 # to this node and failed.
304 # The number of consecutive times that SLURM has reported
305 # a node failure since the last successful task.
307 # Don't dispatch work to this node until this time
308 # (in seconds since the epoch) has passed.
310 foreach my $cpu (1..$ncpus)
312 push @slot, { node => $node,
316 push @node, @nodelist;
321 # Ensure that we get one jobstep running on each allocated node before
322 # we start overloading nodes with concurrent steps
324 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
327 $Job->update_attributes(
328 'tasks_summary' => { 'failed' => 0,
333 Log (undef, "start");
334 $SIG{'INT'} = sub { $main::please_freeze = 1; };
335 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
336 $SIG{'TERM'} = \&croak;
337 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
338 $SIG{'ALRM'} = sub { $main::please_info = 1; };
339 $SIG{'CONT'} = sub { $main::please_continue = 1; };
340 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
342 $main::please_freeze = 0;
343 $main::please_info = 0;
344 $main::please_continue = 0;
345 $main::please_refresh = 0;
346 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
348 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
349 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
350 $ENV{"JOB_UUID"} = $job_id;
353 my @jobstep_todo = ();
354 my @jobstep_done = ();
355 my @jobstep_tomerge = ();
356 my $jobstep_tomerge_level = 0;
357 my $squeue_checked = 0;
358 my $sinfo_checked = 0;
359 my $latest_refresh = scalar time;
363 if (defined $Job->{thawedfromkey})
365 thaw ($Job->{thawedfromkey});
369 my $first_task = api_call("job_tasks/create", job_task => {
370 'job_uuid' => $Job->{'uuid'},
375 push @jobstep, { 'level' => 0,
377 'arvados_task' => $first_task,
379 push @jobstep_todo, 0;
385 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
388 my $build_script = handle_readall(\*DATA);
389 my $nodelist = join(",", @node);
390 my $git_tar_count = 0;
392 if (!defined $no_clear_tmp) {
393 # Find FUSE mounts under $CRUNCH_TMP and unmount them. Then clean
394 # up work directories crunch_tmp/work, crunch_tmp/opt,
397 # TODO: When #5036 is done and widely deployed, we can limit mount's
398 # -t option to simply fuse.keep.
399 my ($exited, $stdout, $stderr) = srun_sync(
400 ["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
401 ['bash', '-ec', '-o', 'pipefail', 'mount -t fuse,fuse.keep | awk "(index(\$3, \"$CRUNCH_TMP\") == 1){print \$3}" | xargs -r -n 1 fusermount -u -z; sleep 1; rm -rf $JOB_WORK $CRUNCH_INSTALL $CRUNCH_TMP/task $CRUNCH_TMP/src* $CRUNCH_TMP/*.cid'],
402 {label => "clean work dirs"});
404 exit(EX_RETRY_UNLOCKED);
408 # If this job requires a Docker image, install that.
409 my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem, $dockeruserarg);
410 if ($docker_locator = $Job->{docker_image_locator}) {
411 Log (undef, "Install docker image $docker_locator");
412 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
415 croak("No Docker image hash found from locator $docker_locator");
417 Log (undef, "docker image hash is $docker_hash");
418 $docker_stream =~ s/^\.//;
419 my $docker_install_script = qq{
420 if $docker_bin images -q --no-trunc --all | grep -xF \Q$docker_hash\E >/dev/null; then
423 declare -a exit_codes=("\${PIPESTATUS[@]}")
424 if [ 0 != "\${exit_codes[0]}" ]; then
425 exit "\${exit_codes[0]}" # `docker images` failed
426 elif [ 1 != "\${exit_codes[1]}" ]; then
427 exit "\${exit_codes[1]}" # `grep` encountered an error
429 # Everything worked fine, but grep didn't find the image on this host.
430 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
434 my ($exited, $stdout, $stderr) = srun_sync(
435 ["srun", "--nodelist=" . join(',', @node)],
436 ["/bin/bash", "-o", "pipefail", "-ec", $docker_install_script],
437 {label => "load docker image"});
440 exit(EX_RETRY_UNLOCKED);
443 # Determine whether this version of Docker supports memory+swap limits.
444 ($exited, $stdout, $stderr) = srun_sync(
445 ["srun", "--nodes=1"],
446 [$docker_bin, 'run', '--help'],
447 {label => "check --memory-swap feature"});
448 $docker_limitmem = ($stdout =~ /--memory-swap/);
450 # Find a non-root Docker user to use.
451 # Tries the default user for the container, then 'crunch', then 'nobody',
452 # testing for whether the actual user id is non-zero. This defends against
453 # mistakes but not malice, but we intend to harden the security in the future
454 # so we don't want anyone getting used to their jobs running as root in their
456 my @tryusers = ("", "crunch", "nobody");
457 foreach my $try_user (@tryusers) {
460 if ($try_user eq "") {
461 $label = "check whether default user is UID 0";
464 $label = "check whether user '$try_user' is UID 0";
465 $try_user_arg = "--user=$try_user";
467 my ($exited, $stdout, $stderr) = srun_sync(
468 ["srun", "--nodes=1"],
470 "$docker_bin run $docker_run_args $try_user_arg $docker_hash id --user"],
473 if ($exited == 0 && $stdout =~ /^\d+$/ && $stdout > 0) {
474 $dockeruserarg = $try_user_arg;
475 if ($try_user eq "") {
476 Log(undef, "Container will run with default user");
478 Log(undef, "Container will run with $dockeruserarg");
484 if (!defined $dockeruserarg) {
485 croak("Could not find a user in container that is not UID 0 (tried default user, @tryusers) or there was a problem running 'id' in the container.");
488 if ($Job->{arvados_sdk_version}) {
489 # The job also specifies an Arvados SDK version. Add the SDKs to the
490 # tar file for the build script to install.
491 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
492 $Job->{arvados_sdk_version}));
493 add_git_archive("git", "--git-dir=$git_dir", "archive",
494 "--prefix=.arvados.sdk/",
495 $Job->{arvados_sdk_version}, "sdk");
499 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
500 # If script_version looks like an absolute path, *and* the --git-dir
501 # argument was not given -- which implies we were not invoked by
502 # crunch-dispatch -- we will use the given path as a working
503 # directory instead of resolving script_version to a git commit (or
504 # doing anything else with git).
505 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
506 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
509 # Resolve the given script_version to a git commit sha1. Also, if
510 # the repository is remote, clone it into our local filesystem: this
511 # ensures "git archive" will work, and is necessary to reliably
512 # resolve a symbolic script_version like "master^".
513 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
515 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
517 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
519 # If we're running under crunch-dispatch, it will have already
520 # pulled the appropriate source tree into its own repository, and
521 # given us that repo's path as $git_dir.
523 # If we're running a "local" job, we might have to fetch content
524 # from a remote repository.
526 # (Currently crunch-dispatch gives a local path with --git-dir, but
527 # we might as well accept URLs there too in case it changes its
529 my $repo = $git_dir || $Job->{'repository'};
531 # Repository can be remote or local. If remote, we'll need to fetch it
532 # to a local dir before doing `git log` et al.
535 if ($repo =~ m{://|^[^/]*:}) {
536 # $repo is a git url we can clone, like git:// or https:// or
537 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
538 # not recognized here because distinguishing that from a local
539 # path is too fragile. If you really need something strange here,
540 # use the ssh:// form.
541 $repo_location = 'remote';
542 } elsif ($repo =~ m{^\.*/}) {
543 # $repo is a local path to a git index. We'll also resolve ../foo
544 # to ../foo/.git if the latter is a directory. To help
545 # disambiguate local paths from named hosted repositories, this
546 # form must be given as ./ or ../ if it's a relative path.
547 if (-d "$repo/.git") {
548 $repo = "$repo/.git";
550 $repo_location = 'local';
552 # $repo is none of the above. It must be the name of a hosted
554 my $arv_repo_list = api_call("repositories/list",
555 'filters' => [['name','=',$repo]]);
556 my @repos_found = @{$arv_repo_list->{'items'}};
557 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
559 Log(undef, "Repository '$repo' -> "
560 . join(", ", map { $_->{'uuid'} } @repos_found));
563 croak("Error: Found $n_found repositories with name '$repo'.");
565 $repo = $repos_found[0]->{'fetch_url'};
566 $repo_location = 'remote';
568 Log(undef, "Using $repo_location repository '$repo'");
569 $ENV{"CRUNCH_SRC_URL"} = $repo;
571 # Resolve given script_version (we'll call that $treeish here) to a
572 # commit sha1 ($commit).
573 my $treeish = $Job->{'script_version'};
575 if ($repo_location eq 'remote') {
576 # We minimize excess object-fetching by re-using the same bare
577 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
578 # just keep adding remotes to it as needed.
579 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
580 my $gitcmd = "git --git-dir=\Q$local_repo\E";
582 # Set up our local repo for caching remote objects, making
584 if (!-d $local_repo) {
585 make_path($local_repo) or croak("Error: could not create $local_repo");
587 # This works (exits 0 and doesn't delete fetched objects) even
588 # if $local_repo is already initialized:
589 `$gitcmd init --bare`;
591 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
594 # If $treeish looks like a hash (or abbrev hash) we look it up in
595 # our local cache first, since that's cheaper. (We don't want to
596 # do that with tags/branches though -- those change over time, so
597 # they should always be resolved by the remote repo.)
598 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
599 # Hide stderr because it's normal for this to fail:
600 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
602 # Careful not to resolve a branch named abcdeff to commit 1234567:
603 $sha1 =~ /^$treeish/ &&
604 $sha1 =~ /^([0-9a-f]{40})$/s) {
606 Log(undef, "Commit $commit already present in $local_repo");
610 if (!defined $commit) {
611 # If $treeish isn't just a hash or abbrev hash, or isn't here
612 # yet, we need to fetch the remote to resolve it correctly.
614 # First, remove all local heads. This prevents a name that does
615 # not exist on the remote from resolving to (or colliding with)
616 # a previously fetched branch or tag (possibly from a different
618 remove_tree("$local_repo/refs/heads", {keep_root => 1});
620 Log(undef, "Fetching objects from $repo to $local_repo");
621 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
623 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
627 # Now that the data is all here, we will use our local repo for
628 # the rest of our git activities.
632 my $gitcmd = "git --git-dir=\Q$repo\E";
633 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
634 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
635 croak("`$gitcmd rev-list` exited "
637 .", '$treeish' not found, giving up");
640 Log(undef, "Version $treeish is commit $commit");
642 if ($commit ne $Job->{'script_version'}) {
643 # Record the real commit id in the database, frozentokey, logs,
644 # etc. -- instead of an abbreviation or a branch name which can
645 # become ambiguous or point to a different commit in the future.
646 if (!$Job->update_attributes('script_version' => $commit)) {
647 croak("Error: failed to update job's script_version attribute");
651 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
652 add_git_archive("$gitcmd archive ''\Q$commit\E");
655 my $git_archive = combined_git_archive();
656 if (!defined $git_archive) {
657 Log(undef, "Skip install phase (no git archive)");
659 Log(undef, "Warning: This probably means workers have no source tree!");
664 my $install_script_tries_left = 3;
665 for (my $attempts = 0; $attempts < 3; $attempts++) {
666 my @srunargs = ("srun",
667 "--nodelist=$nodelist",
668 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
669 my @execargs = ("sh", "-c",
670 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
672 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
673 my ($stdout, $stderr);
674 ($exited, $stdout, $stderr) = srun_sync(
675 \@srunargs, \@execargs,
676 {label => "run install script on all workers"},
677 $build_script . $git_archive);
679 my $stderr_anything_from_script = 0;
680 for my $line (split(/\n/, $stderr)) {
681 if ($line !~ /^(srun: error: |starting: \[)/) {
682 $stderr_anything_from_script = 1;
686 last if $exited == 0 || $main::please_freeze;
688 # If the install script fails but doesn't print an error message,
689 # the next thing anyone is likely to do is just run it again in
690 # case it was a transient problem like "slurm communication fails
691 # because the network isn't reliable enough". So we'll just do
692 # that ourselves (up to 3 attempts in total). OTOH, if there is an
693 # error message, the problem is more likely to have a real fix and
694 # we should fail the job so the fixing process can start, instead
695 # of doing 2 more attempts.
696 last if $stderr_anything_from_script;
699 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
700 unlink($tar_filename);
708 foreach (qw (script script_version script_parameters runtime_constraints))
712 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
714 foreach (split (/\n/, $Job->{knobs}))
716 Log (undef, "knob " . $_);
720 'filters' => [['hostname', 'in', \@node]],
721 'order' => 'hostname',
722 'limit' => scalar(@node),
724 for my $n (@{$resp->{items}}) {
725 Log(undef, "$n->{hostname} $n->{uuid} ".JSON::encode_json($n->{properties}));
730 $main::success = undef;
736 my $thisround_succeeded = 0;
737 my $thisround_failed = 0;
738 my $thisround_failed_multiple = 0;
739 my $working_slot_count = scalar(@slot);
741 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
742 or $a <=> $b } @jobstep_todo;
743 my $level = $jobstep[$jobstep_todo[0]]->{level};
745 my $initial_tasks_this_level = 0;
746 foreach my $id (@jobstep_todo) {
747 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
750 # If the number of tasks scheduled at this level #T is smaller than the number
751 # of slots available #S, only use the first #T slots, or the first slot on
752 # each node, whichever number is greater.
754 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
755 # based on these numbers. Using fewer slots makes more resources available
756 # to each individual task, which should normally be a better strategy when
757 # there are fewer of them running with less parallelism.
759 # Note that this calculation is not redone if the initial tasks at
760 # this level queue more tasks at the same level. This may harm
761 # overall task throughput for that level.
763 if ($initial_tasks_this_level < @node) {
764 @freeslot = (0..$#node);
765 } elsif ($initial_tasks_this_level < @slot) {
766 @freeslot = (0..$initial_tasks_this_level - 1);
768 @freeslot = (0..$#slot);
770 my $round_num_freeslots = scalar(@freeslot);
771 print STDERR "crunch-job have ${round_num_freeslots} free slots for ${initial_tasks_this_level} initial tasks at this level, ".scalar(@node)." nodes, and ".scalar(@slot)." slots\n";
773 my %round_max_slots = ();
774 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
775 my $this_slot = $slot[$freeslot[$ii]];
776 my $node_name = $this_slot->{node}->{name};
777 $round_max_slots{$node_name} ||= $this_slot->{cpu};
778 last if (scalar(keys(%round_max_slots)) >= @node);
781 Log(undef, "start level $level with $round_num_freeslots slots");
784 my $progress_is_dirty = 1;
785 my $progress_stats_updated = 0;
787 update_progress_stats();
791 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
793 # Don't create new tasks if we already know the job's final result.
794 last if defined($main::success);
796 my $id = $jobstep_todo[$todo_ptr];
797 my $Jobstep = $jobstep[$id];
798 if ($Jobstep->{level} != $level)
803 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
804 set_nonblocking($reader{$id});
806 my $childslot = $freeslot[0];
807 my $childnode = $slot[$childslot]->{node};
808 my $childslotname = join (".",
809 $slot[$childslot]->{node}->{name},
810 $slot[$childslot]->{cpu});
812 my $childpid = fork();
815 $SIG{'INT'} = 'DEFAULT';
816 $SIG{'QUIT'} = 'DEFAULT';
817 $SIG{'TERM'} = 'DEFAULT';
819 foreach (values (%reader))
823 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
824 open(STDOUT,">&writer");
825 open(STDERR,">&writer");
830 delete $ENV{"GNUPGHOME"};
831 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
832 $ENV{"TASK_QSEQUENCE"} = $id;
833 $ENV{"TASK_SEQUENCE"} = $level;
834 $ENV{"JOB_SCRIPT"} = $Job->{script};
835 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
836 $param =~ tr/a-z/A-Z/;
837 $ENV{"JOB_PARAMETER_$param"} = $value;
839 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
840 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
841 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
842 $ENV{"HOME"} = $ENV{"TASK_WORK"};
843 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
844 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
845 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
847 my $keep_mnt = $ENV{"TASK_WORK"}.".keep";
853 "--nodelist=".$childnode->{name},
854 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
855 "--job-name=$job_id.$id.$$",
858 my $stdbuf = " stdbuf --output=0 --error=0 ";
860 my $arv_file_cache = "";
861 if (defined($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'})) {
862 $arv_file_cache = "--file-cache=" . ($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'} * 1024 * 1024);
866 "if [ -e \Q$ENV{TASK_WORK}\E ]; then rm -rf \Q$ENV{TASK_WORK}\E; fi; "
867 ."mkdir -p \Q$ENV{CRUNCH_TMP}\E \Q$ENV{JOB_WORK}\E \Q$ENV{TASK_WORK}\E \Q$keep_mnt\E "
868 ."&& cd \Q$ENV{CRUNCH_TMP}\E "
869 # These environment variables get used explicitly later in
870 # $command. No tool is expected to read these values directly.
871 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
872 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
873 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
874 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP "
875 .q{&& declare -a VOLUMES=() }
876 .q{&& if which crunchrunner >/dev/null ; then VOLUMES+=("--volume=$(which crunchrunner):/usr/local/bin/crunchrunner:ro") ; fi }
877 .q{&& if test -f /etc/ssl/certs/ca-certificates.crt ; then VOLUMES+=("--volume=/etc/ssl/certs/ca-certificates.crt:/etc/arvados/ca-certificates.crt:ro") ; }
878 .q{elif test -f /etc/pki/tls/certs/ca-bundle.crt ; then VOLUMES+=("--volume=/etc/pki/tls/certs/ca-bundle.crt:/etc/arvados/ca-certificates.crt:ro") ; fi };
880 $command .= "&& exec arv-mount --read-write --mount-by-pdh=by_pdh --mount-tmp=tmp --crunchstat-interval=10 --allow-other $arv_file_cache \Q$keep_mnt\E --exec ";
881 $ENV{TASK_KEEPMOUNT} = "$keep_mnt/by_pdh";
882 $ENV{TASK_KEEPMOUNT_TMP} = "$keep_mnt/tmp";
886 my $containername = "$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}";
887 my $cidfile = "$ENV{CRUNCH_TMP}/$containername.cid";
888 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
889 $command .= "$docker_bin run $docker_run_args --name=$containername --attach=stdout --attach=stderr --attach=stdin -i \Q$dockeruserarg\E --cidfile=$cidfile --sig-proxy ";
890 # We only set memory limits if Docker lets us limit both memory and swap.
891 # Memory limits alone have been supported longer, but subprocesses tend
892 # to get SIGKILL if they exceed that without any swap limit set.
893 # See #5642 for additional background.
894 if ($docker_limitmem) {
895 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
898 # The source tree and $destdir directory (which we have
899 # installed on the worker host) are available in the container,
900 # under the same path.
901 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
902 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
904 # Currently, we make the "by_pdh" directory in arv-mount's mount
905 # point appear at /keep inside the container (instead of using
906 # the same path as the host like we do with CRUNCH_SRC and
907 # CRUNCH_INSTALL). However, crunch scripts and utilities must
908 # not rely on this. They must use $TASK_KEEPMOUNT.
909 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
910 $ENV{TASK_KEEPMOUNT} = "/keep";
912 # Ditto TASK_KEEPMOUNT_TMP, as /keep_tmp.
913 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT_TMP}:/keep_tmp\E ";
914 $ENV{TASK_KEEPMOUNT_TMP} = "/keep_tmp";
916 # TASK_WORK is almost exactly like a docker data volume: it
917 # starts out empty, is writable, and persists until no
918 # containers use it any more. We don't use --volumes-from to
919 # share it with other containers: it is only accessible to this
920 # task, and it goes away when this task stops.
922 # However, a docker data volume is writable only by root unless
923 # the mount point already happens to exist in the container with
924 # different permissions. Therefore, we [1] assume /tmp already
925 # exists in the image and is writable by the crunch user; [2]
926 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
927 # writable if they are created by docker while setting up the
928 # other --volumes); and [3] create $TASK_WORK inside the
929 # container using $build_script.
930 $command .= "--volume=/tmp ";
931 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
932 $ENV{"HOME"} = $ENV{"TASK_WORK"};
933 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
935 # TODO: Share a single JOB_WORK volume across all task
936 # containers on a given worker node, and delete it when the job
937 # ends (and, in case that doesn't work, when the next job
940 # For now, use the same approach as TASK_WORK above.
941 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
943 # Bind mount the crunchrunner binary and host TLS certificates file into
945 $command .= '"${VOLUMES[@]}" ';
947 while (my ($env_key, $env_val) = each %ENV)
949 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
950 $command .= "--env=\Q$env_key=$env_val\E ";
953 $command .= "--env=\QHOME=$ENV{HOME}\E ";
954 $command .= "\Q$docker_hash\E ";
956 if ($Job->{arvados_sdk_version}) {
958 $command .= "perl - \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E";
960 $command .= "/bin/sh -c \'python -c " .
961 '"from pkg_resources import get_distribution as get; print \"Using Arvados SDK version\", get(\"arvados-python-client\").version"' .
962 ">&2 2>/dev/null; " .
963 "mkdir -p \"$ENV{JOB_WORK}\" \"$ENV{TASK_WORK}\" && " .
964 "if which stdbuf >/dev/null ; then " .
965 " exec $stdbuf \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
967 " exec \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
972 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -poll=10000 ";
974 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
977 my @execargs = ('bash', '-c', $command);
978 srun (\@srunargs, \@execargs, undef, $build_script);
979 # exec() failed, we assume nothing happened.
980 die "srun() failed on build script\n";
983 if (!defined $childpid)
994 jobstepname => "$job_id.$id.$childpid",
996 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
997 $slot[$childslot]->{pid} = $childpid;
999 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
1000 Log ($id, "child $childpid started on $childslotname");
1001 $Jobstep->{starttime} = time;
1002 $Jobstep->{node} = $childnode->{name};
1003 $Jobstep->{slotindex} = $childslot;
1004 delete $Jobstep->{stderr};
1005 delete $Jobstep->{finishtime};
1006 delete $Jobstep->{tempfail};
1008 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
1009 $Jobstep->{'arvados_task'}->save;
1011 splice @jobstep_todo, $todo_ptr, 1;
1014 $progress_is_dirty = 1;
1018 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
1020 last THISROUND if $main::please_freeze;
1021 if ($main::please_info)
1023 $main::please_info = 0;
1025 create_output_collection();
1027 update_progress_stats();
1032 if (!$gotsome || ($latest_refresh + 2 < scalar time))
1034 check_refresh_wanted();
1036 update_progress_stats();
1038 elsif (time - $progress_stats_updated >= 30 || $progress_is_dirty)
1040 update_progress_stats();
1043 select (undef, undef, undef, 0.1);
1045 $working_slot_count = scalar(grep { $_->{node}->{fail_count} == 0 &&
1046 $_->{node}->{hold_count} < 4 } @slot);
1047 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
1048 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
1050 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
1051 .($thisround_failed+$thisround_succeeded)
1052 .") -- giving up on this round";
1053 Log (undef, $message);
1057 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1058 for (my $i=$#freeslot; $i>=0; $i--) {
1059 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1060 push @holdslot, (splice @freeslot, $i, 1);
1063 for (my $i=$#holdslot; $i>=0; $i--) {
1064 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1065 push @freeslot, (splice @holdslot, $i, 1);
1069 # give up if no nodes are succeeding
1070 if ($working_slot_count < 1) {
1071 Log(undef, "Every node has failed -- giving up");
1078 push @freeslot, splice @holdslot;
1079 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1082 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1085 if ($main::please_continue) {
1086 $main::please_continue = 0;
1089 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1091 if (!reapchildren())
1093 check_refresh_wanted();
1095 update_progress_stats();
1096 select (undef, undef, undef, 0.1);
1097 killem (keys %proc) if $main::please_freeze;
1101 update_progress_stats();
1102 freeze_if_want_freeze();
1105 if (!defined $main::success)
1107 if (!@jobstep_todo) {
1109 } elsif ($working_slot_count < 1) {
1110 save_output_collection();
1112 exit(EX_RETRY_UNLOCKED);
1113 } elsif ($thisround_succeeded == 0 &&
1114 ($thisround_failed == 0 || $thisround_failed > 4)) {
1115 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1116 Log (undef, $message);
1121 goto ONELEVEL if !defined $main::success;
1124 release_allocation();
1126 my $collated_output = save_output_collection();
1127 Log (undef, "finish");
1132 if ($collated_output && $main::success) {
1133 $final_state = 'Complete';
1135 $final_state = 'Failed';
1137 $Job->update_attributes('state' => $final_state);
1139 exit (($final_state eq 'Complete') ? 0 : 1);
1143 sub update_progress_stats
1145 $progress_stats_updated = time;
1146 return if !$progress_is_dirty;
1147 my ($todo, $done, $running) = (scalar @jobstep_todo,
1148 scalar @jobstep_done,
1149 scalar keys(%proc));
1150 $Job->{'tasks_summary'} ||= {};
1151 $Job->{'tasks_summary'}->{'todo'} = $todo;
1152 $Job->{'tasks_summary'}->{'done'} = $done;
1153 $Job->{'tasks_summary'}->{'running'} = $running;
1154 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1155 Log (undef, "status: $done done, $running running, $todo todo");
1156 $progress_is_dirty = 0;
1163 my $children_reaped = 0;
1164 my @successful_task_uuids = ();
1166 while((my $pid = waitpid (-1, WNOHANG)) > 0)
1168 my $childstatus = $?;
1170 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1172 . $slot[$proc{$pid}->{slot}]->{cpu});
1173 my $jobstepidx = $proc{$pid}->{jobstepidx};
1176 my $elapsed = time - $proc{$pid}->{time};
1177 my $Jobstep = $jobstep[$jobstepidx];
1179 my $exitvalue = $childstatus >> 8;
1180 my $exitinfo = "exit ".exit_status_s($childstatus);
1181 $Jobstep->{'arvados_task'}->reload;
1182 my $task_success = $Jobstep->{'arvados_task'}->{success};
1184 Log ($jobstepidx, "child $pid on $whatslot $exitinfo success=$task_success");
1186 if (!defined $task_success) {
1187 # task did not indicate one way or the other --> fail
1188 Log($jobstepidx, sprintf(
1189 "ERROR: Task process exited %s, but never updated its task record to indicate success and record its output.",
1190 exit_status_s($childstatus)));
1191 $Jobstep->{'arvados_task'}->{success} = 0;
1192 $Jobstep->{'arvados_task'}->save;
1199 $temporary_fail ||= $Jobstep->{tempfail};
1200 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1202 ++$thisround_failed;
1203 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1205 # Check for signs of a failed or misconfigured node
1206 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1207 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1208 # Don't count this against jobstep failure thresholds if this
1209 # node is already suspected faulty and srun exited quickly
1210 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1212 Log ($jobstepidx, "blaming failure on suspect node " .
1213 $slot[$proc{$pid}->{slot}]->{node}->{name});
1214 $temporary_fail ||= 1;
1216 ban_node_by_slot($proc{$pid}->{slot});
1219 Log ($jobstepidx, sprintf('failure (#%d, %s) after %d seconds',
1220 ++$Jobstep->{'failures'},
1221 $temporary_fail ? 'temporary' : 'permanent',
1224 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1225 # Give up on this task, and the whole job
1228 # Put this task back on the todo queue
1229 push @jobstep_todo, $jobstepidx;
1230 $Job->{'tasks_summary'}->{'failed'}++;
1234 push @successful_task_uuids, $Jobstep->{'arvados_task'}->{uuid};
1235 ++$thisround_succeeded;
1236 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1237 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1238 $slot[$proc{$pid}->{slot}]->{node}->{fail_count} = 0;
1239 push @jobstep_done, $jobstepidx;
1240 Log ($jobstepidx, "success in $elapsed seconds");
1242 $Jobstep->{exitcode} = $childstatus;
1243 $Jobstep->{finishtime} = time;
1244 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1245 $Jobstep->{'arvados_task'}->save;
1246 process_stderr_final ($jobstepidx);
1247 Log ($jobstepidx, sprintf("task output (%d bytes): %s",
1248 length($Jobstep->{'arvados_task'}->{output}),
1249 $Jobstep->{'arvados_task'}->{output}));
1251 close $reader{$jobstepidx};
1252 delete $reader{$jobstepidx};
1253 delete $slot[$proc{$pid}->{slot}]->{pid};
1254 push @freeslot, $proc{$pid}->{slot};
1257 $progress_is_dirty = 1;
1260 if (scalar(@successful_task_uuids) > 0)
1262 Log (undef, sprintf("%d tasks exited (%d succeeded), checking for new tasks from API server.", $children_reaped, scalar(@successful_task_uuids)));
1264 my $newtask_list = [];
1265 my $newtask_results;
1267 $newtask_results = api_call(
1269 'filters' => [["created_by_job_task_uuid","in",\@successful_task_uuids]],
1270 'order' => 'qsequence',
1271 'offset' => scalar(@$newtask_list),
1273 push(@$newtask_list, @{$newtask_results->{items}});
1274 } while (@{$newtask_results->{items}});
1275 Log (undef, sprintf("Got %d new tasks from API server.", scalar(@$newtask_list)));
1276 foreach my $arvados_task (@$newtask_list) {
1278 'level' => $arvados_task->{'sequence'},
1280 'arvados_task' => $arvados_task
1282 push @jobstep, $jobstep;
1283 push @jobstep_todo, $#jobstep;
1287 return $children_reaped;
1290 sub check_refresh_wanted
1292 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1294 $stat[9] > $latest_refresh &&
1295 # ...and we have actually locked the job record...
1296 $job_id eq $Job->{'uuid'}) {
1297 $latest_refresh = scalar time;
1298 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1299 for my $attr ('cancelled_at',
1300 'cancelled_by_user_uuid',
1301 'cancelled_by_client_uuid',
1303 $Job->{$attr} = $Job2->{$attr};
1305 if ($Job->{'state'} ne "Running") {
1306 if ($Job->{'state'} eq "Cancelled") {
1307 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1309 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1312 $main::please_freeze = 1;
1319 my $last_squeue_check = $squeue_checked;
1321 # Do not call `squeue` or check the kill list more than once every
1323 return if $last_squeue_check > time - 15;
1324 $squeue_checked = time;
1326 # Look for children from which we haven't received stderr data since
1327 # the last squeue check. If no such children exist, all procs are
1328 # alive and there's no need to even look at squeue.
1330 # As long as the crunchstat poll interval (10s) is shorter than the
1331 # squeue check interval (15s) this should make the squeue check an
1333 my $silent_procs = 0;
1334 for my $js (map {$jobstep[$_->{jobstepidx}]} values %proc)
1336 if (!exists($js->{stderr_at}))
1338 $js->{stderr_at} = 0;
1340 if ($js->{stderr_at} < $last_squeue_check)
1345 return if $silent_procs == 0;
1347 # use killem() on procs whose killtime is reached
1348 while (my ($pid, $procinfo) = each %proc)
1350 my $js = $jobstep[$procinfo->{jobstepidx}];
1351 if (exists $procinfo->{killtime}
1352 && $procinfo->{killtime} <= time
1353 && $js->{stderr_at} < $last_squeue_check)
1356 if ($js->{stderr_at}) {
1357 $sincewhen = " in last " . (time - $js->{stderr_at}) . "s";
1359 Log($procinfo->{jobstepidx}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
1366 # here is an opportunity to check for mysterious problems with local procs
1370 # Get a list of steps still running. Note: squeue(1) says --steps
1371 # selects a format (which we override anyway) and allows us to
1372 # specify which steps we're interested in (which we don't).
1373 # Importantly, it also changes the meaning of %j from "job name" to
1374 # "step name" and (although this isn't mentioned explicitly in the
1375 # docs) switches from "one line per job" mode to "one line per step"
1376 # mode. Without it, we'd just get a list of one job, instead of a
1378 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%j' --noheader`;
1381 Log(undef, "warning: squeue exit status $? ($!)");
1386 # which of my jobsteps are running, according to squeue?
1388 for my $jobstepname (@squeue)
1390 $ok{$jobstepname} = 1;
1393 # Check for child procs >60s old and not mentioned by squeue.
1394 while (my ($pid, $procinfo) = each %proc)
1396 if ($procinfo->{time} < time - 60
1397 && $procinfo->{jobstepname}
1398 && !exists $ok{$procinfo->{jobstepname}}
1399 && !exists $procinfo->{killtime})
1401 # According to slurm, this task has ended (successfully or not)
1402 # -- but our srun child hasn't exited. First we must wait (30
1403 # seconds) in case this is just a race between communication
1404 # channels. Then, if our srun child process still hasn't
1405 # terminated, we'll conclude some slurm communication
1406 # error/delay has caused the task to die without notifying srun,
1407 # and we'll kill srun ourselves.
1408 $procinfo->{killtime} = time + 30;
1409 Log($procinfo->{jobstepidx}, "notice: task is not in slurm queue but srun process $pid has not exited");
1416 # If a node fails in a multi-node "srun" call during job setup, the call
1417 # may hang instead of exiting with a nonzero code. This function checks
1418 # "sinfo" for the health of the nodes that were allocated and ensures that
1419 # they are all still in the "alloc" state. If a node that is allocated to
1420 # this job is not in "alloc" state, then set please_freeze.
1422 # This is only called from srun_sync() for node configuration. If a
1423 # node fails doing actual work, there are other recovery mechanisms.
1425 # Do not call `sinfo` more than once every 15 seconds.
1426 return if $sinfo_checked > time - 15;
1427 $sinfo_checked = time;
1429 # The output format "%t" means output node states.
1430 my @sinfo = `sinfo --nodes=\Q$ENV{SLURM_NODELIST}\E --noheader -o "%t"`;
1433 Log(undef, "warning: sinfo exit status $? ($!)");
1440 if ($_ != "alloc" && $_ != "alloc*") {
1441 $main::please_freeze = 1;
1446 sub release_allocation
1450 Log (undef, "release job allocation");
1451 system "scancel $ENV{SLURM_JOB_ID}";
1460 my $sel = IO::Select->new();
1461 foreach my $jobstepidx (keys %reader)
1463 my $fd = $reader{$jobstepidx};
1465 $fd_job{$fd} = $jobstepidx;
1467 if (my $stdout_fd = $jobstep[$jobstepidx]->{stdout_r}) {
1468 $sel->add($stdout_fd);
1469 $fd_job{$stdout_fd} = $jobstepidx;
1472 # select on all reader fds with 0.1s timeout
1473 my @ready_fds = $sel->can_read(0.1);
1474 foreach my $fd (@ready_fds)
1477 if (0 < sysread ($fd, $buf, 65536))
1480 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1482 my $jobstepidx = $fd_job{$fd};
1483 if ($jobstep[$jobstepidx]->{stdout_r} == $fd) {
1484 $jobstep[$jobstepidx]->{stdout_captured} .= $buf;
1488 $jobstep[$jobstepidx]->{stderr_at} = time;
1489 $jobstep[$jobstepidx]->{stderr} .= $buf;
1491 # Consume everything up to the last \n
1492 preprocess_stderr ($jobstepidx);
1494 if (length ($jobstep[$jobstepidx]->{stderr}) > 16384)
1496 # If we get a lot of stderr without a newline, chop off the
1497 # front to avoid letting our buffer grow indefinitely.
1498 substr ($jobstep[$jobstepidx]->{stderr},
1499 0, length($jobstep[$jobstepidx]->{stderr}) - 8192) = "";
1507 # Consume all full lines of stderr for a jobstep. Everything after the
1508 # last newline will remain in $jobstep[$jobstepidx]->{stderr} after
1510 sub preprocess_stderr
1512 my $jobstepidx = shift;
1513 # slotindex is only defined for children running Arvados job tasks.
1514 # Be prepared to handle the undef case (for setup srun calls, etc.).
1515 my $job_slot_index = $jobstep[$jobstepidx]->{slotindex};
1517 while ($jobstep[$jobstepidx]->{stderr} =~ /^(.*?)\n/) {
1519 substr $jobstep[$jobstepidx]->{stderr}, 0, 1+length($line), "";
1520 Log ($jobstepidx, "stderr $line");
1521 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/i) {
1522 # If the allocation is revoked, we can't possibly continue, so mark all
1523 # nodes as failed. This will cause the overall exit code to be
1524 # EX_RETRY_UNLOCKED instead of failure so that crunch_dispatch can re-run
1526 $main::please_freeze = 1;
1527 foreach my $st (@slot) {
1528 $st->{node}->{fail_count}++;
1531 elsif ($line =~ /srun: error: .*?\b(Node failure on|Aborting, .*?\bio error\b)/i) {
1532 $jobstep[$jobstepidx]->{tempfail} = 1;
1533 if (defined($job_slot_index)) {
1534 $slot[$job_slot_index]->{node}->{fail_count}++;
1535 ban_node_by_slot($job_slot_index);
1538 elsif ($line =~ /srun: error: (Unable to create job step|.*?: Communication connection failure)/i) {
1539 $jobstep[$jobstepidx]->{tempfail} = 1;
1540 ban_node_by_slot($job_slot_index) if (defined($job_slot_index));
1542 elsif ($line =~ /\bKeep(Read|Write|Request)Error:/) {
1543 $jobstep[$jobstepidx]->{tempfail} = 1;
1549 sub process_stderr_final
1551 my $jobstepidx = shift;
1552 preprocess_stderr ($jobstepidx);
1555 Log ($jobstepidx, "stderr $_");
1556 } split ("\n", $jobstep[$jobstepidx]->{stderr});
1557 $jobstep[$jobstepidx]->{stderr} = '';
1564 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1565 Log(undef, "fetch_block run error from arv-get $hash: $!");
1568 my $output_block = "";
1571 my $bytes = sysread($keep, $buf, 1024 * 1024);
1572 if (!defined $bytes) {
1573 Log(undef, "fetch_block read error from arv-get: $!");
1574 $output_block = undef;
1576 } elsif ($bytes == 0) {
1577 # sysread returns 0 at the end of the pipe.
1580 # some bytes were read into buf.
1581 $output_block .= $buf;
1586 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1587 $output_block = undef;
1589 return $output_block;
1592 # Create a collection by concatenating the output of all tasks (each
1593 # task's output is either a manifest fragment, a locator for a
1594 # manifest fragment stored in Keep, or nothing at all). Return the
1595 # portable_data_hash of the new collection.
1596 sub create_output_collection
1598 Log (undef, "collate");
1600 my ($child_out, $child_in);
1601 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1604 print (arvados.api("v1").collections().
1605 create(body={"manifest_text": sys.stdin.read(),
1606 "owner_uuid": sys.argv[2]}).
1607 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1608 }, retry_count(), $Job->{owner_uuid});
1611 my $manifest_size = 0;
1615 my $output = $_->{'arvados_task'}->{output};
1616 next if (!defined($output));
1618 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1619 $next_write = fetch_block($output);
1621 $next_write = $output;
1623 if (defined($next_write)) {
1624 if (!defined(syswrite($child_in, $next_write))) {
1625 # There's been an error writing. Stop the loop.
1626 # We'll log details about the exit code later.
1629 $manifest_size += length($next_write);
1632 my $uuid = $_->{'arvados_task'}->{'uuid'};
1633 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1638 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1641 my $s = IO::Select->new($child_out);
1642 if ($s->can_read(120)) {
1643 sysread($child_out, $joboutput, 1024 * 1024);
1646 Log(undef, "output collection creation exited " . exit_status_s($?));
1652 Log (undef, "timed out while creating output collection");
1653 foreach my $signal (2, 2, 2, 15, 15, 9) {
1654 kill($signal, $pid);
1655 last if waitpid($pid, WNOHANG) == -1;
1664 # Calls create_output_collection, logs the result, and returns it.
1665 # If that was successful, save that as the output in the job record.
1666 sub save_output_collection {
1667 my $collated_output = create_output_collection();
1669 if (!$collated_output) {
1670 Log(undef, "Failed to write output collection");
1673 Log(undef, "job output $collated_output");
1674 $Job->update_attributes('output' => $collated_output);
1676 return $collated_output;
1683 my $sig = 2; # SIGINT first
1684 if (exists $proc{$_}->{"sent_$sig"} &&
1685 time - $proc{$_}->{"sent_$sig"} > 4)
1687 $sig = 15; # SIGTERM if SIGINT doesn't work
1689 if (exists $proc{$_}->{"sent_$sig"} &&
1690 time - $proc{$_}->{"sent_$sig"} > 4)
1692 $sig = 9; # SIGKILL if SIGTERM doesn't work
1694 if (!exists $proc{$_}->{"sent_$sig"})
1696 Log ($proc{$_}->{jobstepidx}, "sending 2x signal $sig to pid $_");
1698 select (undef, undef, undef, 0.1);
1701 kill $sig, $_; # srun wants two SIGINT to really interrupt
1703 $proc{$_}->{"sent_$sig"} = time;
1704 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1714 vec($bits,fileno($_),1) = 1;
1720 # Send log output to Keep via arv-put.
1722 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1723 # $log_pipe_out_buf is a string containing all output read from arv-put so far.
1724 # $log_pipe_out_select is an IO::Select object around $log_pipe_out.
1725 # $log_pipe_pid is the pid of the arv-put subprocess.
1727 # The only functions that should access these variables directly are:
1729 # log_writer_start($logfilename)
1730 # Starts an arv-put pipe, reading data on stdin and writing it to
1731 # a $logfilename file in an output collection.
1733 # log_writer_read_output([$timeout])
1734 # Read output from $log_pipe_out and append it to $log_pipe_out_buf.
1735 # Passes $timeout to the select() call, with a default of 0.01.
1736 # Returns the result of the last read() call on $log_pipe_out, or
1737 # -1 if read() wasn't called because select() timed out.
1738 # Only other log_writer_* functions should need to call this.
1740 # log_writer_send($txt)
1741 # Writes $txt to the output log collection.
1743 # log_writer_finish()
1744 # Closes the arv-put pipe and returns the output that it produces.
1746 # log_writer_is_active()
1747 # Returns a true value if there is currently a live arv-put
1748 # process, false otherwise.
1750 my ($log_pipe_in, $log_pipe_out, $log_pipe_out_buf, $log_pipe_out_select,
1753 sub log_writer_start($)
1755 my $logfilename = shift;
1756 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1760 '--filename', $logfilename,
1762 $log_pipe_out_buf = "";
1763 $log_pipe_out_select = IO::Select->new($log_pipe_out);
1766 sub log_writer_read_output {
1767 my $timeout = shift || 0.01;
1769 while ($read && $log_pipe_out_select->can_read($timeout)) {
1770 $read = read($log_pipe_out, $log_pipe_out_buf, 65536,
1771 length($log_pipe_out_buf));
1773 if (!defined($read)) {
1774 Log(undef, "error reading log manifest from arv-put: $!");
1779 sub log_writer_send($)
1782 print $log_pipe_in $txt;
1783 log_writer_read_output();
1786 sub log_writer_finish()
1788 return unless $log_pipe_pid;
1790 close($log_pipe_in);
1792 my $logger_failed = 0;
1793 my $read_result = log_writer_read_output(600);
1794 if ($read_result == -1) {
1795 $logger_failed = -1;
1796 Log (undef, "timed out reading from 'arv-put'");
1797 } elsif ($read_result != 0) {
1798 $logger_failed = -2;
1799 Log(undef, "failed to read arv-put log manifest to EOF");
1802 waitpid($log_pipe_pid, 0);
1804 $logger_failed ||= $?;
1805 Log(undef, "log_writer_finish: arv-put exited " . exit_status_s($?))
1808 close($log_pipe_out);
1809 my $arv_put_output = $logger_failed ? undef : $log_pipe_out_buf;
1810 $log_pipe_pid = $log_pipe_in = $log_pipe_out = $log_pipe_out_buf =
1811 $log_pipe_out_select = undef;
1813 return $arv_put_output;
1816 sub log_writer_is_active() {
1817 return $log_pipe_pid;
1820 sub Log # ($jobstepidx, $logmessage)
1822 my ($jobstepidx, $logmessage) = @_;
1823 if ($logmessage =~ /\n/) {
1824 for my $line (split (/\n/, $_[1])) {
1825 Log ($jobstepidx, $line);
1829 my $fh = select STDERR; $|=1; select $fh;
1831 if (defined($jobstepidx) && exists($jobstep[$jobstepidx]->{arvados_task})) {
1832 $task_qseq = $jobstepidx;
1834 my $message = sprintf ("%s %d %s %s", $job_id, $$, $task_qseq, $logmessage);
1835 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1838 if (log_writer_is_active() || -t STDERR) {
1839 my @gmtime = gmtime;
1840 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1841 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1843 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1845 if (log_writer_is_active()) {
1846 log_writer_send($datetime . " " . $message);
1853 my ($package, $file, $line) = caller;
1854 my $message = "@_ at $file line $line\n";
1855 Log (undef, $message);
1856 release_allocation();
1857 freeze() if @jobstep_todo;
1858 create_output_collection() if @jobstep_todo;
1868 if ($Job->{'state'} eq 'Cancelled') {
1869 $Job->update_attributes('finished_at' => scalar gmtime);
1871 $Job->update_attributes('state' => 'Failed');
1878 my $justcheckpoint = shift; # false if this will be the last meta saved
1879 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1880 return unless log_writer_is_active();
1881 my $log_manifest = log_writer_finish();
1882 return unless defined($log_manifest);
1885 my $prev_log_coll = api_call("collections/get", uuid => $Job->{log});
1886 $log_manifest = $prev_log_coll->{manifest_text} . $log_manifest;
1889 my $log_coll = api_call(
1890 "collections/create", ensure_unique_name => 1, collection => {
1891 manifest_text => $log_manifest,
1892 owner_uuid => $Job->{owner_uuid},
1893 name => sprintf("Log from %s job %s", $Job->{script}, $Job->{uuid}),
1895 Log(undef, "log collection is " . $log_coll->{portable_data_hash});
1896 $Job->update_attributes('log' => $log_coll->{portable_data_hash});
1900 sub freeze_if_want_freeze
1902 if ($main::please_freeze)
1904 release_allocation();
1907 # kill some srun procs before freeze+stop
1908 map { $proc{$_} = {} } @_;
1911 killem (keys %proc);
1912 select (undef, undef, undef, 0.1);
1914 while (($died = waitpid (-1, WNOHANG)) > 0)
1916 delete $proc{$died};
1921 create_output_collection();
1931 Log (undef, "Freeze not implemented");
1938 croak ("Thaw not implemented");
1954 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1960 my $srunargs = shift;
1961 my $execargs = shift;
1962 my $opts = shift || {};
1965 my $label = exists $opts->{label} ? $opts->{label} : "@$execargs";
1966 Log (undef, "$label: start");
1968 my ($stderr_r, $stderr_w);
1969 pipe $stderr_r, $stderr_w or croak("pipe() failed: $!");
1971 my ($stdout_r, $stdout_w);
1972 pipe $stdout_r, $stdout_w or croak("pipe() failed: $!");
1974 my $srunpid = fork();
1979 fcntl($stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
1980 fcntl($stdout_w, F_SETFL, 0) or croak($!);
1981 open(STDERR, ">&", $stderr_w);
1982 open(STDOUT, ">&", $stdout_w);
1983 srun ($srunargs, $execargs, $opts, $stdin);
1989 set_nonblocking($stderr_r);
1990 set_nonblocking($stdout_r);
1992 # Add entries to @jobstep and %proc so check_squeue() and
1993 # freeze_if_want_freeze() can treat it like a job task process.
1997 stderr_captured => '',
1998 stdout_r => $stdout_r,
1999 stdout_captured => '',
2001 my $jobstepidx = $#jobstep;
2003 jobstepidx => $jobstepidx,
2005 $reader{$jobstepidx} = $stderr_r;
2007 while ($srunpid != waitpid ($srunpid, WNOHANG)) {
2008 my $busy = readfrompipes();
2009 if (!$busy || ($latest_refresh + 2 < scalar time)) {
2010 check_refresh_wanted();
2015 select(undef, undef, undef, 0.1);
2017 killem(keys %proc) if $main::please_freeze;
2021 1 while readfrompipes();
2022 process_stderr_final ($jobstepidx);
2024 Log (undef, "$label: exit ".exit_status_s($exited));
2028 delete $proc{$srunpid};
2029 delete $reader{$jobstepidx};
2031 my $j = pop @jobstep;
2032 # If the srun showed signs of tempfail, ensure the caller treats that as a
2034 if ($main::please_freeze || $j->{tempfail}) {
2037 return ($exited, $j->{stdout_captured}, $j->{stderr_captured});
2043 my $srunargs = shift;
2044 my $execargs = shift;
2045 my $opts = shift || {};
2047 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
2049 $Data::Dumper::Terse = 1;
2050 $Data::Dumper::Indent = 0;
2051 my $show_cmd = Dumper($args);
2052 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
2053 $show_cmd =~ s/\n/ /g;
2054 if ($opts->{fork}) {
2055 Log(undef, "starting: $show_cmd");
2057 # This is a child process: parent is in charge of reading our
2058 # stderr and copying it to Log() if needed.
2059 warn "starting: $show_cmd\n";
2062 if (defined $stdin) {
2063 my $child = open STDIN, "-|";
2064 defined $child or die "no fork: $!";
2066 print $stdin or die $!;
2067 close STDOUT or die $!;
2072 return system (@$args) if $opts->{fork};
2075 warn "ENV size is ".length(join(" ",%ENV));
2076 die "exec failed: $!: @$args";
2080 sub ban_node_by_slot {
2081 # Don't start any new jobsteps on this node for 60 seconds
2083 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
2084 $slot[$slotid]->{node}->{hold_count}++;
2085 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
2090 my ($lockfile, $error_message) = @_;
2091 open L, ">", $lockfile or croak("$lockfile: $!");
2092 if (!flock L, LOCK_EX|LOCK_NB) {
2093 croak("Can't lock $lockfile: $error_message\n");
2097 sub find_docker_image {
2098 # Given a Keep locator, check to see if it contains a Docker image.
2099 # If so, return its stream name and Docker hash.
2100 # If not, return undef for both values.
2101 my $locator = shift;
2102 my ($streamname, $filename);
2103 my $image = api_call("collections/get", uuid => $locator);
2105 foreach my $line (split(/\n/, $image->{manifest_text})) {
2106 my @tokens = split(/\s+/, $line);
2108 $streamname = shift(@tokens);
2109 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
2110 if (defined($filename)) {
2111 return (undef, undef); # More than one file in the Collection.
2113 $filename = (split(/:/, $filedata, 3))[2];
2118 if (defined($filename) and ($filename =~ /^((?:sha256:)?[0-9A-Fa-f]{64})\.tar$/)) {
2119 return ($streamname, $1);
2121 return (undef, undef);
2126 # Calculate the number of times an operation should be retried,
2127 # assuming exponential backoff, and that we're willing to retry as
2128 # long as tasks have been running. Enforce a minimum of 3 retries.
2129 my ($starttime, $endtime, $timediff, $retries);
2131 $starttime = $jobstep[0]->{starttime};
2132 $endtime = $jobstep[-1]->{finishtime};
2134 if (!defined($starttime)) {
2136 } elsif (!defined($endtime)) {
2137 $timediff = time - $starttime;
2139 $timediff = ($endtime - $starttime) - (time - $endtime);
2141 if ($timediff > 0) {
2142 $retries = int(log($timediff) / log(2));
2144 $retries = 1; # Use the minimum.
2146 return ($retries > 3) ? $retries : 3;
2150 # Pass in two function references.
2151 # This method will be called with the remaining arguments.
2152 # If it dies, retry it with exponential backoff until it succeeds,
2153 # or until the current retry_count is exhausted. After each failure
2154 # that can be retried, the second function will be called with
2155 # the current try count (0-based), next try time, and error message.
2156 my $operation = shift;
2157 my $retry_callback = shift;
2158 my $retries = retry_count();
2159 foreach my $try_count (0..$retries) {
2160 my $next_try = time + (2 ** $try_count);
2161 my $result = eval { $operation->(@_); };
2164 } elsif ($try_count < $retries) {
2165 $retry_callback->($try_count, $next_try, $@);
2166 my $sleep_time = $next_try - time;
2167 sleep($sleep_time) if ($sleep_time > 0);
2170 # Ensure the error message ends in a newline, so Perl doesn't add
2171 # retry_op's line number to it.
2177 # Pass in a /-separated API method name, and arguments for it.
2178 # This function will call that method, retrying as needed until
2179 # the current retry_count is exhausted, with a log on the first failure.
2180 my $method_name = shift;
2181 my $log_api_retry = sub {
2182 my ($try_count, $next_try_at, $errmsg) = @_;
2183 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
2184 $errmsg =~ s/\s/ /g;
2185 $errmsg =~ s/\s+$//;
2187 if ($next_try_at < time) {
2188 $retry_msg = "Retrying.";
2190 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
2191 $retry_msg = "Retrying at $next_try_fmt.";
2193 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
2196 foreach my $key (split(/\//, $method_name)) {
2197 $method = $method->{$key};
2199 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
2203 # Given a $?, return a human-readable exit code string like "0" or
2204 # "1" or "0 with signal 1" or "1 with signal 11".
2205 my $exitcode = shift;
2206 my $s = $exitcode >> 8;
2207 if ($exitcode & 0x7f) {
2208 $s .= " with signal " . ($exitcode & 0x7f);
2210 if ($exitcode & 0x80) {
2211 $s .= " with core dump";
2216 sub handle_readall {
2217 # Pass in a glob reference to a file handle.
2218 # Read all its contents and return them as a string.
2219 my $fh_glob_ref = shift;
2221 return <$fh_glob_ref>;
2224 sub tar_filename_n {
2226 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
2229 sub add_git_archive {
2230 # Pass in a git archive command as a string or list, a la system().
2231 # This method will save its output to be included in the archive sent to the
2235 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
2236 croak("Failed to save git archive: $!");
2238 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
2240 waitpid($git_pid, 0);
2243 croak("Failed to save git archive: git exited " . exit_status_s($?));
2247 sub combined_git_archive {
2248 # Combine all saved tar archives into a single archive, then return its
2249 # contents in a string. Return undef if no archives have been saved.
2250 if ($git_tar_count < 1) {
2253 my $base_tar_name = tar_filename_n(1);
2254 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
2255 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
2256 if ($tar_exit != 0) {
2257 croak("Error preparing build archive: tar -A exited " .
2258 exit_status_s($tar_exit));
2261 if (!open(GIT_TAR, "<", $base_tar_name)) {
2262 croak("Could not open build archive: $!");
2264 my $tar_contents = handle_readall(\*GIT_TAR);
2266 return $tar_contents;
2269 sub set_nonblocking {
2271 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
2272 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
2278 # This is crunch-job's internal dispatch script. crunch-job running on the API
2279 # server invokes this script on individual compute nodes, or localhost if we're
2280 # running a job locally. It gets called in two modes:
2282 # * No arguments: Installation mode. Read a tar archive from the DATA
2283 # file handle; it includes the Crunch script's source code, and
2284 # maybe SDKs as well. Those should be installed in the proper
2285 # locations. This runs outside of any Docker container, so don't try to
2286 # introspect Crunch's runtime environment.
2288 # * With arguments: Crunch script run mode. This script should set up the
2289 # environment, then run the command specified in the arguments. This runs
2290 # inside any Docker container.
2293 use File::Path qw( make_path remove_tree );
2294 use POSIX qw(getcwd);
2296 use constant TASK_TEMPFAIL => 111;
2298 # Map SDK subdirectories to the path environments they belong to.
2299 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2301 my $destdir = $ENV{"CRUNCH_SRC"};
2302 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2303 my $repo = $ENV{"CRUNCH_SRC_URL"};
2304 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2305 my $job_work = $ENV{"JOB_WORK"};
2306 my $task_work = $ENV{"TASK_WORK"};
2308 open(STDOUT_ORIG, ">&", STDOUT);
2309 open(STDERR_ORIG, ">&", STDERR);
2311 for my $dir ($destdir, $job_work, $task_work) {
2314 -e $dir or die "Failed to create temporary directory ($dir): $!";
2319 remove_tree($task_work, {keep_root => 1});
2322 ### Crunch script run mode
2324 # We want to do routine logging during task 0 only. This gives the user
2325 # the information they need, but avoids repeating the information for every
2328 if ($ENV{TASK_SEQUENCE} eq "0") {
2331 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2337 my $python_src = "$install_dir/python";
2338 my $venv_dir = "$job_work/.arvados.venv";
2339 my $venv_built = -e "$venv_dir/bin/activate";
2340 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2341 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2342 "--python=python2.7", $venv_dir);
2343 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2345 $Log->("Built Python SDK virtualenv");
2348 my @pysdk_version_cmd = ("python", "-c",
2349 "from pkg_resources import get_distribution as get; print get('arvados-python-client').version");
2351 $Log->("Running in Python SDK virtualenv");
2352 @pysdk_version_cmd = ();
2353 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2354 @ARGV = ("/bin/sh", "-ec",
2355 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2356 } elsif (-d $python_src) {
2357 $Log->("Warning: virtualenv not found inside Docker container default " .
2358 "\$PATH. Can't install Python SDK.");
2361 if (@pysdk_version_cmd) {
2362 open(my $pysdk_version_pipe, "-|", @pysdk_version_cmd);
2363 my $pysdk_version = <$pysdk_version_pipe>;
2364 close($pysdk_version_pipe);
2366 chomp($pysdk_version);
2367 $Log->("Using Arvados SDK version $pysdk_version");
2369 # A lot could've gone wrong here, but pretty much all of it means that
2370 # Python won't be able to load the Arvados SDK.
2371 $Log->("Warning: Arvados SDK not found");
2375 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2376 my $sdk_path = "$install_dir/$sdk_dir";
2378 if ($ENV{$sdk_envkey}) {
2379 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2381 $ENV{$sdk_envkey} = $sdk_path;
2383 $Log->("Arvados SDK added to %s", $sdk_envkey);
2388 die "Cannot exec `@ARGV`: $!";
2391 ### Installation mode
2392 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2394 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2395 # This exact git archive (source + arvados sdk) is already installed
2396 # here, so there's no need to reinstall it.
2398 # We must consume our DATA section, though: otherwise the process
2399 # feeding it to us will get SIGPIPE.
2401 while (read(DATA, $buf, 65536)) { }
2406 unlink "$destdir.archive_hash";
2410 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2411 local $SIG{PIPE} = "IGNORE";
2412 warn "Extracting archive: $archive_hash\n";
2413 # --ignore-zeros is necessary sometimes: depending on how much NUL
2414 # padding tar -A put on our combined archive (which in turn depends
2415 # on the length of the component archives) tar without
2416 # --ignore-zeros will exit before consuming stdin and cause close()
2417 # to fail on the resulting SIGPIPE.
2418 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2419 die "Error launching 'tar -xC $destdir': $!";
2421 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2422 # get SIGPIPE. We must feed it data incrementally.
2424 while (read(DATA, $tar_input, 65536)) {
2425 print TARX $tar_input;
2428 die "'tar -xC $destdir' exited $?: $!";
2434 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2436 foreach my $sdk_lang (("python",
2437 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2438 if (-d "$sdk_root/$sdk_lang") {
2439 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2440 die "Failed to install $sdk_lang SDK: $!";
2446 my $python_dir = "$install_dir/python";
2447 if ((-d $python_dir) and can_run("python2.7")) {
2448 open(my $egg_info_pipe, "-|",
2449 "python2.7 \Q$python_dir/setup.py\E egg_info 2>&1 >/dev/null");
2450 my @egg_info_errors = <$egg_info_pipe>;
2451 close($egg_info_pipe);
2454 if (@egg_info_errors and (($egg_info_errors[-1] =~ /\bgit\b/) or ($egg_info_errors[-1] =~ /\[Errno 2\]/))) {
2455 # egg_info apparently failed because it couldn't ask git for a build tag.
2456 # Specify no build tag.
2457 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2458 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2461 my $egg_info_exit = $? >> 8;
2462 foreach my $errline (@egg_info_errors) {
2465 warn "python setup.py egg_info failed: exit $egg_info_exit";
2466 exit ($egg_info_exit || 1);
2471 # Hide messages from the install script (unless it fails: shell_or_die
2472 # will show $destdir.log in that case).
2473 open(STDOUT, ">>", "$destdir.log");
2474 open(STDERR, ">&", STDOUT);
2476 if (-e "$destdir/crunch_scripts/install") {
2477 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2478 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2480 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2481 } elsif (-e "./install.sh") {
2482 shell_or_die (undef, "./install.sh", $install_dir);
2485 if ($archive_hash) {
2486 unlink "$destdir.archive_hash.new";
2487 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2488 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2494 my $command_name = shift;
2495 open(my $which, "-|", "which", $command_name);
2496 while (<$which>) { }
2503 my $exitcode = shift;
2505 if ($ENV{"DEBUG"}) {
2506 print STDERR "@_\n";
2508 if (system (@_) != 0) {
2511 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2512 open STDERR, ">&STDERR_ORIG";
2513 system ("cat $destdir.log >&2");
2514 warn "@_ failed ($err): $exitstatus";
2515 if (defined($exitcode)) {
2519 exit (($code >> 8) || 1);