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 $latest_refresh = scalar time;
362 if (defined $Job->{thawedfromkey})
364 thaw ($Job->{thawedfromkey});
368 my $first_task = api_call("job_tasks/create", job_task => {
369 'job_uuid' => $Job->{'uuid'},
374 push @jobstep, { 'level' => 0,
376 'arvados_task' => $first_task,
378 push @jobstep_todo, 0;
384 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
387 my $build_script = handle_readall(\*DATA);
388 my $nodelist = join(",", @node);
389 my $git_tar_count = 0;
391 if (!defined $no_clear_tmp) {
392 # Find FUSE mounts under $CRUNCH_TMP and unmount them. Then clean
393 # up work directories crunch_tmp/work, crunch_tmp/opt,
396 # TODO: When #5036 is done and widely deployed, we can limit mount's
397 # -t option to simply fuse.keep.
398 my ($exited, $stdout, $stderr) = srun_sync(
399 ["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
400 ['bash', '-ec', '-o', 'pipefail', 'mount -t fuse,fuse.keep | awk "(index(\$3, \"$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* $CRUNCH_TMP/*.cid'],
401 {label => "clean work dirs"});
403 exit(EX_RETRY_UNLOCKED);
407 # If this job requires a Docker image, install that.
408 my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem, $dockeruserarg);
409 if ($docker_locator = $Job->{docker_image_locator}) {
410 Log (undef, "Install docker image $docker_locator");
411 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
414 croak("No Docker image hash found from locator $docker_locator");
416 Log (undef, "docker image hash is $docker_hash");
417 $docker_stream =~ s/^\.//;
418 my $docker_install_script = qq{
419 if $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
422 declare -a exit_codes=("\${PIPESTATUS[@]}")
423 if [ 0 != "\${exit_codes[0]}" ]; then
424 exit "\${exit_codes[0]}" # `docker images` failed
425 elif [ 1 != "\${exit_codes[1]}" ]; then
426 exit "\${exit_codes[1]}" # `grep` encountered an error
428 # Everything worked fine, but grep didn't find the image on this host.
429 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
433 my ($exited, $stdout, $stderr) = srun_sync(
434 ["srun", "--nodelist=" . join(',', @node)],
435 ["/bin/bash", "-o", "pipefail", "-ec", $docker_install_script],
436 {label => "load docker image"});
439 exit(EX_RETRY_UNLOCKED);
442 # Determine whether this version of Docker supports memory+swap limits.
443 ($exited, $stdout, $stderr) = srun_sync(
444 ["srun", "--nodes=1"],
445 [$docker_bin, 'run', '--help'],
446 {label => "check --memory-swap feature"});
447 $docker_limitmem = ($stdout =~ /--memory-swap/);
449 # Find a non-root Docker user to use.
450 # Tries the default user for the container, then 'crunch', then 'nobody',
451 # testing for whether the actual user id is non-zero. This defends against
452 # mistakes but not malice, but we intend to harden the security in the future
453 # so we don't want anyone getting used to their jobs running as root in their
455 my @tryusers = ("", "crunch", "nobody");
456 foreach my $try_user (@tryusers) {
459 if ($try_user eq "") {
460 $label = "check whether default user is UID 0";
463 $label = "check whether user '$try_user' is UID 0";
464 $try_user_arg = "--user=$try_user";
466 my ($exited, $stdout, $stderr) = srun_sync(
467 ["srun", "--nodes=1"],
469 "$docker_bin run $docker_run_args $try_user_arg $docker_hash id --user"],
472 if ($exited == 0 && $stdout =~ /^\d+$/ && $stdout > 0) {
473 $dockeruserarg = $try_user_arg;
474 if ($try_user eq "") {
475 Log(undef, "Container will run with default user");
477 Log(undef, "Container will run with $dockeruserarg");
483 if (!defined $dockeruserarg) {
484 croak("Could not find a user in container that is not UID 0 (tried default user, @tryusers) or there was a problem running 'id' in the container.");
487 if ($Job->{arvados_sdk_version}) {
488 # The job also specifies an Arvados SDK version. Add the SDKs to the
489 # tar file for the build script to install.
490 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
491 $Job->{arvados_sdk_version}));
492 add_git_archive("git", "--git-dir=$git_dir", "archive",
493 "--prefix=.arvados.sdk/",
494 $Job->{arvados_sdk_version}, "sdk");
498 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
499 # If script_version looks like an absolute path, *and* the --git-dir
500 # argument was not given -- which implies we were not invoked by
501 # crunch-dispatch -- we will use the given path as a working
502 # directory instead of resolving script_version to a git commit (or
503 # doing anything else with git).
504 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
505 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
508 # Resolve the given script_version to a git commit sha1. Also, if
509 # the repository is remote, clone it into our local filesystem: this
510 # ensures "git archive" will work, and is necessary to reliably
511 # resolve a symbolic script_version like "master^".
512 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
514 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
516 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
518 # If we're running under crunch-dispatch, it will have already
519 # pulled the appropriate source tree into its own repository, and
520 # given us that repo's path as $git_dir.
522 # If we're running a "local" job, we might have to fetch content
523 # from a remote repository.
525 # (Currently crunch-dispatch gives a local path with --git-dir, but
526 # we might as well accept URLs there too in case it changes its
528 my $repo = $git_dir || $Job->{'repository'};
530 # Repository can be remote or local. If remote, we'll need to fetch it
531 # to a local dir before doing `git log` et al.
534 if ($repo =~ m{://|^[^/]*:}) {
535 # $repo is a git url we can clone, like git:// or https:// or
536 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
537 # not recognized here because distinguishing that from a local
538 # path is too fragile. If you really need something strange here,
539 # use the ssh:// form.
540 $repo_location = 'remote';
541 } elsif ($repo =~ m{^\.*/}) {
542 # $repo is a local path to a git index. We'll also resolve ../foo
543 # to ../foo/.git if the latter is a directory. To help
544 # disambiguate local paths from named hosted repositories, this
545 # form must be given as ./ or ../ if it's a relative path.
546 if (-d "$repo/.git") {
547 $repo = "$repo/.git";
549 $repo_location = 'local';
551 # $repo is none of the above. It must be the name of a hosted
553 my $arv_repo_list = api_call("repositories/list",
554 'filters' => [['name','=',$repo]]);
555 my @repos_found = @{$arv_repo_list->{'items'}};
556 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
558 Log(undef, "Repository '$repo' -> "
559 . join(", ", map { $_->{'uuid'} } @repos_found));
562 croak("Error: Found $n_found repositories with name '$repo'.");
564 $repo = $repos_found[0]->{'fetch_url'};
565 $repo_location = 'remote';
567 Log(undef, "Using $repo_location repository '$repo'");
568 $ENV{"CRUNCH_SRC_URL"} = $repo;
570 # Resolve given script_version (we'll call that $treeish here) to a
571 # commit sha1 ($commit).
572 my $treeish = $Job->{'script_version'};
574 if ($repo_location eq 'remote') {
575 # We minimize excess object-fetching by re-using the same bare
576 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
577 # just keep adding remotes to it as needed.
578 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
579 my $gitcmd = "git --git-dir=\Q$local_repo\E";
581 # Set up our local repo for caching remote objects, making
583 if (!-d $local_repo) {
584 make_path($local_repo) or croak("Error: could not create $local_repo");
586 # This works (exits 0 and doesn't delete fetched objects) even
587 # if $local_repo is already initialized:
588 `$gitcmd init --bare`;
590 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
593 # If $treeish looks like a hash (or abbrev hash) we look it up in
594 # our local cache first, since that's cheaper. (We don't want to
595 # do that with tags/branches though -- those change over time, so
596 # they should always be resolved by the remote repo.)
597 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
598 # Hide stderr because it's normal for this to fail:
599 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
601 # Careful not to resolve a branch named abcdeff to commit 1234567:
602 $sha1 =~ /^$treeish/ &&
603 $sha1 =~ /^([0-9a-f]{40})$/s) {
605 Log(undef, "Commit $commit already present in $local_repo");
609 if (!defined $commit) {
610 # If $treeish isn't just a hash or abbrev hash, or isn't here
611 # yet, we need to fetch the remote to resolve it correctly.
613 # First, remove all local heads. This prevents a name that does
614 # not exist on the remote from resolving to (or colliding with)
615 # a previously fetched branch or tag (possibly from a different
617 remove_tree("$local_repo/refs/heads", {keep_root => 1});
619 Log(undef, "Fetching objects from $repo to $local_repo");
620 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
622 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
626 # Now that the data is all here, we will use our local repo for
627 # the rest of our git activities.
631 my $gitcmd = "git --git-dir=\Q$repo\E";
632 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
633 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
634 croak("`$gitcmd rev-list` exited "
636 .", '$treeish' not found, giving up");
639 Log(undef, "Version $treeish is commit $commit");
641 if ($commit ne $Job->{'script_version'}) {
642 # Record the real commit id in the database, frozentokey, logs,
643 # etc. -- instead of an abbreviation or a branch name which can
644 # become ambiguous or point to a different commit in the future.
645 if (!$Job->update_attributes('script_version' => $commit)) {
646 croak("Error: failed to update job's script_version attribute");
650 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
651 add_git_archive("$gitcmd archive ''\Q$commit\E");
654 my $git_archive = combined_git_archive();
655 if (!defined $git_archive) {
656 Log(undef, "Skip install phase (no git archive)");
658 Log(undef, "Warning: This probably means workers have no source tree!");
663 my $install_script_tries_left = 3;
664 for (my $attempts = 0; $attempts < 3; $attempts++) {
665 my @srunargs = ("srun",
666 "--nodelist=$nodelist",
667 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
668 my @execargs = ("sh", "-c",
669 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
671 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
672 my ($stdout, $stderr);
673 ($exited, $stdout, $stderr) = srun_sync(
674 \@srunargs, \@execargs,
675 {label => "run install script on all workers"},
676 $build_script . $git_archive);
678 my $stderr_anything_from_script = 0;
679 for my $line (split(/\n/, $stderr)) {
680 if ($line !~ /^(srun: error: |starting: \[)/) {
681 $stderr_anything_from_script = 1;
685 last if $exited == 0 || $main::please_freeze;
687 # If the install script fails but doesn't print an error message,
688 # the next thing anyone is likely to do is just run it again in
689 # case it was a transient problem like "slurm communication fails
690 # because the network isn't reliable enough". So we'll just do
691 # that ourselves (up to 3 attempts in total). OTOH, if there is an
692 # error message, the problem is more likely to have a real fix and
693 # we should fail the job so the fixing process can start, instead
694 # of doing 2 more attempts.
695 last if $stderr_anything_from_script;
698 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
699 unlink($tar_filename);
707 foreach (qw (script script_version script_parameters runtime_constraints))
711 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
713 foreach (split (/\n/, $Job->{knobs}))
715 Log (undef, "knob " . $_);
720 $main::success = undef;
726 my $thisround_succeeded = 0;
727 my $thisround_failed = 0;
728 my $thisround_failed_multiple = 0;
729 my $working_slot_count = scalar(@slot);
731 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
732 or $a <=> $b } @jobstep_todo;
733 my $level = $jobstep[$jobstep_todo[0]]->{level};
735 my $initial_tasks_this_level = 0;
736 foreach my $id (@jobstep_todo) {
737 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
740 # If the number of tasks scheduled at this level #T is smaller than the number
741 # of slots available #S, only use the first #T slots, or the first slot on
742 # each node, whichever number is greater.
744 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
745 # based on these numbers. Using fewer slots makes more resources available
746 # to each individual task, which should normally be a better strategy when
747 # there are fewer of them running with less parallelism.
749 # Note that this calculation is not redone if the initial tasks at
750 # this level queue more tasks at the same level. This may harm
751 # overall task throughput for that level.
753 if ($initial_tasks_this_level < @node) {
754 @freeslot = (0..$#node);
755 } elsif ($initial_tasks_this_level < @slot) {
756 @freeslot = (0..$initial_tasks_this_level - 1);
758 @freeslot = (0..$#slot);
760 my $round_num_freeslots = scalar(@freeslot);
761 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";
763 my %round_max_slots = ();
764 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
765 my $this_slot = $slot[$freeslot[$ii]];
766 my $node_name = $this_slot->{node}->{name};
767 $round_max_slots{$node_name} ||= $this_slot->{cpu};
768 last if (scalar(keys(%round_max_slots)) >= @node);
771 Log(undef, "start level $level with $round_num_freeslots slots");
774 my $progress_is_dirty = 1;
775 my $progress_stats_updated = 0;
777 update_progress_stats();
781 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
783 # Don't create new tasks if we already know the job's final result.
784 last if defined($main::success);
786 my $id = $jobstep_todo[$todo_ptr];
787 my $Jobstep = $jobstep[$id];
788 if ($Jobstep->{level} != $level)
793 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
794 set_nonblocking($reader{$id});
796 my $childslot = $freeslot[0];
797 my $childnode = $slot[$childslot]->{node};
798 my $childslotname = join (".",
799 $slot[$childslot]->{node}->{name},
800 $slot[$childslot]->{cpu});
802 my $childpid = fork();
805 $SIG{'INT'} = 'DEFAULT';
806 $SIG{'QUIT'} = 'DEFAULT';
807 $SIG{'TERM'} = 'DEFAULT';
809 foreach (values (%reader))
813 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
814 open(STDOUT,">&writer");
815 open(STDERR,">&writer");
820 delete $ENV{"GNUPGHOME"};
821 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
822 $ENV{"TASK_QSEQUENCE"} = $id;
823 $ENV{"TASK_SEQUENCE"} = $level;
824 $ENV{"JOB_SCRIPT"} = $Job->{script};
825 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
826 $param =~ tr/a-z/A-Z/;
827 $ENV{"JOB_PARAMETER_$param"} = $value;
829 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
830 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
831 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
832 $ENV{"HOME"} = $ENV{"TASK_WORK"};
833 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
834 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
835 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
837 my $keep_mnt = $ENV{"TASK_WORK"}.".keep";
843 "--nodelist=".$childnode->{name},
844 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
845 "--job-name=$job_id.$id.$$",
848 my $stdbuf = " stdbuf --output=0 --error=0 ";
850 my $arv_file_cache = "";
851 if (defined($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'})) {
852 $arv_file_cache = "--file-cache=" . ($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'} * 1024 * 1024);
856 "if [ -e \Q$ENV{TASK_WORK}\E ]; then rm -rf \Q$ENV{TASK_WORK}\E; fi; "
857 ."mkdir -p \Q$ENV{CRUNCH_TMP}\E \Q$ENV{JOB_WORK}\E \Q$ENV{TASK_WORK}\E \Q$keep_mnt\E "
858 ."&& cd \Q$ENV{CRUNCH_TMP}\E "
859 # These environment variables get used explicitly later in
860 # $command. No tool is expected to read these values directly.
861 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
862 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
863 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
864 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP "
865 # $VOLUME_CRUNCHRUNNER and $VOLUME_CERTS will be passed unquoted as
866 # arguments to `docker run`. They must contain their own quoting.
867 .q{&& VOLUME_CRUNCHRUNNER="" VOLUME_CERTS="" }
868 .q{&& if which crunchrunner >/dev/null ; then VOLUME_CRUNCHRUNNER=--volume=$(which crunchrunner):/usr/local/bin/crunchrunner ; fi }
869 .q{&& if test -f /etc/ssl/certs/ca-certificates.crt ; then VOLUME_CERTS=--volume=/etc/ssl/certs/ca-certificates.crt:/etc/arvados/ca-certificates.crt ; }
870 .q{elif test -f /etc/pki/tls/certs/ca-bundle.crt ; then VOLUME_CERTS=--volume=/etc/pki/tls/certs/ca-bundle.crt:/etc/arvados/ca-certificates.crt ; fi };
872 $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 ";
873 $ENV{TASK_KEEPMOUNT} = "$keep_mnt/by_pdh";
874 $ENV{TASK_KEEPMOUNT_TMP} = "$keep_mnt/tmp";
878 my $containername = "$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}";
879 my $cidfile = "$ENV{CRUNCH_TMP}/$containername.cid";
880 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
881 $command .= "$docker_bin run $docker_run_args --name=$containername --attach=stdout --attach=stderr --attach=stdin -i \Q$dockeruserarg\E --cidfile=$cidfile --sig-proxy ";
882 # We only set memory limits if Docker lets us limit both memory and swap.
883 # Memory limits alone have been supported longer, but subprocesses tend
884 # to get SIGKILL if they exceed that without any swap limit set.
885 # See #5642 for additional background.
886 if ($docker_limitmem) {
887 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
890 # The source tree and $destdir directory (which we have
891 # installed on the worker host) are available in the container,
892 # under the same path.
893 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
894 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
896 # Currently, we make the "by_pdh" directory in arv-mount's mount
897 # point appear at /keep inside the container (instead of using
898 # the same path as the host like we do with CRUNCH_SRC and
899 # CRUNCH_INSTALL). However, crunch scripts and utilities must
900 # not rely on this. They must use $TASK_KEEPMOUNT.
901 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
902 $ENV{TASK_KEEPMOUNT} = "/keep";
904 # Ditto TASK_KEEPMOUNT_TMP, as /keep_tmp.
905 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT_TMP}:/keep_tmp\E ";
906 $ENV{TASK_KEEPMOUNT_TMP} = "/keep_tmp";
908 # TASK_WORK is almost exactly like a docker data volume: it
909 # starts out empty, is writable, and persists until no
910 # containers use it any more. We don't use --volumes-from to
911 # share it with other containers: it is only accessible to this
912 # task, and it goes away when this task stops.
914 # However, a docker data volume is writable only by root unless
915 # the mount point already happens to exist in the container with
916 # different permissions. Therefore, we [1] assume /tmp already
917 # exists in the image and is writable by the crunch user; [2]
918 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
919 # writable if they are created by docker while setting up the
920 # other --volumes); and [3] create $TASK_WORK inside the
921 # container using $build_script.
922 $command .= "--volume=/tmp ";
923 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
924 $ENV{"HOME"} = $ENV{"TASK_WORK"};
925 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
927 # TODO: Share a single JOB_WORK volume across all task
928 # containers on a given worker node, and delete it when the job
929 # ends (and, in case that doesn't work, when the next job
932 # For now, use the same approach as TASK_WORK above.
933 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
935 # Bind mount the crunchrunner binary and host TLS certificates file into
937 $command .= "\$VOLUME_CRUNCHRUNNER \$VOLUME_CERTS ";
939 while (my ($env_key, $env_val) = each %ENV)
941 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
942 $command .= "--env=\Q$env_key=$env_val\E ";
945 $command .= "--env=\QHOME=$ENV{HOME}\E ";
946 $command .= "\Q$docker_hash\E ";
948 if ($Job->{arvados_sdk_version}) {
950 $command .= "perl - \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E";
952 $command .= "/bin/sh -c \'python -c " .
953 '"from pkg_resources import get_distribution as get; print \"Using Arvados SDK version\", get(\"arvados-python-client\").version"' .
954 ">&2 2>/dev/null; " .
955 "mkdir -p \"$ENV{JOB_WORK}\" \"$ENV{TASK_WORK}\" && " .
956 "if which stdbuf >/dev/null ; then " .
957 " exec $stdbuf \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
959 " exec \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
964 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -poll=10000 ";
966 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
969 my @execargs = ('bash', '-c', $command);
970 srun (\@srunargs, \@execargs, undef, $build_script);
971 # exec() failed, we assume nothing happened.
972 die "srun() failed on build script\n";
975 if (!defined $childpid)
986 jobstepname => "$job_id.$id.$childpid",
988 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
989 $slot[$childslot]->{pid} = $childpid;
991 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
992 Log ($id, "child $childpid started on $childslotname");
993 $Jobstep->{starttime} = time;
994 $Jobstep->{node} = $childnode->{name};
995 $Jobstep->{slotindex} = $childslot;
996 delete $Jobstep->{stderr};
997 delete $Jobstep->{finishtime};
998 delete $Jobstep->{tempfail};
1000 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
1001 $Jobstep->{'arvados_task'}->save;
1003 splice @jobstep_todo, $todo_ptr, 1;
1006 $progress_is_dirty = 1;
1010 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
1012 last THISROUND if $main::please_freeze;
1013 if ($main::please_info)
1015 $main::please_info = 0;
1017 create_output_collection();
1019 update_progress_stats();
1024 if (!$gotsome || ($latest_refresh + 2 < scalar time))
1026 check_refresh_wanted();
1028 update_progress_stats();
1030 elsif (time - $progress_stats_updated >= 30 || $progress_is_dirty)
1032 update_progress_stats();
1035 select (undef, undef, undef, 0.1);
1037 $working_slot_count = scalar(grep { $_->{node}->{fail_count} == 0 &&
1038 $_->{node}->{hold_count} < 4 } @slot);
1039 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
1040 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
1042 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
1043 .($thisround_failed+$thisround_succeeded)
1044 .") -- giving up on this round";
1045 Log (undef, $message);
1049 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1050 for (my $i=$#freeslot; $i>=0; $i--) {
1051 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1052 push @holdslot, (splice @freeslot, $i, 1);
1055 for (my $i=$#holdslot; $i>=0; $i--) {
1056 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1057 push @freeslot, (splice @holdslot, $i, 1);
1061 # give up if no nodes are succeeding
1062 if ($working_slot_count < 1) {
1063 Log(undef, "Every node has failed -- giving up");
1070 push @freeslot, splice @holdslot;
1071 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1074 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1077 if ($main::please_continue) {
1078 $main::please_continue = 0;
1081 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1083 if (!reapchildren())
1085 check_refresh_wanted();
1087 update_progress_stats();
1088 select (undef, undef, undef, 0.1);
1089 killem (keys %proc) if $main::please_freeze;
1093 update_progress_stats();
1094 freeze_if_want_freeze();
1097 if (!defined $main::success)
1099 if (!@jobstep_todo) {
1101 } elsif ($working_slot_count < 1) {
1102 save_output_collection();
1104 exit(EX_RETRY_UNLOCKED);
1105 } elsif ($thisround_succeeded == 0 &&
1106 ($thisround_failed == 0 || $thisround_failed > 4)) {
1107 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1108 Log (undef, $message);
1113 goto ONELEVEL if !defined $main::success;
1116 release_allocation();
1118 my $collated_output = save_output_collection();
1119 Log (undef, "finish");
1124 if ($collated_output && $main::success) {
1125 $final_state = 'Complete';
1127 $final_state = 'Failed';
1129 $Job->update_attributes('state' => $final_state);
1131 exit (($final_state eq 'Complete') ? 0 : 1);
1135 sub update_progress_stats
1137 $progress_stats_updated = time;
1138 return if !$progress_is_dirty;
1139 my ($todo, $done, $running) = (scalar @jobstep_todo,
1140 scalar @jobstep_done,
1141 scalar keys(%proc));
1142 $Job->{'tasks_summary'} ||= {};
1143 $Job->{'tasks_summary'}->{'todo'} = $todo;
1144 $Job->{'tasks_summary'}->{'done'} = $done;
1145 $Job->{'tasks_summary'}->{'running'} = $running;
1146 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1147 Log (undef, "status: $done done, $running running, $todo todo");
1148 $progress_is_dirty = 0;
1155 my $children_reaped = 0;
1156 my @successful_task_uuids = ();
1158 while((my $pid = waitpid (-1, WNOHANG)) > 0)
1160 my $childstatus = $?;
1162 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1164 . $slot[$proc{$pid}->{slot}]->{cpu});
1165 my $jobstepidx = $proc{$pid}->{jobstepidx};
1168 my $elapsed = time - $proc{$pid}->{time};
1169 my $Jobstep = $jobstep[$jobstepidx];
1171 my $exitvalue = $childstatus >> 8;
1172 my $exitinfo = "exit ".exit_status_s($childstatus);
1173 $Jobstep->{'arvados_task'}->reload;
1174 my $task_success = $Jobstep->{'arvados_task'}->{success};
1176 Log ($jobstepidx, "child $pid on $whatslot $exitinfo success=$task_success");
1178 if (!defined $task_success) {
1179 # task did not indicate one way or the other --> fail
1180 Log($jobstepidx, sprintf(
1181 "ERROR: Task process exited %s, but never updated its task record to indicate success and record its output.",
1182 exit_status_s($childstatus)));
1183 $Jobstep->{'arvados_task'}->{success} = 0;
1184 $Jobstep->{'arvados_task'}->save;
1191 $temporary_fail ||= $Jobstep->{tempfail};
1192 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1194 ++$thisround_failed;
1195 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1197 # Check for signs of a failed or misconfigured node
1198 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1199 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1200 # Don't count this against jobstep failure thresholds if this
1201 # node is already suspected faulty and srun exited quickly
1202 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1204 Log ($jobstepidx, "blaming failure on suspect node " .
1205 $slot[$proc{$pid}->{slot}]->{node}->{name});
1206 $temporary_fail ||= 1;
1208 ban_node_by_slot($proc{$pid}->{slot});
1211 Log ($jobstepidx, sprintf('failure (#%d, %s) after %d seconds',
1212 ++$Jobstep->{'failures'},
1213 $temporary_fail ? 'temporary' : 'permanent',
1216 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1217 # Give up on this task, and the whole job
1220 # Put this task back on the todo queue
1221 push @jobstep_todo, $jobstepidx;
1222 $Job->{'tasks_summary'}->{'failed'}++;
1226 push @successful_task_uuids, $Jobstep->{'arvados_task'}->{uuid};
1227 ++$thisround_succeeded;
1228 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1229 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1230 $slot[$proc{$pid}->{slot}]->{node}->{fail_count} = 0;
1231 push @jobstep_done, $jobstepidx;
1232 Log ($jobstepidx, "success in $elapsed seconds");
1234 $Jobstep->{exitcode} = $childstatus;
1235 $Jobstep->{finishtime} = time;
1236 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1237 $Jobstep->{'arvados_task'}->save;
1238 process_stderr_final ($jobstepidx);
1239 Log ($jobstepidx, sprintf("task output (%d bytes): %s",
1240 length($Jobstep->{'arvados_task'}->{output}),
1241 $Jobstep->{'arvados_task'}->{output}));
1243 close $reader{$jobstepidx};
1244 delete $reader{$jobstepidx};
1245 delete $slot[$proc{$pid}->{slot}]->{pid};
1246 push @freeslot, $proc{$pid}->{slot};
1249 $progress_is_dirty = 1;
1252 if (scalar(@successful_task_uuids) > 0)
1254 Log (undef, sprintf("%d tasks exited (%d succeeded), checking for new tasks from API server.", $children_reaped, scalar(@successful_task_uuids)));
1256 my $newtask_list = [];
1257 my $newtask_results;
1259 $newtask_results = api_call(
1261 'filters' => [["created_by_job_task_uuid","in",\@successful_task_uuids]],
1262 'order' => 'qsequence',
1263 'offset' => scalar(@$newtask_list),
1265 push(@$newtask_list, @{$newtask_results->{items}});
1266 } while (@{$newtask_results->{items}});
1267 Log (undef, sprintf("Got %d new tasks from API server.", scalar(@$newtask_list)));
1268 foreach my $arvados_task (@$newtask_list) {
1270 'level' => $arvados_task->{'sequence'},
1272 'arvados_task' => $arvados_task
1274 push @jobstep, $jobstep;
1275 push @jobstep_todo, $#jobstep;
1279 return $children_reaped;
1282 sub check_refresh_wanted
1284 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1286 $stat[9] > $latest_refresh &&
1287 # ...and we have actually locked the job record...
1288 $job_id eq $Job->{'uuid'}) {
1289 $latest_refresh = scalar time;
1290 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1291 for my $attr ('cancelled_at',
1292 'cancelled_by_user_uuid',
1293 'cancelled_by_client_uuid',
1295 $Job->{$attr} = $Job2->{$attr};
1297 if ($Job->{'state'} ne "Running") {
1298 if ($Job->{'state'} eq "Cancelled") {
1299 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1301 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1304 $main::please_freeze = 1;
1311 my $last_squeue_check = $squeue_checked;
1313 # Do not call `squeue` or check the kill list more than once every
1315 return if $last_squeue_check > time - 15;
1316 $squeue_checked = time;
1318 # Look for children from which we haven't received stderr data since
1319 # the last squeue check. If no such children exist, all procs are
1320 # alive and there's no need to even look at squeue.
1322 # As long as the crunchstat poll interval (10s) is shorter than the
1323 # squeue check interval (15s) this should make the squeue check an
1325 my $silent_procs = 0;
1326 for my $js (map {$jobstep[$_->{jobstepidx}]} values %proc)
1328 if (!exists($js->{stderr_at}))
1330 $js->{stderr_at} = 0;
1332 if ($js->{stderr_at} < $last_squeue_check)
1337 return if $silent_procs == 0;
1339 # use killem() on procs whose killtime is reached
1340 while (my ($pid, $procinfo) = each %proc)
1342 my $js = $jobstep[$procinfo->{jobstepidx}];
1343 if (exists $procinfo->{killtime}
1344 && $procinfo->{killtime} <= time
1345 && $js->{stderr_at} < $last_squeue_check)
1348 if ($js->{stderr_at}) {
1349 $sincewhen = " in last " . (time - $js->{stderr_at}) . "s";
1351 Log($procinfo->{jobstepidx}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
1358 # here is an opportunity to check for mysterious problems with local procs
1362 # Get a list of steps still running. Note: squeue(1) says --steps
1363 # selects a format (which we override anyway) and allows us to
1364 # specify which steps we're interested in (which we don't).
1365 # Importantly, it also changes the meaning of %j from "job name" to
1366 # "step name" and (although this isn't mentioned explicitly in the
1367 # docs) switches from "one line per job" mode to "one line per step"
1368 # mode. Without it, we'd just get a list of one job, instead of a
1370 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%j' --noheader`;
1373 Log(undef, "warning: squeue exit status $? ($!)");
1378 # which of my jobsteps are running, according to squeue?
1380 for my $jobstepname (@squeue)
1382 $ok{$jobstepname} = 1;
1385 # Check for child procs >60s old and not mentioned by squeue.
1386 while (my ($pid, $procinfo) = each %proc)
1388 if ($procinfo->{time} < time - 60
1389 && $procinfo->{jobstepname}
1390 && !exists $ok{$procinfo->{jobstepname}}
1391 && !exists $procinfo->{killtime})
1393 # According to slurm, this task has ended (successfully or not)
1394 # -- but our srun child hasn't exited. First we must wait (30
1395 # seconds) in case this is just a race between communication
1396 # channels. Then, if our srun child process still hasn't
1397 # terminated, we'll conclude some slurm communication
1398 # error/delay has caused the task to die without notifying srun,
1399 # and we'll kill srun ourselves.
1400 $procinfo->{killtime} = time + 30;
1401 Log($procinfo->{jobstepidx}, "notice: task is not in slurm queue but srun process $pid has not exited");
1407 sub release_allocation
1411 Log (undef, "release job allocation");
1412 system "scancel $ENV{SLURM_JOB_ID}";
1421 my $sel = IO::Select->new();
1422 foreach my $jobstepidx (keys %reader)
1424 my $fd = $reader{$jobstepidx};
1426 $fd_job{$fd} = $jobstepidx;
1428 if (my $stdout_fd = $jobstep[$jobstepidx]->{stdout_r}) {
1429 $sel->add($stdout_fd);
1430 $fd_job{$stdout_fd} = $jobstepidx;
1433 # select on all reader fds with 0.1s timeout
1434 my @ready_fds = $sel->can_read(0.1);
1435 foreach my $fd (@ready_fds)
1438 if (0 < sysread ($fd, $buf, 65536))
1441 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1443 my $jobstepidx = $fd_job{$fd};
1444 if ($jobstep[$jobstepidx]->{stdout_r} == $fd) {
1445 $jobstep[$jobstepidx]->{stdout_captured} .= $buf;
1449 $jobstep[$jobstepidx]->{stderr_at} = time;
1450 $jobstep[$jobstepidx]->{stderr} .= $buf;
1452 # Consume everything up to the last \n
1453 preprocess_stderr ($jobstepidx);
1455 if (length ($jobstep[$jobstepidx]->{stderr}) > 16384)
1457 # If we get a lot of stderr without a newline, chop off the
1458 # front to avoid letting our buffer grow indefinitely.
1459 substr ($jobstep[$jobstepidx]->{stderr},
1460 0, length($jobstep[$jobstepidx]->{stderr}) - 8192) = "";
1468 # Consume all full lines of stderr for a jobstep. Everything after the
1469 # last newline will remain in $jobstep[$jobstepidx]->{stderr} after
1471 sub preprocess_stderr
1473 my $jobstepidx = shift;
1474 # slotindex is only defined for children running Arvados job tasks.
1475 # Be prepared to handle the undef case (for setup srun calls, etc.).
1476 my $job_slot_index = $jobstep[$jobstepidx]->{slotindex};
1478 while ($jobstep[$jobstepidx]->{stderr} =~ /^(.*?)\n/) {
1480 substr $jobstep[$jobstepidx]->{stderr}, 0, 1+length($line), "";
1481 Log ($jobstepidx, "stderr $line");
1482 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1484 $main::please_freeze = 1;
1486 elsif ($line =~ /srun: error: (Node failure on|Aborting, .*\bio error\b)/) {
1487 $jobstep[$jobstepidx]->{tempfail} = 1;
1488 if (defined($job_slot_index)) {
1489 $slot[$job_slot_index]->{node}->{fail_count}++;
1490 ban_node_by_slot($job_slot_index);
1493 elsif ($line =~ /srun: error: (Unable to create job step|.*: Communication connection failure)/) {
1494 $jobstep[$jobstepidx]->{tempfail} = 1;
1495 ban_node_by_slot($job_slot_index) if (defined($job_slot_index));
1497 elsif ($line =~ /\bKeep(Read|Write|Request)Error:/) {
1498 $jobstep[$jobstepidx]->{tempfail} = 1;
1504 sub process_stderr_final
1506 my $jobstepidx = shift;
1507 preprocess_stderr ($jobstepidx);
1510 Log ($jobstepidx, "stderr $_");
1511 } split ("\n", $jobstep[$jobstepidx]->{stderr});
1512 $jobstep[$jobstepidx]->{stderr} = '';
1519 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1520 Log(undef, "fetch_block run error from arv-get $hash: $!");
1523 my $output_block = "";
1526 my $bytes = sysread($keep, $buf, 1024 * 1024);
1527 if (!defined $bytes) {
1528 Log(undef, "fetch_block read error from arv-get: $!");
1529 $output_block = undef;
1531 } elsif ($bytes == 0) {
1532 # sysread returns 0 at the end of the pipe.
1535 # some bytes were read into buf.
1536 $output_block .= $buf;
1541 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1542 $output_block = undef;
1544 return $output_block;
1547 # Create a collection by concatenating the output of all tasks (each
1548 # task's output is either a manifest fragment, a locator for a
1549 # manifest fragment stored in Keep, or nothing at all). Return the
1550 # portable_data_hash of the new collection.
1551 sub create_output_collection
1553 Log (undef, "collate");
1555 my ($child_out, $child_in);
1556 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1559 print (arvados.api("v1").collections().
1560 create(body={"manifest_text": sys.stdin.read()}).
1561 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1565 my $manifest_size = 0;
1569 my $output = $_->{'arvados_task'}->{output};
1570 next if (!defined($output));
1572 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1573 $next_write = fetch_block($output);
1575 $next_write = $output;
1577 if (defined($next_write)) {
1578 if (!defined(syswrite($child_in, $next_write))) {
1579 # There's been an error writing. Stop the loop.
1580 # We'll log details about the exit code later.
1583 $manifest_size += length($next_write);
1586 my $uuid = $_->{'arvados_task'}->{'uuid'};
1587 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1592 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1595 my $s = IO::Select->new($child_out);
1596 if ($s->can_read(120)) {
1597 sysread($child_out, $joboutput, 1024 * 1024);
1600 Log(undef, "output collection creation exited " . exit_status_s($?));
1606 Log (undef, "timed out while creating output collection");
1607 foreach my $signal (2, 2, 2, 15, 15, 9) {
1608 kill($signal, $pid);
1609 last if waitpid($pid, WNOHANG) == -1;
1618 # Calls create_output_collection, logs the result, and returns it.
1619 # If that was successful, save that as the output in the job record.
1620 sub save_output_collection {
1621 my $collated_output = create_output_collection();
1623 if (!$collated_output) {
1624 Log(undef, "Failed to write output collection");
1627 Log(undef, "job output $collated_output");
1628 $Job->update_attributes('output' => $collated_output);
1630 return $collated_output;
1637 my $sig = 2; # SIGINT first
1638 if (exists $proc{$_}->{"sent_$sig"} &&
1639 time - $proc{$_}->{"sent_$sig"} > 4)
1641 $sig = 15; # SIGTERM if SIGINT doesn't work
1643 if (exists $proc{$_}->{"sent_$sig"} &&
1644 time - $proc{$_}->{"sent_$sig"} > 4)
1646 $sig = 9; # SIGKILL if SIGTERM doesn't work
1648 if (!exists $proc{$_}->{"sent_$sig"})
1650 Log ($proc{$_}->{jobstepidx}, "sending 2x signal $sig to pid $_");
1652 select (undef, undef, undef, 0.1);
1655 kill $sig, $_; # srun wants two SIGINT to really interrupt
1657 $proc{$_}->{"sent_$sig"} = time;
1658 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1668 vec($bits,fileno($_),1) = 1;
1674 # Send log output to Keep via arv-put.
1676 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1677 # $log_pipe_out_buf is a string containing all output read from arv-put so far.
1678 # $log_pipe_out_select is an IO::Select object around $log_pipe_out.
1679 # $log_pipe_pid is the pid of the arv-put subprocess.
1681 # The only functions that should access these variables directly are:
1683 # log_writer_start($logfilename)
1684 # Starts an arv-put pipe, reading data on stdin and writing it to
1685 # a $logfilename file in an output collection.
1687 # log_writer_read_output([$timeout])
1688 # Read output from $log_pipe_out and append it to $log_pipe_out_buf.
1689 # Passes $timeout to the select() call, with a default of 0.01.
1690 # Returns the result of the last read() call on $log_pipe_out, or
1691 # -1 if read() wasn't called because select() timed out.
1692 # Only other log_writer_* functions should need to call this.
1694 # log_writer_send($txt)
1695 # Writes $txt to the output log collection.
1697 # log_writer_finish()
1698 # Closes the arv-put pipe and returns the output that it produces.
1700 # log_writer_is_active()
1701 # Returns a true value if there is currently a live arv-put
1702 # process, false otherwise.
1704 my ($log_pipe_in, $log_pipe_out, $log_pipe_out_buf, $log_pipe_out_select,
1707 sub log_writer_start($)
1709 my $logfilename = shift;
1710 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1714 '--filename', $logfilename,
1716 $log_pipe_out_buf = "";
1717 $log_pipe_out_select = IO::Select->new($log_pipe_out);
1720 sub log_writer_read_output {
1721 my $timeout = shift || 0.01;
1723 while ($read && $log_pipe_out_select->can_read($timeout)) {
1724 $read = read($log_pipe_out, $log_pipe_out_buf, 65536,
1725 length($log_pipe_out_buf));
1727 if (!defined($read)) {
1728 Log(undef, "error reading log manifest from arv-put: $!");
1733 sub log_writer_send($)
1736 print $log_pipe_in $txt;
1737 log_writer_read_output();
1740 sub log_writer_finish()
1742 return unless $log_pipe_pid;
1744 close($log_pipe_in);
1746 my $logger_failed = 0;
1747 my $read_result = log_writer_read_output(120);
1748 if ($read_result == -1) {
1749 $logger_failed = -1;
1750 Log (undef, "timed out reading from 'arv-put'");
1751 } elsif ($read_result != 0) {
1752 $logger_failed = -2;
1753 Log(undef, "failed to read arv-put log manifest to EOF");
1756 waitpid($log_pipe_pid, 0);
1758 $logger_failed ||= $?;
1759 Log(undef, "log_writer_finish: arv-put exited " . exit_status_s($?))
1762 close($log_pipe_out);
1763 my $arv_put_output = $logger_failed ? undef : $log_pipe_out_buf;
1764 $log_pipe_pid = $log_pipe_in = $log_pipe_out = $log_pipe_out_buf =
1765 $log_pipe_out_select = undef;
1767 return $arv_put_output;
1770 sub log_writer_is_active() {
1771 return $log_pipe_pid;
1774 sub Log # ($jobstepidx, $logmessage)
1776 my ($jobstepidx, $logmessage) = @_;
1777 if ($logmessage =~ /\n/) {
1778 for my $line (split (/\n/, $_[1])) {
1779 Log ($jobstepidx, $line);
1783 my $fh = select STDERR; $|=1; select $fh;
1785 if (defined($jobstepidx) && exists($jobstep[$jobstepidx]->{arvados_task})) {
1786 $task_qseq = $jobstepidx;
1788 my $message = sprintf ("%s %d %s %s", $job_id, $$, $task_qseq, $logmessage);
1789 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1792 if (log_writer_is_active() || -t STDERR) {
1793 my @gmtime = gmtime;
1794 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1795 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1797 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1799 if (log_writer_is_active()) {
1800 log_writer_send($datetime . " " . $message);
1807 my ($package, $file, $line) = caller;
1808 my $message = "@_ at $file line $line\n";
1809 Log (undef, $message);
1810 freeze() if @jobstep_todo;
1811 create_output_collection() if @jobstep_todo;
1821 if ($Job->{'state'} eq 'Cancelled') {
1822 $Job->update_attributes('finished_at' => scalar gmtime);
1824 $Job->update_attributes('state' => 'Failed');
1831 my $justcheckpoint = shift; # false if this will be the last meta saved
1832 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1833 return unless log_writer_is_active();
1834 my $log_manifest = log_writer_finish();
1835 return unless defined($log_manifest);
1838 my $prev_log_coll = api_call("collections/get", uuid => $Job->{log});
1839 $log_manifest = $prev_log_coll->{manifest_text} . $log_manifest;
1842 my $log_coll = api_call(
1843 "collections/create", ensure_unique_name => 1, collection => {
1844 manifest_text => $log_manifest,
1845 owner_uuid => $Job->{owner_uuid},
1846 name => sprintf("Log from %s job %s", $Job->{script}, $Job->{uuid}),
1848 Log(undef, "log collection is " . $log_coll->{portable_data_hash});
1849 $Job->update_attributes('log' => $log_coll->{portable_data_hash});
1853 sub freeze_if_want_freeze
1855 if ($main::please_freeze)
1857 release_allocation();
1860 # kill some srun procs before freeze+stop
1861 map { $proc{$_} = {} } @_;
1864 killem (keys %proc);
1865 select (undef, undef, undef, 0.1);
1867 while (($died = waitpid (-1, WNOHANG)) > 0)
1869 delete $proc{$died};
1874 create_output_collection();
1884 Log (undef, "Freeze not implemented");
1891 croak ("Thaw not implemented");
1907 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1914 my $srunargs = shift;
1915 my $execargs = shift;
1916 my $opts = shift || {};
1919 my $label = exists $opts->{label} ? $opts->{label} : "@$execargs";
1920 Log (undef, "$label: start");
1922 my ($stderr_r, $stderr_w);
1923 pipe $stderr_r, $stderr_w or croak("pipe() failed: $!");
1925 my ($stdout_r, $stdout_w);
1926 pipe $stdout_r, $stdout_w or croak("pipe() failed: $!");
1928 my $srunpid = fork();
1933 fcntl($stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
1934 fcntl($stdout_w, F_SETFL, 0) or croak($!);
1935 open(STDERR, ">&", $stderr_w);
1936 open(STDOUT, ">&", $stdout_w);
1937 srun ($srunargs, $execargs, $opts, $stdin);
1943 set_nonblocking($stderr_r);
1944 set_nonblocking($stdout_r);
1946 # Add entries to @jobstep and %proc so check_squeue() and
1947 # freeze_if_want_freeze() can treat it like a job task process.
1951 stderr_captured => '',
1952 stdout_r => $stdout_r,
1953 stdout_captured => '',
1955 my $jobstepidx = $#jobstep;
1957 jobstepidx => $jobstepidx,
1959 $reader{$jobstepidx} = $stderr_r;
1961 while ($srunpid != waitpid ($srunpid, WNOHANG)) {
1962 my $busy = readfrompipes();
1963 if (!$busy || ($latest_refresh + 2 < scalar time)) {
1964 check_refresh_wanted();
1968 select(undef, undef, undef, 0.1);
1970 killem(keys %proc) if $main::please_freeze;
1974 1 while readfrompipes();
1975 process_stderr_final ($jobstepidx);
1977 Log (undef, "$label: exit ".exit_status_s($exited));
1981 delete $proc{$srunpid};
1982 delete $reader{$jobstepidx};
1984 my $j = pop @jobstep;
1985 # If the srun showed signs of tempfail, ensure the caller treats that as a
1987 if ($main::please_freeze || $j->{tempfail}) {
1990 return ($exited, $j->{stdout_captured}, $j->{stderr_captured});
1996 my $srunargs = shift;
1997 my $execargs = shift;
1998 my $opts = shift || {};
2000 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
2002 $Data::Dumper::Terse = 1;
2003 $Data::Dumper::Indent = 0;
2004 my $show_cmd = Dumper($args);
2005 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
2006 $show_cmd =~ s/\n/ /g;
2007 if ($opts->{fork}) {
2008 Log(undef, "starting: $show_cmd");
2010 # This is a child process: parent is in charge of reading our
2011 # stderr and copying it to Log() if needed.
2012 warn "starting: $show_cmd\n";
2015 if (defined $stdin) {
2016 my $child = open STDIN, "-|";
2017 defined $child or die "no fork: $!";
2019 print $stdin or die $!;
2020 close STDOUT or die $!;
2025 return system (@$args) if $opts->{fork};
2028 warn "ENV size is ".length(join(" ",%ENV));
2029 die "exec failed: $!: @$args";
2033 sub ban_node_by_slot {
2034 # Don't start any new jobsteps on this node for 60 seconds
2036 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
2037 $slot[$slotid]->{node}->{hold_count}++;
2038 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
2043 my ($lockfile, $error_message) = @_;
2044 open L, ">", $lockfile or croak("$lockfile: $!");
2045 if (!flock L, LOCK_EX|LOCK_NB) {
2046 croak("Can't lock $lockfile: $error_message\n");
2050 sub find_docker_image {
2051 # Given a Keep locator, check to see if it contains a Docker image.
2052 # If so, return its stream name and Docker hash.
2053 # If not, return undef for both values.
2054 my $locator = shift;
2055 my ($streamname, $filename);
2056 my $image = api_call("collections/get", uuid => $locator);
2058 foreach my $line (split(/\n/, $image->{manifest_text})) {
2059 my @tokens = split(/\s+/, $line);
2061 $streamname = shift(@tokens);
2062 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
2063 if (defined($filename)) {
2064 return (undef, undef); # More than one file in the Collection.
2066 $filename = (split(/:/, $filedata, 3))[2];
2071 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
2072 return ($streamname, $1);
2074 return (undef, undef);
2079 # Calculate the number of times an operation should be retried,
2080 # assuming exponential backoff, and that we're willing to retry as
2081 # long as tasks have been running. Enforce a minimum of 3 retries.
2082 my ($starttime, $endtime, $timediff, $retries);
2084 $starttime = $jobstep[0]->{starttime};
2085 $endtime = $jobstep[-1]->{finishtime};
2087 if (!defined($starttime)) {
2089 } elsif (!defined($endtime)) {
2090 $timediff = time - $starttime;
2092 $timediff = ($endtime - $starttime) - (time - $endtime);
2094 if ($timediff > 0) {
2095 $retries = int(log($timediff) / log(2));
2097 $retries = 1; # Use the minimum.
2099 return ($retries > 3) ? $retries : 3;
2103 # Pass in two function references.
2104 # This method will be called with the remaining arguments.
2105 # If it dies, retry it with exponential backoff until it succeeds,
2106 # or until the current retry_count is exhausted. After each failure
2107 # that can be retried, the second function will be called with
2108 # the current try count (0-based), next try time, and error message.
2109 my $operation = shift;
2110 my $retry_callback = shift;
2111 my $retries = retry_count();
2112 foreach my $try_count (0..$retries) {
2113 my $next_try = time + (2 ** $try_count);
2114 my $result = eval { $operation->(@_); };
2117 } elsif ($try_count < $retries) {
2118 $retry_callback->($try_count, $next_try, $@);
2119 my $sleep_time = $next_try - time;
2120 sleep($sleep_time) if ($sleep_time > 0);
2123 # Ensure the error message ends in a newline, so Perl doesn't add
2124 # retry_op's line number to it.
2130 # Pass in a /-separated API method name, and arguments for it.
2131 # This function will call that method, retrying as needed until
2132 # the current retry_count is exhausted, with a log on the first failure.
2133 my $method_name = shift;
2134 my $log_api_retry = sub {
2135 my ($try_count, $next_try_at, $errmsg) = @_;
2136 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
2137 $errmsg =~ s/\s/ /g;
2138 $errmsg =~ s/\s+$//;
2140 if ($next_try_at < time) {
2141 $retry_msg = "Retrying.";
2143 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
2144 $retry_msg = "Retrying at $next_try_fmt.";
2146 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
2149 foreach my $key (split(/\//, $method_name)) {
2150 $method = $method->{$key};
2152 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
2156 # Given a $?, return a human-readable exit code string like "0" or
2157 # "1" or "0 with signal 1" or "1 with signal 11".
2158 my $exitcode = shift;
2159 my $s = $exitcode >> 8;
2160 if ($exitcode & 0x7f) {
2161 $s .= " with signal " . ($exitcode & 0x7f);
2163 if ($exitcode & 0x80) {
2164 $s .= " with core dump";
2169 sub handle_readall {
2170 # Pass in a glob reference to a file handle.
2171 # Read all its contents and return them as a string.
2172 my $fh_glob_ref = shift;
2174 return <$fh_glob_ref>;
2177 sub tar_filename_n {
2179 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
2182 sub add_git_archive {
2183 # Pass in a git archive command as a string or list, a la system().
2184 # This method will save its output to be included in the archive sent to the
2188 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
2189 croak("Failed to save git archive: $!");
2191 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
2193 waitpid($git_pid, 0);
2196 croak("Failed to save git archive: git exited " . exit_status_s($?));
2200 sub combined_git_archive {
2201 # Combine all saved tar archives into a single archive, then return its
2202 # contents in a string. Return undef if no archives have been saved.
2203 if ($git_tar_count < 1) {
2206 my $base_tar_name = tar_filename_n(1);
2207 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
2208 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
2209 if ($tar_exit != 0) {
2210 croak("Error preparing build archive: tar -A exited " .
2211 exit_status_s($tar_exit));
2214 if (!open(GIT_TAR, "<", $base_tar_name)) {
2215 croak("Could not open build archive: $!");
2217 my $tar_contents = handle_readall(\*GIT_TAR);
2219 return $tar_contents;
2222 sub set_nonblocking {
2224 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
2225 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
2231 # This is crunch-job's internal dispatch script. crunch-job running on the API
2232 # server invokes this script on individual compute nodes, or localhost if we're
2233 # running a job locally. It gets called in two modes:
2235 # * No arguments: Installation mode. Read a tar archive from the DATA
2236 # file handle; it includes the Crunch script's source code, and
2237 # maybe SDKs as well. Those should be installed in the proper
2238 # locations. This runs outside of any Docker container, so don't try to
2239 # introspect Crunch's runtime environment.
2241 # * With arguments: Crunch script run mode. This script should set up the
2242 # environment, then run the command specified in the arguments. This runs
2243 # inside any Docker container.
2246 use File::Path qw( make_path remove_tree );
2247 use POSIX qw(getcwd);
2249 use constant TASK_TEMPFAIL => 111;
2251 # Map SDK subdirectories to the path environments they belong to.
2252 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2254 my $destdir = $ENV{"CRUNCH_SRC"};
2255 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2256 my $repo = $ENV{"CRUNCH_SRC_URL"};
2257 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2258 my $job_work = $ENV{"JOB_WORK"};
2259 my $task_work = $ENV{"TASK_WORK"};
2261 open(STDOUT_ORIG, ">&", STDOUT);
2262 open(STDERR_ORIG, ">&", STDERR);
2264 for my $dir ($destdir, $job_work, $task_work) {
2267 -e $dir or die "Failed to create temporary directory ($dir): $!";
2272 remove_tree($task_work, {keep_root => 1});
2275 ### Crunch script run mode
2277 # We want to do routine logging during task 0 only. This gives the user
2278 # the information they need, but avoids repeating the information for every
2281 if ($ENV{TASK_SEQUENCE} eq "0") {
2284 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2290 my $python_src = "$install_dir/python";
2291 my $venv_dir = "$job_work/.arvados.venv";
2292 my $venv_built = -e "$venv_dir/bin/activate";
2293 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2294 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2295 "--python=python2.7", $venv_dir);
2296 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2298 $Log->("Built Python SDK virtualenv");
2301 my @pysdk_version_cmd = ("python", "-c",
2302 "from pkg_resources import get_distribution as get; print get('arvados-python-client').version");
2304 $Log->("Running in Python SDK virtualenv");
2305 @pysdk_version_cmd = ();
2306 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2307 @ARGV = ("/bin/sh", "-ec",
2308 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2309 } elsif (-d $python_src) {
2310 $Log->("Warning: virtualenv not found inside Docker container default " .
2311 "\$PATH. Can't install Python SDK.");
2314 if (@pysdk_version_cmd) {
2315 open(my $pysdk_version_pipe, "-|", @pysdk_version_cmd);
2316 my $pysdk_version = <$pysdk_version_pipe>;
2317 close($pysdk_version_pipe);
2319 chomp($pysdk_version);
2320 $Log->("Using Arvados SDK version $pysdk_version");
2322 # A lot could've gone wrong here, but pretty much all of it means that
2323 # Python won't be able to load the Arvados SDK.
2324 $Log->("Warning: Arvados SDK not found");
2328 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2329 my $sdk_path = "$install_dir/$sdk_dir";
2331 if ($ENV{$sdk_envkey}) {
2332 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2334 $ENV{$sdk_envkey} = $sdk_path;
2336 $Log->("Arvados SDK added to %s", $sdk_envkey);
2341 die "Cannot exec `@ARGV`: $!";
2344 ### Installation mode
2345 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2347 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2348 # This exact git archive (source + arvados sdk) is already installed
2349 # here, so there's no need to reinstall it.
2351 # We must consume our DATA section, though: otherwise the process
2352 # feeding it to us will get SIGPIPE.
2354 while (read(DATA, $buf, 65536)) { }
2359 unlink "$destdir.archive_hash";
2363 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2364 local $SIG{PIPE} = "IGNORE";
2365 warn "Extracting archive: $archive_hash\n";
2366 # --ignore-zeros is necessary sometimes: depending on how much NUL
2367 # padding tar -A put on our combined archive (which in turn depends
2368 # on the length of the component archives) tar without
2369 # --ignore-zeros will exit before consuming stdin and cause close()
2370 # to fail on the resulting SIGPIPE.
2371 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2372 die "Error launching 'tar -xC $destdir': $!";
2374 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2375 # get SIGPIPE. We must feed it data incrementally.
2377 while (read(DATA, $tar_input, 65536)) {
2378 print TARX $tar_input;
2381 die "'tar -xC $destdir' exited $?: $!";
2387 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2389 foreach my $sdk_lang (("python",
2390 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2391 if (-d "$sdk_root/$sdk_lang") {
2392 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2393 die "Failed to install $sdk_lang SDK: $!";
2399 my $python_dir = "$install_dir/python";
2400 if ((-d $python_dir) and can_run("python2.7")) {
2401 open(my $egg_info_pipe, "-|",
2402 "python2.7 \Q$python_dir/setup.py\E egg_info 2>&1 >/dev/null");
2403 my @egg_info_errors = <$egg_info_pipe>;
2404 close($egg_info_pipe);
2407 if (@egg_info_errors and (($egg_info_errors[-1] =~ /\bgit\b/) or ($egg_info_errors[-1] =~ /\[Errno 2\]/))) {
2408 # egg_info apparently failed because it couldn't ask git for a build tag.
2409 # Specify no build tag.
2410 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2411 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2414 my $egg_info_exit = $? >> 8;
2415 foreach my $errline (@egg_info_errors) {
2418 warn "python setup.py egg_info failed: exit $egg_info_exit";
2419 exit ($egg_info_exit || 1);
2424 # Hide messages from the install script (unless it fails: shell_or_die
2425 # will show $destdir.log in that case).
2426 open(STDOUT, ">>", "$destdir.log");
2427 open(STDERR, ">&", STDOUT);
2429 if (-e "$destdir/crunch_scripts/install") {
2430 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2431 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2433 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2434 } elsif (-e "./install.sh") {
2435 shell_or_die (undef, "./install.sh", $install_dir);
2438 if ($archive_hash) {
2439 unlink "$destdir.archive_hash.new";
2440 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2441 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2447 my $command_name = shift;
2448 open(my $which, "-|", "which", $command_name);
2449 while (<$which>) { }
2456 my $exitcode = shift;
2458 if ($ENV{"DEBUG"}) {
2459 print STDERR "@_\n";
2461 if (system (@_) != 0) {
2464 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2465 open STDERR, ">&STDERR_ORIG";
2466 system ("cat $destdir.log >&2");
2467 warn "@_ failed ($err): $exitstatus";
2468 if (defined($exitcode)) {
2472 exit (($code >> 8) || 1);