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 $docker_install_script = qq{
421 id=\$($docker_bin inspect --format="{{.ID}}" \Q$docker_hash\E) || return 1
422 echo "image ID is \$id"
423 [[ \${id} = \Q$docker_hash\E ]]
425 if loaded >&2 2>/dev/null; then
426 echo >&2 "image is already present"
429 echo >&2 "docker image is not present; loading"
430 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
431 if ! loaded >&2; then
432 echo >&2 "`docker load` exited 0, but image is not found (!)"
435 echo >&2 "image loaded successfully"
438 my ($exited, $stdout, $stderr) = srun_sync(
439 ["srun", "--nodelist=" . join(',', @node)],
440 ["/bin/bash", "-o", "pipefail", "-ec", $docker_install_script],
441 {label => "load docker image"});
444 exit(EX_RETRY_UNLOCKED);
447 # Determine whether this version of Docker supports memory+swap limits.
448 ($exited, $stdout, $stderr) = srun_sync(
449 ["srun", "--nodes=1"],
450 [$docker_bin, 'run', '--help'],
451 {label => "check --memory-swap feature"});
452 $docker_limitmem = ($stdout =~ /--memory-swap/);
454 # Find a non-root Docker user to use.
455 # Tries the default user for the container, then 'crunch', then 'nobody',
456 # testing for whether the actual user id is non-zero. This defends against
457 # mistakes but not malice, but we intend to harden the security in the future
458 # so we don't want anyone getting used to their jobs running as root in their
460 my @tryusers = ("", "crunch", "nobody");
461 foreach my $try_user (@tryusers) {
464 if ($try_user eq "") {
465 $label = "check whether default user is UID 0";
468 $label = "check whether user '$try_user' is UID 0";
469 $try_user_arg = "--user=$try_user";
471 my ($exited, $stdout, $stderr) = srun_sync(
472 ["srun", "--nodes=1"],
474 "$docker_bin run $docker_run_args $try_user_arg $docker_hash id --user"],
477 if ($exited == 0 && $stdout =~ /^\d+$/ && $stdout > 0) {
478 $dockeruserarg = $try_user_arg;
479 if ($try_user eq "") {
480 Log(undef, "Container will run with default user");
482 Log(undef, "Container will run with $dockeruserarg");
488 if (!defined $dockeruserarg) {
489 croak("Could not find a user in container that is not UID 0 (tried default user, @tryusers) or there was a problem running 'id' in the container.");
492 if ($Job->{arvados_sdk_version}) {
493 # The job also specifies an Arvados SDK version. Add the SDKs to the
494 # tar file for the build script to install.
495 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
496 $Job->{arvados_sdk_version}));
497 add_git_archive("git", "--git-dir=$git_dir", "archive",
498 "--prefix=.arvados.sdk/",
499 $Job->{arvados_sdk_version}, "sdk");
503 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
504 # If script_version looks like an absolute path, *and* the --git-dir
505 # argument was not given -- which implies we were not invoked by
506 # crunch-dispatch -- we will use the given path as a working
507 # directory instead of resolving script_version to a git commit (or
508 # doing anything else with git).
509 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
510 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
513 # Resolve the given script_version to a git commit sha1. Also, if
514 # the repository is remote, clone it into our local filesystem: this
515 # ensures "git archive" will work, and is necessary to reliably
516 # resolve a symbolic script_version like "master^".
517 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
519 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
521 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
523 # If we're running under crunch-dispatch, it will have already
524 # pulled the appropriate source tree into its own repository, and
525 # given us that repo's path as $git_dir.
527 # If we're running a "local" job, we might have to fetch content
528 # from a remote repository.
530 # (Currently crunch-dispatch gives a local path with --git-dir, but
531 # we might as well accept URLs there too in case it changes its
533 my $repo = $git_dir || $Job->{'repository'};
535 # Repository can be remote or local. If remote, we'll need to fetch it
536 # to a local dir before doing `git log` et al.
539 if ($repo =~ m{://|^[^/]*:}) {
540 # $repo is a git url we can clone, like git:// or https:// or
541 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
542 # not recognized here because distinguishing that from a local
543 # path is too fragile. If you really need something strange here,
544 # use the ssh:// form.
545 $repo_location = 'remote';
546 } elsif ($repo =~ m{^\.*/}) {
547 # $repo is a local path to a git index. We'll also resolve ../foo
548 # to ../foo/.git if the latter is a directory. To help
549 # disambiguate local paths from named hosted repositories, this
550 # form must be given as ./ or ../ if it's a relative path.
551 if (-d "$repo/.git") {
552 $repo = "$repo/.git";
554 $repo_location = 'local';
556 # $repo is none of the above. It must be the name of a hosted
558 my $arv_repo_list = api_call("repositories/list",
559 'filters' => [['name','=',$repo]]);
560 my @repos_found = @{$arv_repo_list->{'items'}};
561 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
563 Log(undef, "Repository '$repo' -> "
564 . join(", ", map { $_->{'uuid'} } @repos_found));
567 croak("Error: Found $n_found repositories with name '$repo'.");
569 $repo = $repos_found[0]->{'fetch_url'};
570 $repo_location = 'remote';
572 Log(undef, "Using $repo_location repository '$repo'");
573 $ENV{"CRUNCH_SRC_URL"} = $repo;
575 # Resolve given script_version (we'll call that $treeish here) to a
576 # commit sha1 ($commit).
577 my $treeish = $Job->{'script_version'};
579 if ($repo_location eq 'remote') {
580 # We minimize excess object-fetching by re-using the same bare
581 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
582 # just keep adding remotes to it as needed.
583 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
584 my $gitcmd = "git --git-dir=\Q$local_repo\E";
586 # Set up our local repo for caching remote objects, making
588 if (!-d $local_repo) {
589 make_path($local_repo) or croak("Error: could not create $local_repo");
591 # This works (exits 0 and doesn't delete fetched objects) even
592 # if $local_repo is already initialized:
593 `$gitcmd init --bare`;
595 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
598 # If $treeish looks like a hash (or abbrev hash) we look it up in
599 # our local cache first, since that's cheaper. (We don't want to
600 # do that with tags/branches though -- those change over time, so
601 # they should always be resolved by the remote repo.)
602 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
603 # Hide stderr because it's normal for this to fail:
604 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
606 # Careful not to resolve a branch named abcdeff to commit 1234567:
607 $sha1 =~ /^$treeish/ &&
608 $sha1 =~ /^([0-9a-f]{40})$/s) {
610 Log(undef, "Commit $commit already present in $local_repo");
614 if (!defined $commit) {
615 # If $treeish isn't just a hash or abbrev hash, or isn't here
616 # yet, we need to fetch the remote to resolve it correctly.
618 # First, remove all local heads. This prevents a name that does
619 # not exist on the remote from resolving to (or colliding with)
620 # a previously fetched branch or tag (possibly from a different
622 remove_tree("$local_repo/refs/heads", {keep_root => 1});
624 Log(undef, "Fetching objects from $repo to $local_repo");
625 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
627 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
631 # Now that the data is all here, we will use our local repo for
632 # the rest of our git activities.
636 my $gitcmd = "git --git-dir=\Q$repo\E";
637 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
638 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
639 croak("`$gitcmd rev-list` exited "
641 .", '$treeish' not found, giving up");
644 Log(undef, "Version $treeish is commit $commit");
646 if ($commit ne $Job->{'script_version'}) {
647 # Record the real commit id in the database, frozentokey, logs,
648 # etc. -- instead of an abbreviation or a branch name which can
649 # become ambiguous or point to a different commit in the future.
650 if (!$Job->update_attributes('script_version' => $commit)) {
651 croak("Error: failed to update job's script_version attribute");
655 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
656 add_git_archive("$gitcmd archive ''\Q$commit\E");
659 my $git_archive = combined_git_archive();
660 if (!defined $git_archive) {
661 Log(undef, "Skip install phase (no git archive)");
663 Log(undef, "Warning: This probably means workers have no source tree!");
668 my $install_script_tries_left = 3;
669 for (my $attempts = 0; $attempts < 3; $attempts++) {
670 my @srunargs = ("srun",
671 "--nodelist=$nodelist",
672 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
673 my @execargs = ("sh", "-c",
674 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
676 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
677 my ($stdout, $stderr);
678 ($exited, $stdout, $stderr) = srun_sync(
679 \@srunargs, \@execargs,
680 {label => "run install script on all workers"},
681 $build_script . $git_archive);
683 my $stderr_anything_from_script = 0;
684 for my $line (split(/\n/, $stderr)) {
685 if ($line !~ /^(srun: error: |starting: \[)/) {
686 $stderr_anything_from_script = 1;
690 last if $exited == 0 || $main::please_freeze;
692 # If the install script fails but doesn't print an error message,
693 # the next thing anyone is likely to do is just run it again in
694 # case it was a transient problem like "slurm communication fails
695 # because the network isn't reliable enough". So we'll just do
696 # that ourselves (up to 3 attempts in total). OTOH, if there is an
697 # error message, the problem is more likely to have a real fix and
698 # we should fail the job so the fixing process can start, instead
699 # of doing 2 more attempts.
700 last if $stderr_anything_from_script;
703 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
704 unlink($tar_filename);
712 foreach (qw (script script_version script_parameters runtime_constraints))
716 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
718 foreach (split (/\n/, $Job->{knobs}))
720 Log (undef, "knob " . $_);
724 'filters' => [['hostname', 'in', \@node]],
725 'order' => 'hostname',
726 'limit' => scalar(@node),
728 for my $n (@{$resp->{items}}) {
729 Log(undef, "$n->{hostname} $n->{uuid} ".JSON::encode_json($n->{properties}));
734 $main::success = undef;
740 my $thisround_succeeded = 0;
741 my $thisround_failed = 0;
742 my $thisround_failed_multiple = 0;
743 my $working_slot_count = scalar(@slot);
745 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
746 or $a <=> $b } @jobstep_todo;
747 my $level = $jobstep[$jobstep_todo[0]]->{level};
749 my $initial_tasks_this_level = 0;
750 foreach my $id (@jobstep_todo) {
751 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
754 # If the number of tasks scheduled at this level #T is smaller than the number
755 # of slots available #S, only use the first #T slots, or the first slot on
756 # each node, whichever number is greater.
758 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
759 # based on these numbers. Using fewer slots makes more resources available
760 # to each individual task, which should normally be a better strategy when
761 # there are fewer of them running with less parallelism.
763 # Note that this calculation is not redone if the initial tasks at
764 # this level queue more tasks at the same level. This may harm
765 # overall task throughput for that level.
767 if ($initial_tasks_this_level < @node) {
768 @freeslot = (0..$#node);
769 } elsif ($initial_tasks_this_level < @slot) {
770 @freeslot = (0..$initial_tasks_this_level - 1);
772 @freeslot = (0..$#slot);
774 my $round_num_freeslots = scalar(@freeslot);
775 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";
777 my %round_max_slots = ();
778 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
779 my $this_slot = $slot[$freeslot[$ii]];
780 my $node_name = $this_slot->{node}->{name};
781 $round_max_slots{$node_name} ||= $this_slot->{cpu};
782 last if (scalar(keys(%round_max_slots)) >= @node);
785 Log(undef, "start level $level with $round_num_freeslots slots");
788 my $progress_is_dirty = 1;
789 my $progress_stats_updated = 0;
791 update_progress_stats();
795 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
797 # Don't create new tasks if we already know the job's final result.
798 last if defined($main::success);
800 my $id = $jobstep_todo[$todo_ptr];
801 my $Jobstep = $jobstep[$id];
802 if ($Jobstep->{level} != $level)
807 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
808 set_nonblocking($reader{$id});
810 my $childslot = $freeslot[0];
811 my $childnode = $slot[$childslot]->{node};
812 my $childslotname = join (".",
813 $slot[$childslot]->{node}->{name},
814 $slot[$childslot]->{cpu});
816 my $childpid = fork();
819 $SIG{'INT'} = 'DEFAULT';
820 $SIG{'QUIT'} = 'DEFAULT';
821 $SIG{'TERM'} = 'DEFAULT';
823 foreach (values (%reader))
827 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
828 open(STDOUT,">&writer");
829 open(STDERR,">&writer");
834 delete $ENV{"GNUPGHOME"};
835 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
836 $ENV{"TASK_QSEQUENCE"} = $id;
837 $ENV{"TASK_SEQUENCE"} = $level;
838 $ENV{"JOB_SCRIPT"} = $Job->{script};
839 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
840 $param =~ tr/a-z/A-Z/;
841 $ENV{"JOB_PARAMETER_$param"} = $value;
843 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
844 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
845 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
846 $ENV{"HOME"} = $ENV{"TASK_WORK"};
847 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
848 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
849 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
851 my $keep_mnt = $ENV{"TASK_WORK"}.".keep";
857 "--nodelist=".$childnode->{name},
858 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
859 "--job-name=$job_id.$id.$$",
862 my $stdbuf = " stdbuf --output=0 --error=0 ";
864 my $arv_file_cache = "";
865 if (defined($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'})) {
866 $arv_file_cache = "--file-cache=" . ($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'} * 1024 * 1024);
870 "if [ -e \Q$ENV{TASK_WORK}\E ]; then rm -rf \Q$ENV{TASK_WORK}\E; fi; "
871 ."mkdir -p \Q$ENV{CRUNCH_TMP}\E \Q$ENV{JOB_WORK}\E \Q$ENV{TASK_WORK}\E \Q$keep_mnt\E "
872 ."&& cd \Q$ENV{CRUNCH_TMP}\E "
873 # These environment variables get used explicitly later in
874 # $command. No tool is expected to read these values directly.
875 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
876 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
877 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
878 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP "
879 .q{&& declare -a VOLUMES=() }
880 .q{&& if which crunchrunner >/dev/null ; then VOLUMES+=("--volume=$(which crunchrunner):/usr/local/bin/crunchrunner:ro") ; fi }
881 .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") ; }
882 .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 };
884 $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 ";
885 $ENV{TASK_KEEPMOUNT} = "$keep_mnt/by_pdh";
886 $ENV{TASK_KEEPMOUNT_TMP} = "$keep_mnt/tmp";
890 my $containername = "$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}";
891 my $cidfile = "$ENV{CRUNCH_TMP}/$containername.cid";
892 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
893 $command .= "$docker_bin run $docker_run_args --name=$containername --attach=stdout --attach=stderr --attach=stdin -i \Q$dockeruserarg\E --cidfile=$cidfile --sig-proxy ";
894 # We only set memory limits if Docker lets us limit both memory and swap.
895 # Memory limits alone have been supported longer, but subprocesses tend
896 # to get SIGKILL if they exceed that without any swap limit set.
897 # See #5642 for additional background.
898 if ($docker_limitmem) {
899 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
902 # The source tree and $destdir directory (which we have
903 # installed on the worker host) are available in the container,
904 # under the same path.
905 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
906 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
908 # Currently, we make the "by_pdh" directory in arv-mount's mount
909 # point appear at /keep inside the container (instead of using
910 # the same path as the host like we do with CRUNCH_SRC and
911 # CRUNCH_INSTALL). However, crunch scripts and utilities must
912 # not rely on this. They must use $TASK_KEEPMOUNT.
913 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
914 $ENV{TASK_KEEPMOUNT} = "/keep";
916 # Ditto TASK_KEEPMOUNT_TMP, as /keep_tmp.
917 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT_TMP}:/keep_tmp\E ";
918 $ENV{TASK_KEEPMOUNT_TMP} = "/keep_tmp";
920 # TASK_WORK is almost exactly like a docker data volume: it
921 # starts out empty, is writable, and persists until no
922 # containers use it any more. We don't use --volumes-from to
923 # share it with other containers: it is only accessible to this
924 # task, and it goes away when this task stops.
926 # However, a docker data volume is writable only by root unless
927 # the mount point already happens to exist in the container with
928 # different permissions. Therefore, we [1] assume /tmp already
929 # exists in the image and is writable by the crunch user; [2]
930 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
931 # writable if they are created by docker while setting up the
932 # other --volumes); and [3] create $TASK_WORK inside the
933 # container using $build_script.
934 $command .= "--volume=/tmp ";
935 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
936 $ENV{"HOME"} = $ENV{"TASK_WORK"};
937 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
939 # TODO: Share a single JOB_WORK volume across all task
940 # containers on a given worker node, and delete it when the job
941 # ends (and, in case that doesn't work, when the next job
944 # For now, use the same approach as TASK_WORK above.
945 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
947 # Bind mount the crunchrunner binary and host TLS certificates file into
949 $command .= '"${VOLUMES[@]}" ';
951 while (my ($env_key, $env_val) = each %ENV)
953 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
954 $command .= "--env=\Q$env_key=$env_val\E ";
957 $command .= "--env=\QHOME=$ENV{HOME}\E ";
958 $command .= "\Q$docker_hash\E ";
960 if ($Job->{arvados_sdk_version}) {
962 $command .= "perl - \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E";
964 $command .= "/bin/sh -c \'python -c " .
965 '"from pkg_resources import get_distribution as get; print \"Using Arvados SDK version\", get(\"arvados-python-client\").version"' .
966 ">&2 2>/dev/null; " .
967 "mkdir -p \"$ENV{JOB_WORK}\" \"$ENV{TASK_WORK}\" && " .
968 "if which stdbuf >/dev/null ; then " .
969 " exec $stdbuf \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
971 " exec \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
976 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -poll=10000 ";
978 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
981 my @execargs = ('bash', '-c', $command);
982 srun (\@srunargs, \@execargs, undef, $build_script);
983 # exec() failed, we assume nothing happened.
984 die "srun() failed on build script\n";
987 if (!defined $childpid)
998 jobstepname => "$job_id.$id.$childpid",
1000 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
1001 $slot[$childslot]->{pid} = $childpid;
1003 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
1004 Log ($id, "child $childpid started on $childslotname");
1005 $Jobstep->{starttime} = time;
1006 $Jobstep->{node} = $childnode->{name};
1007 $Jobstep->{slotindex} = $childslot;
1008 delete $Jobstep->{stderr};
1009 delete $Jobstep->{finishtime};
1010 delete $Jobstep->{tempfail};
1012 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
1013 $Jobstep->{'arvados_task'}->save;
1015 splice @jobstep_todo, $todo_ptr, 1;
1018 $progress_is_dirty = 1;
1022 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
1024 last THISROUND if $main::please_freeze;
1025 if ($main::please_info)
1027 $main::please_info = 0;
1029 create_output_collection();
1031 update_progress_stats();
1036 if (!$gotsome || ($latest_refresh + 2 < scalar time))
1038 check_refresh_wanted();
1040 update_progress_stats();
1042 elsif (time - $progress_stats_updated >= 30 || $progress_is_dirty)
1044 update_progress_stats();
1047 select (undef, undef, undef, 0.1);
1049 $working_slot_count = scalar(grep { $_->{node}->{fail_count} == 0 &&
1050 $_->{node}->{hold_count} < 4 } @slot);
1051 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
1052 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
1054 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
1055 .($thisround_failed+$thisround_succeeded)
1056 .") -- giving up on this round";
1057 Log (undef, $message);
1061 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1062 for (my $i=$#freeslot; $i>=0; $i--) {
1063 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1064 push @holdslot, (splice @freeslot, $i, 1);
1067 for (my $i=$#holdslot; $i>=0; $i--) {
1068 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1069 push @freeslot, (splice @holdslot, $i, 1);
1073 # give up if no nodes are succeeding
1074 if ($working_slot_count < 1) {
1075 Log(undef, "Every node has failed -- giving up");
1082 push @freeslot, splice @holdslot;
1083 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1086 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1089 if ($main::please_continue) {
1090 $main::please_continue = 0;
1093 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1095 if (!reapchildren())
1097 check_refresh_wanted();
1099 update_progress_stats();
1100 select (undef, undef, undef, 0.1);
1101 killem (keys %proc) if $main::please_freeze;
1105 update_progress_stats();
1106 freeze_if_want_freeze();
1109 if (!defined $main::success)
1111 if (!@jobstep_todo) {
1113 } elsif ($working_slot_count < 1) {
1114 save_output_collection();
1116 exit(EX_RETRY_UNLOCKED);
1117 } elsif ($thisround_succeeded == 0 &&
1118 ($thisround_failed == 0 || $thisround_failed > 4)) {
1119 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1120 Log (undef, $message);
1125 goto ONELEVEL if !defined $main::success;
1128 release_allocation();
1130 my $collated_output = save_output_collection();
1131 Log (undef, "finish");
1136 if ($collated_output && $main::success) {
1137 $final_state = 'Complete';
1139 $final_state = 'Failed';
1141 $Job->update_attributes('state' => $final_state);
1143 exit (($final_state eq 'Complete') ? 0 : 1);
1147 sub update_progress_stats
1149 $progress_stats_updated = time;
1150 return if !$progress_is_dirty;
1151 my ($todo, $done, $running) = (scalar @jobstep_todo,
1152 scalar @jobstep_done,
1153 scalar keys(%proc));
1154 $Job->{'tasks_summary'} ||= {};
1155 $Job->{'tasks_summary'}->{'todo'} = $todo;
1156 $Job->{'tasks_summary'}->{'done'} = $done;
1157 $Job->{'tasks_summary'}->{'running'} = $running;
1158 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1159 Log (undef, "status: $done done, $running running, $todo todo");
1160 $progress_is_dirty = 0;
1167 my $children_reaped = 0;
1168 my @successful_task_uuids = ();
1170 while((my $pid = waitpid (-1, WNOHANG)) > 0)
1172 my $childstatus = $?;
1174 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1176 . $slot[$proc{$pid}->{slot}]->{cpu});
1177 my $jobstepidx = $proc{$pid}->{jobstepidx};
1180 my $elapsed = time - $proc{$pid}->{time};
1181 my $Jobstep = $jobstep[$jobstepidx];
1183 my $exitvalue = $childstatus >> 8;
1184 my $exitinfo = "exit ".exit_status_s($childstatus);
1185 $Jobstep->{'arvados_task'}->reload;
1186 my $task_success = $Jobstep->{'arvados_task'}->{success};
1188 Log ($jobstepidx, "child $pid on $whatslot $exitinfo success=$task_success");
1190 if (!defined $task_success) {
1191 # task did not indicate one way or the other --> fail
1192 Log($jobstepidx, sprintf(
1193 "ERROR: Task process exited %s, but never updated its task record to indicate success and record its output.",
1194 exit_status_s($childstatus)));
1195 $Jobstep->{'arvados_task'}->{success} = 0;
1196 $Jobstep->{'arvados_task'}->save;
1203 $temporary_fail ||= $Jobstep->{tempfail};
1204 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1206 ++$thisround_failed;
1207 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1209 # Check for signs of a failed or misconfigured node
1210 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1211 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1212 # Don't count this against jobstep failure thresholds if this
1213 # node is already suspected faulty and srun exited quickly
1214 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1216 Log ($jobstepidx, "blaming failure on suspect node " .
1217 $slot[$proc{$pid}->{slot}]->{node}->{name});
1218 $temporary_fail ||= 1;
1220 ban_node_by_slot($proc{$pid}->{slot});
1223 Log ($jobstepidx, sprintf('failure (#%d, %s) after %d seconds',
1224 ++$Jobstep->{'failures'},
1225 $temporary_fail ? 'temporary' : 'permanent',
1228 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1229 # Give up on this task, and the whole job
1232 # Put this task back on the todo queue
1233 push @jobstep_todo, $jobstepidx;
1234 $Job->{'tasks_summary'}->{'failed'}++;
1238 push @successful_task_uuids, $Jobstep->{'arvados_task'}->{uuid};
1239 ++$thisround_succeeded;
1240 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1241 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1242 $slot[$proc{$pid}->{slot}]->{node}->{fail_count} = 0;
1243 push @jobstep_done, $jobstepidx;
1244 Log ($jobstepidx, "success in $elapsed seconds");
1246 $Jobstep->{exitcode} = $childstatus;
1247 $Jobstep->{finishtime} = time;
1248 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1249 $Jobstep->{'arvados_task'}->save;
1250 process_stderr_final ($jobstepidx);
1251 Log ($jobstepidx, sprintf("task output (%d bytes): %s",
1252 length($Jobstep->{'arvados_task'}->{output}),
1253 $Jobstep->{'arvados_task'}->{output}));
1255 close $reader{$jobstepidx};
1256 delete $reader{$jobstepidx};
1257 delete $slot[$proc{$pid}->{slot}]->{pid};
1258 push @freeslot, $proc{$pid}->{slot};
1261 $progress_is_dirty = 1;
1264 if (scalar(@successful_task_uuids) > 0)
1266 Log (undef, sprintf("%d tasks exited (%d succeeded), checking for new tasks from API server.", $children_reaped, scalar(@successful_task_uuids)));
1268 my $newtask_list = [];
1269 my $newtask_results;
1271 $newtask_results = api_call(
1273 'filters' => [["created_by_job_task_uuid","in",\@successful_task_uuids]],
1274 'order' => 'qsequence',
1275 'offset' => scalar(@$newtask_list),
1277 push(@$newtask_list, @{$newtask_results->{items}});
1278 } while (@{$newtask_results->{items}});
1279 Log (undef, sprintf("Got %d new tasks from API server.", scalar(@$newtask_list)));
1280 foreach my $arvados_task (@$newtask_list) {
1282 'level' => $arvados_task->{'sequence'},
1284 'arvados_task' => $arvados_task
1286 push @jobstep, $jobstep;
1287 push @jobstep_todo, $#jobstep;
1291 return $children_reaped;
1294 sub check_refresh_wanted
1296 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1298 $stat[9] > $latest_refresh &&
1299 # ...and we have actually locked the job record...
1300 $job_id eq $Job->{'uuid'}) {
1301 $latest_refresh = scalar time;
1302 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1303 for my $attr ('cancelled_at',
1304 'cancelled_by_user_uuid',
1305 'cancelled_by_client_uuid',
1307 $Job->{$attr} = $Job2->{$attr};
1309 if ($Job->{'state'} ne "Running") {
1310 if ($Job->{'state'} eq "Cancelled") {
1311 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1313 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1316 $main::please_freeze = 1;
1323 my $last_squeue_check = $squeue_checked;
1325 # Do not call `squeue` or check the kill list more than once every
1327 return if $last_squeue_check > time - 15;
1328 $squeue_checked = time;
1330 # Look for children from which we haven't received stderr data since
1331 # the last squeue check. If no such children exist, all procs are
1332 # alive and there's no need to even look at squeue.
1334 # As long as the crunchstat poll interval (10s) is shorter than the
1335 # squeue check interval (15s) this should make the squeue check an
1337 my $silent_procs = 0;
1338 for my $js (map {$jobstep[$_->{jobstepidx}]} values %proc)
1340 if (!exists($js->{stderr_at}))
1342 $js->{stderr_at} = 0;
1344 if ($js->{stderr_at} < $last_squeue_check)
1349 return if $silent_procs == 0;
1351 # use killem() on procs whose killtime is reached
1352 while (my ($pid, $procinfo) = each %proc)
1354 my $js = $jobstep[$procinfo->{jobstepidx}];
1355 if (exists $procinfo->{killtime}
1356 && $procinfo->{killtime} <= time
1357 && $js->{stderr_at} < $last_squeue_check)
1360 if ($js->{stderr_at}) {
1361 $sincewhen = " in last " . (time - $js->{stderr_at}) . "s";
1363 Log($procinfo->{jobstepidx}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
1370 # here is an opportunity to check for mysterious problems with local procs
1374 # Get a list of steps still running. Note: squeue(1) says --steps
1375 # selects a format (which we override anyway) and allows us to
1376 # specify which steps we're interested in (which we don't).
1377 # Importantly, it also changes the meaning of %j from "job name" to
1378 # "step name" and (although this isn't mentioned explicitly in the
1379 # docs) switches from "one line per job" mode to "one line per step"
1380 # mode. Without it, we'd just get a list of one job, instead of a
1382 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%j' --noheader`;
1385 Log(undef, "warning: squeue exit status $? ($!)");
1390 # which of my jobsteps are running, according to squeue?
1392 for my $jobstepname (@squeue)
1394 $ok{$jobstepname} = 1;
1397 # Check for child procs >60s old and not mentioned by squeue.
1398 while (my ($pid, $procinfo) = each %proc)
1400 if ($procinfo->{time} < time - 60
1401 && $procinfo->{jobstepname}
1402 && !exists $ok{$procinfo->{jobstepname}}
1403 && !exists $procinfo->{killtime})
1405 # According to slurm, this task has ended (successfully or not)
1406 # -- but our srun child hasn't exited. First we must wait (30
1407 # seconds) in case this is just a race between communication
1408 # channels. Then, if our srun child process still hasn't
1409 # terminated, we'll conclude some slurm communication
1410 # error/delay has caused the task to die without notifying srun,
1411 # and we'll kill srun ourselves.
1412 $procinfo->{killtime} = time + 30;
1413 Log($procinfo->{jobstepidx}, "notice: task is not in slurm queue but srun process $pid has not exited");
1420 # If a node fails in a multi-node "srun" call during job setup, the call
1421 # may hang instead of exiting with a nonzero code. This function checks
1422 # "sinfo" for the health of the nodes that were allocated and ensures that
1423 # they are all still in the "alloc" state. If a node that is allocated to
1424 # this job is not in "alloc" state, then set please_freeze.
1426 # This is only called from srun_sync() for node configuration. If a
1427 # node fails doing actual work, there are other recovery mechanisms.
1429 # Do not call `sinfo` more than once every 15 seconds.
1430 return if $sinfo_checked > time - 15;
1431 $sinfo_checked = time;
1433 # The output format "%t" means output node states.
1434 my @sinfo = `sinfo --nodes=\Q$ENV{SLURM_NODELIST}\E --noheader -o "%t"`;
1437 Log(undef, "warning: sinfo exit status $? ($!)");
1444 if ($_ != "alloc" && $_ != "alloc*") {
1445 $main::please_freeze = 1;
1450 sub release_allocation
1454 Log (undef, "release job allocation");
1455 system "scancel $ENV{SLURM_JOB_ID}";
1464 my $sel = IO::Select->new();
1465 foreach my $jobstepidx (keys %reader)
1467 my $fd = $reader{$jobstepidx};
1469 $fd_job{$fd} = $jobstepidx;
1471 if (my $stdout_fd = $jobstep[$jobstepidx]->{stdout_r}) {
1472 $sel->add($stdout_fd);
1473 $fd_job{$stdout_fd} = $jobstepidx;
1476 # select on all reader fds with 0.1s timeout
1477 my @ready_fds = $sel->can_read(0.1);
1478 foreach my $fd (@ready_fds)
1481 if (0 < sysread ($fd, $buf, 65536))
1484 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1486 my $jobstepidx = $fd_job{$fd};
1487 if ($jobstep[$jobstepidx]->{stdout_r} == $fd) {
1488 $jobstep[$jobstepidx]->{stdout_captured} .= $buf;
1492 $jobstep[$jobstepidx]->{stderr_at} = time;
1493 $jobstep[$jobstepidx]->{stderr} .= $buf;
1495 # Consume everything up to the last \n
1496 preprocess_stderr ($jobstepidx);
1498 if (length ($jobstep[$jobstepidx]->{stderr}) > 16384)
1500 # If we get a lot of stderr without a newline, chop off the
1501 # front to avoid letting our buffer grow indefinitely.
1502 substr ($jobstep[$jobstepidx]->{stderr},
1503 0, length($jobstep[$jobstepidx]->{stderr}) - 8192) = "";
1511 # Consume all full lines of stderr for a jobstep. Everything after the
1512 # last newline will remain in $jobstep[$jobstepidx]->{stderr} after
1514 sub preprocess_stderr
1516 my $jobstepidx = shift;
1517 # slotindex is only defined for children running Arvados job tasks.
1518 # Be prepared to handle the undef case (for setup srun calls, etc.).
1519 my $job_slot_index = $jobstep[$jobstepidx]->{slotindex};
1521 while ($jobstep[$jobstepidx]->{stderr} =~ /^(.*?)\n/) {
1523 substr $jobstep[$jobstepidx]->{stderr}, 0, 1+length($line), "";
1524 Log ($jobstepidx, "stderr $line");
1525 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/i) {
1526 # If the allocation is revoked, we can't possibly continue, so mark all
1527 # nodes as failed. This will cause the overall exit code to be
1528 # EX_RETRY_UNLOCKED instead of failure so that crunch_dispatch can re-run
1530 $main::please_freeze = 1;
1531 foreach my $st (@slot) {
1532 $st->{node}->{fail_count}++;
1535 elsif ($line =~ /srun: error: .*?\b(Node failure on|Aborting, .*?\bio error\b)/i) {
1536 $jobstep[$jobstepidx]->{tempfail} = 1;
1537 if (defined($job_slot_index)) {
1538 $slot[$job_slot_index]->{node}->{fail_count}++;
1539 ban_node_by_slot($job_slot_index);
1542 elsif ($line =~ /srun: error: (Unable to create job step|.*?: Communication connection failure)/i) {
1543 $jobstep[$jobstepidx]->{tempfail} = 1;
1544 ban_node_by_slot($job_slot_index) if (defined($job_slot_index));
1546 elsif ($line =~ /\bKeep(Read|Write|Request)Error:/) {
1547 $jobstep[$jobstepidx]->{tempfail} = 1;
1553 sub process_stderr_final
1555 my $jobstepidx = shift;
1556 preprocess_stderr ($jobstepidx);
1559 Log ($jobstepidx, "stderr $_");
1560 } split ("\n", $jobstep[$jobstepidx]->{stderr});
1561 $jobstep[$jobstepidx]->{stderr} = '';
1568 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1569 Log(undef, "fetch_block run error from arv-get $hash: $!");
1572 my $output_block = "";
1575 my $bytes = sysread($keep, $buf, 1024 * 1024);
1576 if (!defined $bytes) {
1577 Log(undef, "fetch_block read error from arv-get: $!");
1578 $output_block = undef;
1580 } elsif ($bytes == 0) {
1581 # sysread returns 0 at the end of the pipe.
1584 # some bytes were read into buf.
1585 $output_block .= $buf;
1590 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1591 $output_block = undef;
1593 return $output_block;
1596 # Create a collection by concatenating the output of all tasks (each
1597 # task's output is either a manifest fragment, a locator for a
1598 # manifest fragment stored in Keep, or nothing at all). Return the
1599 # portable_data_hash of the new collection.
1600 sub create_output_collection
1602 Log (undef, "collate");
1604 my ($child_out, $child_in);
1605 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1608 print (arvados.api("v1").collections().
1609 create(body={"manifest_text": sys.stdin.read(),
1610 "owner_uuid": sys.argv[2]}).
1611 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1612 }, retry_count(), $Job->{owner_uuid});
1615 my $manifest_size = 0;
1619 my $output = $_->{'arvados_task'}->{output};
1620 next if (!defined($output));
1622 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1623 $next_write = fetch_block($output);
1625 $next_write = $output;
1627 if (defined($next_write)) {
1628 if (!defined(syswrite($child_in, $next_write))) {
1629 # There's been an error writing. Stop the loop.
1630 # We'll log details about the exit code later.
1633 $manifest_size += length($next_write);
1636 my $uuid = $_->{'arvados_task'}->{'uuid'};
1637 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1642 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1645 my $s = IO::Select->new($child_out);
1646 if ($s->can_read(120)) {
1647 sysread($child_out, $joboutput, 1024 * 1024);
1650 Log(undef, "output collection creation exited " . exit_status_s($?));
1656 Log (undef, "timed out while creating output collection");
1657 foreach my $signal (2, 2, 2, 15, 15, 9) {
1658 kill($signal, $pid);
1659 last if waitpid($pid, WNOHANG) == -1;
1668 # Calls create_output_collection, logs the result, and returns it.
1669 # If that was successful, save that as the output in the job record.
1670 sub save_output_collection {
1671 my $collated_output = create_output_collection();
1673 if (!$collated_output) {
1674 Log(undef, "Failed to write output collection");
1677 Log(undef, "job output $collated_output");
1678 $Job->update_attributes('output' => $collated_output);
1680 return $collated_output;
1687 my $sig = 2; # SIGINT first
1688 if (exists $proc{$_}->{"sent_$sig"} &&
1689 time - $proc{$_}->{"sent_$sig"} > 4)
1691 $sig = 15; # SIGTERM if SIGINT doesn't work
1693 if (exists $proc{$_}->{"sent_$sig"} &&
1694 time - $proc{$_}->{"sent_$sig"} > 4)
1696 $sig = 9; # SIGKILL if SIGTERM doesn't work
1698 if (!exists $proc{$_}->{"sent_$sig"})
1700 Log ($proc{$_}->{jobstepidx}, "sending 2x signal $sig to pid $_");
1702 select (undef, undef, undef, 0.1);
1705 kill $sig, $_; # srun wants two SIGINT to really interrupt
1707 $proc{$_}->{"sent_$sig"} = time;
1708 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1718 vec($bits,fileno($_),1) = 1;
1724 # Send log output to Keep via arv-put.
1726 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1727 # $log_pipe_out_buf is a string containing all output read from arv-put so far.
1728 # $log_pipe_out_select is an IO::Select object around $log_pipe_out.
1729 # $log_pipe_pid is the pid of the arv-put subprocess.
1731 # The only functions that should access these variables directly are:
1733 # log_writer_start($logfilename)
1734 # Starts an arv-put pipe, reading data on stdin and writing it to
1735 # a $logfilename file in an output collection.
1737 # log_writer_read_output([$timeout])
1738 # Read output from $log_pipe_out and append it to $log_pipe_out_buf.
1739 # Passes $timeout to the select() call, with a default of 0.01.
1740 # Returns the result of the last read() call on $log_pipe_out, or
1741 # -1 if read() wasn't called because select() timed out.
1742 # Only other log_writer_* functions should need to call this.
1744 # log_writer_send($txt)
1745 # Writes $txt to the output log collection.
1747 # log_writer_finish()
1748 # Closes the arv-put pipe and returns the output that it produces.
1750 # log_writer_is_active()
1751 # Returns a true value if there is currently a live arv-put
1752 # process, false otherwise.
1754 my ($log_pipe_in, $log_pipe_out, $log_pipe_out_buf, $log_pipe_out_select,
1757 sub log_writer_start($)
1759 my $logfilename = shift;
1760 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1764 '--filename', $logfilename,
1766 $log_pipe_out_buf = "";
1767 $log_pipe_out_select = IO::Select->new($log_pipe_out);
1770 sub log_writer_read_output {
1771 my $timeout = shift || 0.01;
1773 while ($read && $log_pipe_out_select->can_read($timeout)) {
1774 $read = read($log_pipe_out, $log_pipe_out_buf, 65536,
1775 length($log_pipe_out_buf));
1777 if (!defined($read)) {
1778 Log(undef, "error reading log manifest from arv-put: $!");
1783 sub log_writer_send($)
1786 print $log_pipe_in $txt;
1787 log_writer_read_output();
1790 sub log_writer_finish()
1792 return unless $log_pipe_pid;
1794 close($log_pipe_in);
1796 my $logger_failed = 0;
1797 my $read_result = log_writer_read_output(600);
1798 if ($read_result == -1) {
1799 $logger_failed = -1;
1800 Log (undef, "timed out reading from 'arv-put'");
1801 } elsif ($read_result != 0) {
1802 $logger_failed = -2;
1803 Log(undef, "failed to read arv-put log manifest to EOF");
1806 waitpid($log_pipe_pid, 0);
1808 $logger_failed ||= $?;
1809 Log(undef, "log_writer_finish: arv-put exited " . exit_status_s($?))
1812 close($log_pipe_out);
1813 my $arv_put_output = $logger_failed ? undef : $log_pipe_out_buf;
1814 $log_pipe_pid = $log_pipe_in = $log_pipe_out = $log_pipe_out_buf =
1815 $log_pipe_out_select = undef;
1817 return $arv_put_output;
1820 sub log_writer_is_active() {
1821 return $log_pipe_pid;
1824 sub Log # ($jobstepidx, $logmessage)
1826 my ($jobstepidx, $logmessage) = @_;
1827 if ($logmessage =~ /\n/) {
1828 for my $line (split (/\n/, $_[1])) {
1829 Log ($jobstepidx, $line);
1833 my $fh = select STDERR; $|=1; select $fh;
1835 if (defined($jobstepidx) && exists($jobstep[$jobstepidx]->{arvados_task})) {
1836 $task_qseq = $jobstepidx;
1838 my $message = sprintf ("%s %d %s %s", $job_id, $$, $task_qseq, $logmessage);
1839 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1842 if (log_writer_is_active() || -t STDERR) {
1843 my @gmtime = gmtime;
1844 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1845 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1847 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1849 if (log_writer_is_active()) {
1850 log_writer_send($datetime . " " . $message);
1857 my ($package, $file, $line) = caller;
1858 my $message = "@_ at $file line $line\n";
1859 Log (undef, $message);
1860 release_allocation();
1861 freeze() if @jobstep_todo;
1862 create_output_collection() if @jobstep_todo;
1872 if ($Job->{'state'} eq 'Cancelled') {
1873 $Job->update_attributes('finished_at' => scalar gmtime);
1875 $Job->update_attributes('state' => 'Failed');
1882 my $justcheckpoint = shift; # false if this will be the last meta saved
1883 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1884 return unless log_writer_is_active();
1885 my $log_manifest = log_writer_finish();
1886 return unless defined($log_manifest);
1889 my $prev_log_coll = api_call("collections/get", uuid => $Job->{log});
1890 $log_manifest = $prev_log_coll->{manifest_text} . $log_manifest;
1893 my $log_coll = api_call(
1894 "collections/create", ensure_unique_name => 1, collection => {
1895 manifest_text => $log_manifest,
1896 owner_uuid => $Job->{owner_uuid},
1897 name => sprintf("Log from %s job %s", $Job->{script}, $Job->{uuid}),
1899 Log(undef, "log collection is " . $log_coll->{portable_data_hash});
1900 $Job->update_attributes('log' => $log_coll->{portable_data_hash});
1904 sub freeze_if_want_freeze
1906 if ($main::please_freeze)
1908 release_allocation();
1911 # kill some srun procs before freeze+stop
1912 map { $proc{$_} = {} } @_;
1915 killem (keys %proc);
1916 select (undef, undef, undef, 0.1);
1918 while (($died = waitpid (-1, WNOHANG)) > 0)
1920 delete $proc{$died};
1925 create_output_collection();
1935 Log (undef, "Freeze not implemented");
1942 croak ("Thaw not implemented");
1958 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1964 my $srunargs = shift;
1965 my $execargs = shift;
1966 my $opts = shift || {};
1969 my $label = exists $opts->{label} ? $opts->{label} : "@$execargs";
1970 Log (undef, "$label: start");
1972 my ($stderr_r, $stderr_w);
1973 pipe $stderr_r, $stderr_w or croak("pipe() failed: $!");
1975 my ($stdout_r, $stdout_w);
1976 pipe $stdout_r, $stdout_w or croak("pipe() failed: $!");
1978 my $srunpid = fork();
1983 fcntl($stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
1984 fcntl($stdout_w, F_SETFL, 0) or croak($!);
1985 open(STDERR, ">&", $stderr_w);
1986 open(STDOUT, ">&", $stdout_w);
1987 srun ($srunargs, $execargs, $opts, $stdin);
1993 set_nonblocking($stderr_r);
1994 set_nonblocking($stdout_r);
1996 # Add entries to @jobstep and %proc so check_squeue() and
1997 # freeze_if_want_freeze() can treat it like a job task process.
2001 stderr_captured => '',
2002 stdout_r => $stdout_r,
2003 stdout_captured => '',
2005 my $jobstepidx = $#jobstep;
2007 jobstepidx => $jobstepidx,
2009 $reader{$jobstepidx} = $stderr_r;
2011 while ($srunpid != waitpid ($srunpid, WNOHANG)) {
2012 my $busy = readfrompipes();
2013 if (!$busy || ($latest_refresh + 2 < scalar time)) {
2014 check_refresh_wanted();
2019 select(undef, undef, undef, 0.1);
2021 killem(keys %proc) if $main::please_freeze;
2025 1 while readfrompipes();
2026 process_stderr_final ($jobstepidx);
2028 Log (undef, "$label: exit ".exit_status_s($exited));
2032 delete $proc{$srunpid};
2033 delete $reader{$jobstepidx};
2035 my $j = pop @jobstep;
2036 # If the srun showed signs of tempfail, ensure the caller treats that as a
2038 if ($main::please_freeze || $j->{tempfail}) {
2041 return ($exited, $j->{stdout_captured}, $j->{stderr_captured});
2047 my $srunargs = shift;
2048 my $execargs = shift;
2049 my $opts = shift || {};
2051 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
2053 $Data::Dumper::Terse = 1;
2054 $Data::Dumper::Indent = 0;
2055 my $show_cmd = Dumper($args);
2056 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
2057 $show_cmd =~ s/\n/ /g;
2058 if ($opts->{fork}) {
2059 Log(undef, "starting: $show_cmd");
2061 # This is a child process: parent is in charge of reading our
2062 # stderr and copying it to Log() if needed.
2063 warn "starting: $show_cmd\n";
2066 if (defined $stdin) {
2067 my $child = open STDIN, "-|";
2068 defined $child or die "no fork: $!";
2070 print $stdin or die $!;
2071 close STDOUT or die $!;
2076 return system (@$args) if $opts->{fork};
2079 warn "ENV size is ".length(join(" ",%ENV));
2080 die "exec failed: $!: @$args";
2084 sub ban_node_by_slot {
2085 # Don't start any new jobsteps on this node for 60 seconds
2087 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
2088 $slot[$slotid]->{node}->{hold_count}++;
2089 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
2094 my ($lockfile, $error_message) = @_;
2095 open L, ">", $lockfile or croak("$lockfile: $!");
2096 if (!flock L, LOCK_EX|LOCK_NB) {
2097 croak("Can't lock $lockfile: $error_message\n");
2101 sub find_docker_image {
2102 # Given a Keep locator, check to see if it contains a Docker image.
2103 # If so, return its stream name and Docker hash.
2104 # If not, return undef for both values.
2105 my $locator = shift;
2106 my ($streamname, $filename);
2107 my $image = api_call("collections/get", uuid => $locator);
2109 foreach my $line (split(/\n/, $image->{manifest_text})) {
2110 my @tokens = split(/\s+/, $line);
2112 $streamname = shift(@tokens);
2113 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
2114 if (defined($filename)) {
2115 return (undef, undef); # More than one file in the Collection.
2117 $filename = (split(/:/, $filedata, 3))[2];
2122 if (defined($filename) and ($filename =~ /^((?:sha256:)?[0-9A-Fa-f]{64})\.tar$/)) {
2123 return ($streamname, $1);
2125 return (undef, undef);
2130 # Calculate the number of times an operation should be retried,
2131 # assuming exponential backoff, and that we're willing to retry as
2132 # long as tasks have been running. Enforce a minimum of 3 retries.
2133 my ($starttime, $endtime, $timediff, $retries);
2135 $starttime = $jobstep[0]->{starttime};
2136 $endtime = $jobstep[-1]->{finishtime};
2138 if (!defined($starttime)) {
2140 } elsif (!defined($endtime)) {
2141 $timediff = time - $starttime;
2143 $timediff = ($endtime - $starttime) - (time - $endtime);
2145 if ($timediff > 0) {
2146 $retries = int(log($timediff) / log(2));
2148 $retries = 1; # Use the minimum.
2150 return ($retries > 3) ? $retries : 3;
2154 # Pass in two function references.
2155 # This method will be called with the remaining arguments.
2156 # If it dies, retry it with exponential backoff until it succeeds,
2157 # or until the current retry_count is exhausted. After each failure
2158 # that can be retried, the second function will be called with
2159 # the current try count (0-based), next try time, and error message.
2160 my $operation = shift;
2161 my $retry_callback = shift;
2162 my $retries = retry_count();
2163 foreach my $try_count (0..$retries) {
2164 my $next_try = time + (2 ** $try_count);
2165 my $result = eval { $operation->(@_); };
2168 } elsif ($try_count < $retries) {
2169 $retry_callback->($try_count, $next_try, $@);
2170 my $sleep_time = $next_try - time;
2171 sleep($sleep_time) if ($sleep_time > 0);
2174 # Ensure the error message ends in a newline, so Perl doesn't add
2175 # retry_op's line number to it.
2181 # Pass in a /-separated API method name, and arguments for it.
2182 # This function will call that method, retrying as needed until
2183 # the current retry_count is exhausted, with a log on the first failure.
2184 my $method_name = shift;
2185 my $log_api_retry = sub {
2186 my ($try_count, $next_try_at, $errmsg) = @_;
2187 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
2188 $errmsg =~ s/\s/ /g;
2189 $errmsg =~ s/\s+$//;
2191 if ($next_try_at < time) {
2192 $retry_msg = "Retrying.";
2194 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
2195 $retry_msg = "Retrying at $next_try_fmt.";
2197 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
2200 foreach my $key (split(/\//, $method_name)) {
2201 $method = $method->{$key};
2203 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
2207 # Given a $?, return a human-readable exit code string like "0" or
2208 # "1" or "0 with signal 1" or "1 with signal 11".
2209 my $exitcode = shift;
2210 my $s = $exitcode >> 8;
2211 if ($exitcode & 0x7f) {
2212 $s .= " with signal " . ($exitcode & 0x7f);
2214 if ($exitcode & 0x80) {
2215 $s .= " with core dump";
2220 sub handle_readall {
2221 # Pass in a glob reference to a file handle.
2222 # Read all its contents and return them as a string.
2223 my $fh_glob_ref = shift;
2225 return <$fh_glob_ref>;
2228 sub tar_filename_n {
2230 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
2233 sub add_git_archive {
2234 # Pass in a git archive command as a string or list, a la system().
2235 # This method will save its output to be included in the archive sent to the
2239 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
2240 croak("Failed to save git archive: $!");
2242 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
2244 waitpid($git_pid, 0);
2247 croak("Failed to save git archive: git exited " . exit_status_s($?));
2251 sub combined_git_archive {
2252 # Combine all saved tar archives into a single archive, then return its
2253 # contents in a string. Return undef if no archives have been saved.
2254 if ($git_tar_count < 1) {
2257 my $base_tar_name = tar_filename_n(1);
2258 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
2259 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
2260 if ($tar_exit != 0) {
2261 croak("Error preparing build archive: tar -A exited " .
2262 exit_status_s($tar_exit));
2265 if (!open(GIT_TAR, "<", $base_tar_name)) {
2266 croak("Could not open build archive: $!");
2268 my $tar_contents = handle_readall(\*GIT_TAR);
2270 return $tar_contents;
2273 sub set_nonblocking {
2275 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
2276 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
2282 # This is crunch-job's internal dispatch script. crunch-job running on the API
2283 # server invokes this script on individual compute nodes, or localhost if we're
2284 # running a job locally. It gets called in two modes:
2286 # * No arguments: Installation mode. Read a tar archive from the DATA
2287 # file handle; it includes the Crunch script's source code, and
2288 # maybe SDKs as well. Those should be installed in the proper
2289 # locations. This runs outside of any Docker container, so don't try to
2290 # introspect Crunch's runtime environment.
2292 # * With arguments: Crunch script run mode. This script should set up the
2293 # environment, then run the command specified in the arguments. This runs
2294 # inside any Docker container.
2297 use File::Path qw( make_path remove_tree );
2298 use POSIX qw(getcwd);
2300 use constant TASK_TEMPFAIL => 111;
2302 # Map SDK subdirectories to the path environments they belong to.
2303 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2305 my $destdir = $ENV{"CRUNCH_SRC"};
2306 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2307 my $repo = $ENV{"CRUNCH_SRC_URL"};
2308 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2309 my $job_work = $ENV{"JOB_WORK"};
2310 my $task_work = $ENV{"TASK_WORK"};
2312 open(STDOUT_ORIG, ">&", STDOUT);
2313 open(STDERR_ORIG, ">&", STDERR);
2315 for my $dir ($destdir, $job_work, $task_work) {
2318 -e $dir or die "Failed to create temporary directory ($dir): $!";
2323 remove_tree($task_work, {keep_root => 1});
2326 ### Crunch script run mode
2328 # We want to do routine logging during task 0 only. This gives the user
2329 # the information they need, but avoids repeating the information for every
2332 if ($ENV{TASK_SEQUENCE} eq "0") {
2335 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2341 my $python_src = "$install_dir/python";
2342 my $venv_dir = "$job_work/.arvados.venv";
2343 my $venv_built = -e "$venv_dir/bin/activate";
2344 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2345 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2346 "--python=python2.7", $venv_dir);
2347 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2349 $Log->("Built Python SDK virtualenv");
2352 my @pysdk_version_cmd = ("python", "-c",
2353 "from pkg_resources import get_distribution as get; print get('arvados-python-client').version");
2355 $Log->("Running in Python SDK virtualenv");
2356 @pysdk_version_cmd = ();
2357 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2358 @ARGV = ("/bin/sh", "-ec",
2359 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2360 } elsif (-d $python_src) {
2361 $Log->("Warning: virtualenv not found inside Docker container default " .
2362 "\$PATH. Can't install Python SDK.");
2365 if (@pysdk_version_cmd) {
2366 open(my $pysdk_version_pipe, "-|", @pysdk_version_cmd);
2367 my $pysdk_version = <$pysdk_version_pipe>;
2368 close($pysdk_version_pipe);
2370 chomp($pysdk_version);
2371 $Log->("Using Arvados SDK version $pysdk_version");
2373 # A lot could've gone wrong here, but pretty much all of it means that
2374 # Python won't be able to load the Arvados SDK.
2375 $Log->("Warning: Arvados SDK not found");
2379 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2380 my $sdk_path = "$install_dir/$sdk_dir";
2382 if ($ENV{$sdk_envkey}) {
2383 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2385 $ENV{$sdk_envkey} = $sdk_path;
2387 $Log->("Arvados SDK added to %s", $sdk_envkey);
2392 die "Cannot exec `@ARGV`: $!";
2395 ### Installation mode
2396 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2398 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2399 # This exact git archive (source + arvados sdk) is already installed
2400 # here, so there's no need to reinstall it.
2402 # We must consume our DATA section, though: otherwise the process
2403 # feeding it to us will get SIGPIPE.
2405 while (read(DATA, $buf, 65536)) { }
2410 unlink "$destdir.archive_hash";
2414 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2415 local $SIG{PIPE} = "IGNORE";
2416 warn "Extracting archive: $archive_hash\n";
2417 # --ignore-zeros is necessary sometimes: depending on how much NUL
2418 # padding tar -A put on our combined archive (which in turn depends
2419 # on the length of the component archives) tar without
2420 # --ignore-zeros will exit before consuming stdin and cause close()
2421 # to fail on the resulting SIGPIPE.
2422 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2423 die "Error launching 'tar -xC $destdir': $!";
2425 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2426 # get SIGPIPE. We must feed it data incrementally.
2428 while (read(DATA, $tar_input, 65536)) {
2429 print TARX $tar_input;
2432 die "'tar -xC $destdir' exited $?: $!";
2438 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2440 foreach my $sdk_lang (("python",
2441 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2442 if (-d "$sdk_root/$sdk_lang") {
2443 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2444 die "Failed to install $sdk_lang SDK: $!";
2450 my $python_dir = "$install_dir/python";
2451 if ((-d $python_dir) and can_run("python2.7")) {
2452 open(my $egg_info_pipe, "-|",
2453 "python2.7 \Q$python_dir/setup.py\E egg_info 2>&1 >/dev/null");
2454 my @egg_info_errors = <$egg_info_pipe>;
2455 close($egg_info_pipe);
2458 if (@egg_info_errors and (($egg_info_errors[-1] =~ /\bgit\b/) or ($egg_info_errors[-1] =~ /\[Errno 2\]/))) {
2459 # egg_info apparently failed because it couldn't ask git for a build tag.
2460 # Specify no build tag.
2461 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2462 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2465 my $egg_info_exit = $? >> 8;
2466 foreach my $errline (@egg_info_errors) {
2469 warn "python setup.py egg_info failed: exit $egg_info_exit";
2470 exit ($egg_info_exit || 1);
2475 # Hide messages from the install script (unless it fails: shell_or_die
2476 # will show $destdir.log in that case).
2477 open(STDOUT, ">>", "$destdir.log");
2478 open(STDERR, ">&", STDOUT);
2480 if (-e "$destdir/crunch_scripts/install") {
2481 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2482 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2484 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2485 } elsif (-e "./install.sh") {
2486 shell_or_die (undef, "./install.sh", $install_dir);
2489 if ($archive_hash) {
2490 unlink "$destdir.archive_hash.new";
2491 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2492 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2498 my $command_name = shift;
2499 open(my $which, "-|", "which", $command_name);
2500 while (<$which>) { }
2507 my $exitcode = shift;
2509 if ($ENV{"DEBUG"}) {
2510 print STDERR "@_\n";
2512 if (system (@_) != 0) {
2515 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2516 open STDERR, ">&STDERR_ORIG";
2517 system ("cat $destdir.log >&2");
2518 warn "@_ failed ($err): $exitstatus";
2519 if (defined($exitcode)) {
2523 exit (($code >> 8) || 1);