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
420 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
424 my ($exited, $stdout, $stderr) = srun_sync(
425 ["srun", "--nodelist=" . join(',', @node)],
426 ["/bin/bash", "-o", "pipefail", "-ec", $docker_install_script],
427 {label => "load docker image"});
430 exit(EX_RETRY_UNLOCKED);
433 # Determine whether this version of Docker supports memory+swap limits.
434 ($exited, $stdout, $stderr) = srun_sync(
435 ["srun", "--nodes=1"],
436 [$docker_bin, 'run', '--help'],
437 {label => "check --memory-swap feature"});
438 $docker_limitmem = ($stdout =~ /--memory-swap/);
440 # Find a non-root Docker user to use.
441 # Tries the default user for the container, then 'crunch', then 'nobody',
442 # testing for whether the actual user id is non-zero. This defends against
443 # mistakes but not malice, but we intend to harden the security in the future
444 # so we don't want anyone getting used to their jobs running as root in their
446 my @tryusers = ("", "crunch", "nobody");
447 foreach my $try_user (@tryusers) {
450 if ($try_user eq "") {
451 $label = "check whether default user is UID 0";
454 $label = "check whether user '$try_user' is UID 0";
455 $try_user_arg = "--user=$try_user";
457 my ($exited, $stdout, $stderr) = srun_sync(
458 ["srun", "--nodes=1"],
460 "$docker_bin run $docker_run_args $try_user_arg $docker_hash id --user"],
463 if ($exited == 0 && $stdout =~ /^\d+$/ && $stdout > 0) {
464 $dockeruserarg = $try_user_arg;
465 if ($try_user eq "") {
466 Log(undef, "Container will run with default user");
468 Log(undef, "Container will run with $dockeruserarg");
474 if (!defined $dockeruserarg) {
475 croak("Could not find a user in container that is not UID 0 (tried default user, @tryusers) or there was a problem running 'id' in the container.");
478 if ($Job->{arvados_sdk_version}) {
479 # The job also specifies an Arvados SDK version. Add the SDKs to the
480 # tar file for the build script to install.
481 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
482 $Job->{arvados_sdk_version}));
483 add_git_archive("git", "--git-dir=$git_dir", "archive",
484 "--prefix=.arvados.sdk/",
485 $Job->{arvados_sdk_version}, "sdk");
489 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
490 # If script_version looks like an absolute path, *and* the --git-dir
491 # argument was not given -- which implies we were not invoked by
492 # crunch-dispatch -- we will use the given path as a working
493 # directory instead of resolving script_version to a git commit (or
494 # doing anything else with git).
495 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
496 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
499 # Resolve the given script_version to a git commit sha1. Also, if
500 # the repository is remote, clone it into our local filesystem: this
501 # ensures "git archive" will work, and is necessary to reliably
502 # resolve a symbolic script_version like "master^".
503 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
505 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
507 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
509 # If we're running under crunch-dispatch, it will have already
510 # pulled the appropriate source tree into its own repository, and
511 # given us that repo's path as $git_dir.
513 # If we're running a "local" job, we might have to fetch content
514 # from a remote repository.
516 # (Currently crunch-dispatch gives a local path with --git-dir, but
517 # we might as well accept URLs there too in case it changes its
519 my $repo = $git_dir || $Job->{'repository'};
521 # Repository can be remote or local. If remote, we'll need to fetch it
522 # to a local dir before doing `git log` et al.
525 if ($repo =~ m{://|^[^/]*:}) {
526 # $repo is a git url we can clone, like git:// or https:// or
527 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
528 # not recognized here because distinguishing that from a local
529 # path is too fragile. If you really need something strange here,
530 # use the ssh:// form.
531 $repo_location = 'remote';
532 } elsif ($repo =~ m{^\.*/}) {
533 # $repo is a local path to a git index. We'll also resolve ../foo
534 # to ../foo/.git if the latter is a directory. To help
535 # disambiguate local paths from named hosted repositories, this
536 # form must be given as ./ or ../ if it's a relative path.
537 if (-d "$repo/.git") {
538 $repo = "$repo/.git";
540 $repo_location = 'local';
542 # $repo is none of the above. It must be the name of a hosted
544 my $arv_repo_list = api_call("repositories/list",
545 'filters' => [['name','=',$repo]]);
546 my @repos_found = @{$arv_repo_list->{'items'}};
547 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
549 Log(undef, "Repository '$repo' -> "
550 . join(", ", map { $_->{'uuid'} } @repos_found));
553 croak("Error: Found $n_found repositories with name '$repo'.");
555 $repo = $repos_found[0]->{'fetch_url'};
556 $repo_location = 'remote';
558 Log(undef, "Using $repo_location repository '$repo'");
559 $ENV{"CRUNCH_SRC_URL"} = $repo;
561 # Resolve given script_version (we'll call that $treeish here) to a
562 # commit sha1 ($commit).
563 my $treeish = $Job->{'script_version'};
565 if ($repo_location eq 'remote') {
566 # We minimize excess object-fetching by re-using the same bare
567 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
568 # just keep adding remotes to it as needed.
569 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
570 my $gitcmd = "git --git-dir=\Q$local_repo\E";
572 # Set up our local repo for caching remote objects, making
574 if (!-d $local_repo) {
575 make_path($local_repo) or croak("Error: could not create $local_repo");
577 # This works (exits 0 and doesn't delete fetched objects) even
578 # if $local_repo is already initialized:
579 `$gitcmd init --bare`;
581 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
584 # If $treeish looks like a hash (or abbrev hash) we look it up in
585 # our local cache first, since that's cheaper. (We don't want to
586 # do that with tags/branches though -- those change over time, so
587 # they should always be resolved by the remote repo.)
588 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
589 # Hide stderr because it's normal for this to fail:
590 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
592 # Careful not to resolve a branch named abcdeff to commit 1234567:
593 $sha1 =~ /^$treeish/ &&
594 $sha1 =~ /^([0-9a-f]{40})$/s) {
596 Log(undef, "Commit $commit already present in $local_repo");
600 if (!defined $commit) {
601 # If $treeish isn't just a hash or abbrev hash, or isn't here
602 # yet, we need to fetch the remote to resolve it correctly.
604 # First, remove all local heads. This prevents a name that does
605 # not exist on the remote from resolving to (or colliding with)
606 # a previously fetched branch or tag (possibly from a different
608 remove_tree("$local_repo/refs/heads", {keep_root => 1});
610 Log(undef, "Fetching objects from $repo to $local_repo");
611 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
613 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
617 # Now that the data is all here, we will use our local repo for
618 # the rest of our git activities.
622 my $gitcmd = "git --git-dir=\Q$repo\E";
623 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
624 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
625 croak("`$gitcmd rev-list` exited "
627 .", '$treeish' not found, giving up");
630 Log(undef, "Version $treeish is commit $commit");
632 if ($commit ne $Job->{'script_version'}) {
633 # Record the real commit id in the database, frozentokey, logs,
634 # etc. -- instead of an abbreviation or a branch name which can
635 # become ambiguous or point to a different commit in the future.
636 if (!$Job->update_attributes('script_version' => $commit)) {
637 croak("Error: failed to update job's script_version attribute");
641 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
642 add_git_archive("$gitcmd archive ''\Q$commit\E");
645 my $git_archive = combined_git_archive();
646 if (!defined $git_archive) {
647 Log(undef, "Skip install phase (no git archive)");
649 Log(undef, "Warning: This probably means workers have no source tree!");
654 my $install_script_tries_left = 3;
655 for (my $attempts = 0; $attempts < 3; $attempts++) {
656 my @srunargs = ("srun",
657 "--nodelist=$nodelist",
658 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
659 my @execargs = ("sh", "-c",
660 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
662 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
663 my ($stdout, $stderr);
664 ($exited, $stdout, $stderr) = srun_sync(
665 \@srunargs, \@execargs,
666 {label => "run install script on all workers"},
667 $build_script . $git_archive);
669 my $stderr_anything_from_script = 0;
670 for my $line (split(/\n/, $stderr)) {
671 if ($line !~ /^(srun: error: |starting: \[)/) {
672 $stderr_anything_from_script = 1;
676 last if $exited == 0 || $main::please_freeze;
678 # If the install script fails but doesn't print an error message,
679 # the next thing anyone is likely to do is just run it again in
680 # case it was a transient problem like "slurm communication fails
681 # because the network isn't reliable enough". So we'll just do
682 # that ourselves (up to 3 attempts in total). OTOH, if there is an
683 # error message, the problem is more likely to have a real fix and
684 # we should fail the job so the fixing process can start, instead
685 # of doing 2 more attempts.
686 last if $stderr_anything_from_script;
689 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
690 unlink($tar_filename);
698 foreach (qw (script script_version script_parameters runtime_constraints))
702 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
704 foreach (split (/\n/, $Job->{knobs}))
706 Log (undef, "knob " . $_);
711 $main::success = undef;
717 my $thisround_succeeded = 0;
718 my $thisround_failed = 0;
719 my $thisround_failed_multiple = 0;
720 my $working_slot_count = scalar(@slot);
722 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
723 or $a <=> $b } @jobstep_todo;
724 my $level = $jobstep[$jobstep_todo[0]]->{level};
726 my $initial_tasks_this_level = 0;
727 foreach my $id (@jobstep_todo) {
728 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
731 # If the number of tasks scheduled at this level #T is smaller than the number
732 # of slots available #S, only use the first #T slots, or the first slot on
733 # each node, whichever number is greater.
735 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
736 # based on these numbers. Using fewer slots makes more resources available
737 # to each individual task, which should normally be a better strategy when
738 # there are fewer of them running with less parallelism.
740 # Note that this calculation is not redone if the initial tasks at
741 # this level queue more tasks at the same level. This may harm
742 # overall task throughput for that level.
744 if ($initial_tasks_this_level < @node) {
745 @freeslot = (0..$#node);
746 } elsif ($initial_tasks_this_level < @slot) {
747 @freeslot = (0..$initial_tasks_this_level - 1);
749 @freeslot = (0..$#slot);
751 my $round_num_freeslots = scalar(@freeslot);
752 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";
754 my %round_max_slots = ();
755 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
756 my $this_slot = $slot[$freeslot[$ii]];
757 my $node_name = $this_slot->{node}->{name};
758 $round_max_slots{$node_name} ||= $this_slot->{cpu};
759 last if (scalar(keys(%round_max_slots)) >= @node);
762 Log(undef, "start level $level with $round_num_freeslots slots");
765 my $progress_is_dirty = 1;
766 my $progress_stats_updated = 0;
768 update_progress_stats();
772 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
774 # Don't create new tasks if we already know the job's final result.
775 last if defined($main::success);
777 my $id = $jobstep_todo[$todo_ptr];
778 my $Jobstep = $jobstep[$id];
779 if ($Jobstep->{level} != $level)
784 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
785 set_nonblocking($reader{$id});
787 my $childslot = $freeslot[0];
788 my $childnode = $slot[$childslot]->{node};
789 my $childslotname = join (".",
790 $slot[$childslot]->{node}->{name},
791 $slot[$childslot]->{cpu});
793 my $childpid = fork();
796 $SIG{'INT'} = 'DEFAULT';
797 $SIG{'QUIT'} = 'DEFAULT';
798 $SIG{'TERM'} = 'DEFAULT';
800 foreach (values (%reader))
804 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
805 open(STDOUT,">&writer");
806 open(STDERR,">&writer");
811 delete $ENV{"GNUPGHOME"};
812 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
813 $ENV{"TASK_QSEQUENCE"} = $id;
814 $ENV{"TASK_SEQUENCE"} = $level;
815 $ENV{"JOB_SCRIPT"} = $Job->{script};
816 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
817 $param =~ tr/a-z/A-Z/;
818 $ENV{"JOB_PARAMETER_$param"} = $value;
820 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
821 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
822 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
823 $ENV{"HOME"} = $ENV{"TASK_WORK"};
824 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
825 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
826 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
828 my $keep_mnt = $ENV{"TASK_WORK"}.".keep";
834 "--nodelist=".$childnode->{name},
835 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
836 "--job-name=$job_id.$id.$$",
839 my $stdbuf = " stdbuf --output=0 --error=0 ";
841 my $arv_file_cache = "";
842 if (defined($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'})) {
843 $arv_file_cache = "--file-cache=" . ($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'} * 1024 * 1024);
847 "if [ -e \Q$ENV{TASK_WORK}\E ]; then rm -rf \Q$ENV{TASK_WORK}\E; fi; "
848 ."mkdir -p \Q$ENV{CRUNCH_TMP}\E \Q$ENV{JOB_WORK}\E \Q$ENV{TASK_WORK}\E \Q$keep_mnt\E "
849 ."&& cd \Q$ENV{CRUNCH_TMP}\E "
850 # These environment variables get used explicitly later in
851 # $command. No tool is expected to read these values directly.
852 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
853 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
854 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
855 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
857 $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 ";
858 $ENV{TASK_KEEPMOUNT} = "$keep_mnt/by_pdh";
859 $ENV{TASK_KEEPMOUNT_TMP} = "$keep_mnt/tmp";
863 my $containername = "$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}";
864 my $cidfile = "$ENV{CRUNCH_TMP}/$containername.cid";
865 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
866 $command .= "$docker_bin run $docker_run_args --name=$containername --attach=stdout --attach=stderr --attach=stdin -i \Q$dockeruserarg\E --cidfile=$cidfile --sig-proxy ";
867 # We only set memory limits if Docker lets us limit both memory and swap.
868 # Memory limits alone have been supported longer, but subprocesses tend
869 # to get SIGKILL if they exceed that without any swap limit set.
870 # See #5642 for additional background.
871 if ($docker_limitmem) {
872 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
875 # The source tree and $destdir directory (which we have
876 # installed on the worker host) are available in the container,
877 # under the same path.
878 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
879 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
881 # Currently, we make the "by_pdh" directory in arv-mount's mount
882 # point appear at /keep inside the container (instead of using
883 # the same path as the host like we do with CRUNCH_SRC and
884 # CRUNCH_INSTALL). However, crunch scripts and utilities must
885 # not rely on this. They must use $TASK_KEEPMOUNT.
886 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
887 $ENV{TASK_KEEPMOUNT} = "/keep";
889 # Ditto TASK_KEEPMOUNT_TMP, as /keep_tmp.
890 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT_TMP}:/keep_tmp\E ";
891 $ENV{TASK_KEEPMOUNT_TMP} = "/keep_tmp";
893 # TASK_WORK is almost exactly like a docker data volume: it
894 # starts out empty, is writable, and persists until no
895 # containers use it any more. We don't use --volumes-from to
896 # share it with other containers: it is only accessible to this
897 # task, and it goes away when this task stops.
899 # However, a docker data volume is writable only by root unless
900 # the mount point already happens to exist in the container with
901 # different permissions. Therefore, we [1] assume /tmp already
902 # exists in the image and is writable by the crunch user; [2]
903 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
904 # writable if they are created by docker while setting up the
905 # other --volumes); and [3] create $TASK_WORK inside the
906 # container using $build_script.
907 $command .= "--volume=/tmp ";
908 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
909 $ENV{"HOME"} = $ENV{"TASK_WORK"};
910 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
912 # TODO: Share a single JOB_WORK volume across all task
913 # containers on a given worker node, and delete it when the job
914 # ends (and, in case that doesn't work, when the next job
917 # For now, use the same approach as TASK_WORK above.
918 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
920 while (my ($env_key, $env_val) = each %ENV)
922 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
923 $command .= "--env=\Q$env_key=$env_val\E ";
926 $command .= "--env=\QHOME=$ENV{HOME}\E ";
927 $command .= "\Q$docker_hash\E ";
929 if ($Job->{arvados_sdk_version}) {
931 $command .= "perl - \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E";
933 $command .= "/bin/sh -c \'python -c " .
934 '"from pkg_resources import get_distribution as get; print \"Using Arvados SDK version\", get(\"arvados-python-client\").version"' .
935 ">&2 2>/dev/null; " .
936 "mkdir -p \"$ENV{JOB_WORK}\" \"$ENV{TASK_WORK}\" && " .
937 "if which stdbuf >/dev/null ; then " .
938 " exec $stdbuf \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
940 " exec \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
945 $command .= "crunchstat -cgroup-root=\Q$cgroup_root\E -poll=10000 ";
947 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
950 my @execargs = ('bash', '-c', $command);
951 srun (\@srunargs, \@execargs, undef, $build_script);
952 # exec() failed, we assume nothing happened.
953 die "srun() failed on build script\n";
956 if (!defined $childpid)
967 jobstepname => "$job_id.$id.$childpid",
969 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
970 $slot[$childslot]->{pid} = $childpid;
972 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
973 Log ($id, "child $childpid started on $childslotname");
974 $Jobstep->{starttime} = time;
975 $Jobstep->{node} = $childnode->{name};
976 $Jobstep->{slotindex} = $childslot;
977 delete $Jobstep->{stderr};
978 delete $Jobstep->{finishtime};
979 delete $Jobstep->{tempfail};
981 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
982 $Jobstep->{'arvados_task'}->save;
984 splice @jobstep_todo, $todo_ptr, 1;
987 $progress_is_dirty = 1;
991 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
993 last THISROUND if $main::please_freeze;
994 if ($main::please_info)
996 $main::please_info = 0;
998 create_output_collection();
1000 update_progress_stats();
1005 if (!$gotsome || ($latest_refresh + 2 < scalar time))
1007 check_refresh_wanted();
1009 update_progress_stats();
1011 elsif (time - $progress_stats_updated >= 30 || $progress_is_dirty)
1013 update_progress_stats();
1016 select (undef, undef, undef, 0.1);
1018 $working_slot_count = scalar(grep { $_->{node}->{fail_count} == 0 &&
1019 $_->{node}->{hold_count} < 4 } @slot);
1020 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
1021 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
1023 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
1024 .($thisround_failed+$thisround_succeeded)
1025 .") -- giving up on this round";
1026 Log (undef, $message);
1030 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1031 for (my $i=$#freeslot; $i>=0; $i--) {
1032 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1033 push @holdslot, (splice @freeslot, $i, 1);
1036 for (my $i=$#holdslot; $i>=0; $i--) {
1037 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1038 push @freeslot, (splice @holdslot, $i, 1);
1042 # give up if no nodes are succeeding
1043 if ($working_slot_count < 1) {
1044 Log(undef, "Every node has failed -- giving up");
1051 push @freeslot, splice @holdslot;
1052 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1055 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1058 if ($main::please_continue) {
1059 $main::please_continue = 0;
1062 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1064 if (!reapchildren())
1066 check_refresh_wanted();
1068 update_progress_stats();
1069 select (undef, undef, undef, 0.1);
1070 killem (keys %proc) if $main::please_freeze;
1074 update_progress_stats();
1075 freeze_if_want_freeze();
1078 if (!defined $main::success)
1080 if (!@jobstep_todo) {
1082 } elsif ($working_slot_count < 1) {
1083 save_output_collection();
1085 exit(EX_RETRY_UNLOCKED);
1086 } elsif ($thisround_succeeded == 0 &&
1087 ($thisround_failed == 0 || $thisround_failed > 4)) {
1088 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1089 Log (undef, $message);
1094 goto ONELEVEL if !defined $main::success;
1097 release_allocation();
1099 my $collated_output = save_output_collection();
1100 Log (undef, "finish");
1105 if ($collated_output && $main::success) {
1106 $final_state = 'Complete';
1108 $final_state = 'Failed';
1110 $Job->update_attributes('state' => $final_state);
1112 exit (($final_state eq 'Complete') ? 0 : 1);
1116 sub update_progress_stats
1118 $progress_stats_updated = time;
1119 return if !$progress_is_dirty;
1120 my ($todo, $done, $running) = (scalar @jobstep_todo,
1121 scalar @jobstep_done,
1122 scalar keys(%proc));
1123 $Job->{'tasks_summary'} ||= {};
1124 $Job->{'tasks_summary'}->{'todo'} = $todo;
1125 $Job->{'tasks_summary'}->{'done'} = $done;
1126 $Job->{'tasks_summary'}->{'running'} = $running;
1127 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1128 Log (undef, "status: $done done, $running running, $todo todo");
1129 $progress_is_dirty = 0;
1136 my $children_reaped = 0;
1137 my @successful_task_uuids = ();
1139 while((my $pid = waitpid (-1, WNOHANG)) > 0)
1141 my $childstatus = $?;
1143 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1145 . $slot[$proc{$pid}->{slot}]->{cpu});
1146 my $jobstepidx = $proc{$pid}->{jobstepidx};
1148 if (!WIFEXITED($childstatus))
1150 # child did not exit (may be temporarily stopped)
1151 Log ($jobstepidx, "child $pid did not actually exit in reapchildren, ignoring for now.");
1156 my $elapsed = time - $proc{$pid}->{time};
1157 my $Jobstep = $jobstep[$jobstepidx];
1159 my $exitvalue = $childstatus >> 8;
1160 my $exitinfo = "exit ".exit_status_s($childstatus);
1161 $Jobstep->{'arvados_task'}->reload;
1162 my $task_success = $Jobstep->{'arvados_task'}->{success};
1164 Log ($jobstepidx, "child $pid on $whatslot $exitinfo success=$task_success");
1166 if (!defined $task_success) {
1167 # task did not indicate one way or the other --> fail
1168 Log($jobstepidx, sprintf(
1169 "ERROR: Task process exited %s, but never updated its task record to indicate success and record its output.",
1170 exit_status_s($childstatus)));
1171 $Jobstep->{'arvados_task'}->{success} = 0;
1172 $Jobstep->{'arvados_task'}->save;
1179 $temporary_fail ||= $Jobstep->{tempfail};
1180 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1182 ++$thisround_failed;
1183 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1185 # Check for signs of a failed or misconfigured node
1186 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1187 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1188 # Don't count this against jobstep failure thresholds if this
1189 # node is already suspected faulty and srun exited quickly
1190 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1192 Log ($jobstepidx, "blaming failure on suspect node " .
1193 $slot[$proc{$pid}->{slot}]->{node}->{name});
1194 $temporary_fail ||= 1;
1196 ban_node_by_slot($proc{$pid}->{slot});
1199 Log ($jobstepidx, sprintf('failure (#%d, %s) after %d seconds',
1200 ++$Jobstep->{'failures'},
1201 $temporary_fail ? 'temporary' : 'permanent',
1204 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1205 # Give up on this task, and the whole job
1208 # Put this task back on the todo queue
1209 push @jobstep_todo, $jobstepidx;
1210 $Job->{'tasks_summary'}->{'failed'}++;
1214 push @successful_task_uuids, $Jobstep->{'arvados_task'}->{uuid};
1215 ++$thisround_succeeded;
1216 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1217 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1218 $slot[$proc{$pid}->{slot}]->{node}->{fail_count} = 0;
1219 push @jobstep_done, $jobstepidx;
1220 Log ($jobstepidx, "success in $elapsed seconds");
1222 $Jobstep->{exitcode} = $childstatus;
1223 $Jobstep->{finishtime} = time;
1224 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1225 $Jobstep->{'arvados_task'}->save;
1226 process_stderr_final ($jobstepidx);
1227 Log ($jobstepidx, sprintf("task output (%d bytes): %s",
1228 length($Jobstep->{'arvados_task'}->{output}),
1229 $Jobstep->{'arvados_task'}->{output}));
1231 close $reader{$jobstepidx};
1232 delete $reader{$jobstepidx};
1233 delete $slot[$proc{$pid}->{slot}]->{pid};
1234 push @freeslot, $proc{$pid}->{slot};
1237 $progress_is_dirty = 1;
1240 if (scalar(@successful_task_uuids) > 0)
1242 Log (undef, sprintf("%d tasks exited (%d succeeded), checking for new tasks from API server.", $children_reaped, scalar(@successful_task_uuids)));
1244 my $newtask_list = [];
1245 my $newtask_results;
1247 $newtask_results = api_call(
1249 'filters' => [["created_by_job_task_uuid","in",\@successful_task_uuids]],
1250 'order' => 'qsequence',
1251 'offset' => scalar(@$newtask_list),
1253 push(@$newtask_list, @{$newtask_results->{items}});
1254 } while (@{$newtask_results->{items}});
1255 Log (undef, sprintf("Got %d new tasks from API server.", scalar(@$newtask_list)));
1256 foreach my $arvados_task (@$newtask_list) {
1258 'level' => $arvados_task->{'sequence'},
1260 'arvados_task' => $arvados_task
1262 push @jobstep, $jobstep;
1263 push @jobstep_todo, $#jobstep;
1267 return $children_reaped;
1270 sub check_refresh_wanted
1272 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1274 $stat[9] > $latest_refresh &&
1275 # ...and we have actually locked the job record...
1276 $job_id eq $Job->{'uuid'}) {
1277 $latest_refresh = scalar time;
1278 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1279 for my $attr ('cancelled_at',
1280 'cancelled_by_user_uuid',
1281 'cancelled_by_client_uuid',
1283 $Job->{$attr} = $Job2->{$attr};
1285 if ($Job->{'state'} ne "Running") {
1286 if ($Job->{'state'} eq "Cancelled") {
1287 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1289 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1292 $main::please_freeze = 1;
1299 my $last_squeue_check = $squeue_checked;
1301 # Do not call `squeue` or check the kill list more than once every
1303 return if $last_squeue_check > time - 15;
1304 $squeue_checked = time;
1306 # Look for children from which we haven't received stderr data since
1307 # the last squeue check. If no such children exist, all procs are
1308 # alive and there's no need to even look at squeue.
1310 # As long as the crunchstat poll interval (10s) is shorter than the
1311 # squeue check interval (15s) this should make the squeue check an
1313 my $silent_procs = 0;
1314 for my $js (map {$jobstep[$_->{jobstepidx}]} values %proc)
1316 if (!exists($js->{stderr_at}))
1318 $js->{stderr_at} = 0;
1320 if ($js->{stderr_at} < $last_squeue_check)
1325 return if $silent_procs == 0;
1327 # use killem() on procs whose killtime is reached
1328 while (my ($pid, $procinfo) = each %proc)
1330 my $js = $jobstep[$procinfo->{jobstepidx}];
1331 if (exists $procinfo->{killtime}
1332 && $procinfo->{killtime} <= time
1333 && $js->{stderr_at} < $last_squeue_check)
1336 if ($js->{stderr_at}) {
1337 $sincewhen = " in last " . (time - $js->{stderr_at}) . "s";
1339 Log($procinfo->{jobstepidx}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
1346 # here is an opportunity to check for mysterious problems with local procs
1350 # Get a list of steps still running. Note: squeue(1) says --steps
1351 # selects a format (which we override anyway) and allows us to
1352 # specify which steps we're interested in (which we don't).
1353 # Importantly, it also changes the meaning of %j from "job name" to
1354 # "step name" and (although this isn't mentioned explicitly in the
1355 # docs) switches from "one line per job" mode to "one line per step"
1356 # mode. Without it, we'd just get a list of one job, instead of a
1358 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%j' --noheader`;
1361 Log(undef, "warning: squeue exit status $? ($!)");
1366 # which of my jobsteps are running, according to squeue?
1368 for my $jobstepname (@squeue)
1370 $ok{$jobstepname} = 1;
1373 # Check for child procs >60s old and not mentioned by squeue.
1374 while (my ($pid, $procinfo) = each %proc)
1376 if ($procinfo->{time} < time - 60
1377 && $procinfo->{jobstepname}
1378 && !exists $ok{$procinfo->{jobstepname}}
1379 && !exists $procinfo->{killtime})
1381 # According to slurm, this task has ended (successfully or not)
1382 # -- but our srun child hasn't exited. First we must wait (30
1383 # seconds) in case this is just a race between communication
1384 # channels. Then, if our srun child process still hasn't
1385 # terminated, we'll conclude some slurm communication
1386 # error/delay has caused the task to die without notifying srun,
1387 # and we'll kill srun ourselves.
1388 $procinfo->{killtime} = time + 30;
1389 Log($procinfo->{jobstepidx}, "notice: task is not in slurm queue but srun process $pid has not exited");
1395 sub release_allocation
1399 Log (undef, "release job allocation");
1400 system "scancel $ENV{SLURM_JOB_ID}";
1409 my $sel = IO::Select->new();
1410 foreach my $jobstepidx (keys %reader)
1412 my $fd = $reader{$jobstepidx};
1414 $fd_job{$fd} = $jobstepidx;
1416 if (my $stdout_fd = $jobstep[$jobstepidx]->{stdout_r}) {
1417 $sel->add($stdout_fd);
1418 $fd_job{$stdout_fd} = $jobstepidx;
1421 # select on all reader fds with 0.1s timeout
1422 my @ready_fds = $sel->can_read(0.1);
1423 foreach my $fd (@ready_fds)
1426 if (0 < sysread ($fd, $buf, 65536))
1429 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1431 my $jobstepidx = $fd_job{$fd};
1432 if ($jobstep[$jobstepidx]->{stdout_r} == $fd) {
1433 $jobstep[$jobstepidx]->{stdout_captured} .= $buf;
1437 $jobstep[$jobstepidx]->{stderr_at} = time;
1438 $jobstep[$jobstepidx]->{stderr} .= $buf;
1440 # Consume everything up to the last \n
1441 preprocess_stderr ($jobstepidx);
1443 if (length ($jobstep[$jobstepidx]->{stderr}) > 16384)
1445 # If we get a lot of stderr without a newline, chop off the
1446 # front to avoid letting our buffer grow indefinitely.
1447 substr ($jobstep[$jobstepidx]->{stderr},
1448 0, length($jobstep[$jobstepidx]->{stderr}) - 8192) = "";
1456 # Consume all full lines of stderr for a jobstep. Everything after the
1457 # last newline will remain in $jobstep[$jobstepidx]->{stderr} after
1459 sub preprocess_stderr
1461 my $jobstepidx = shift;
1463 while ($jobstep[$jobstepidx]->{stderr} =~ /^(.*?)\n/) {
1465 substr $jobstep[$jobstepidx]->{stderr}, 0, 1+length($line), "";
1466 Log ($jobstepidx, "stderr $line");
1467 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1469 $main::please_freeze = 1;
1471 elsif (!exists $jobstep[$jobstepidx]->{slotindex}) {
1472 # Skip the following tempfail checks if this srun proc isn't
1473 # attached to a particular worker slot.
1475 elsif ($line =~ /srun: error: (Node failure on|Aborting, .*\bio error\b)/) {
1476 my $job_slot_index = $jobstep[$jobstepidx]->{slotindex};
1477 $slot[$job_slot_index]->{node}->{fail_count}++;
1478 $jobstep[$jobstepidx]->{tempfail} = 1;
1479 ban_node_by_slot($job_slot_index);
1481 elsif ($line =~ /srun: error: (Unable to create job step|.*: Communication connection failure)/) {
1482 $jobstep[$jobstepidx]->{tempfail} = 1;
1483 ban_node_by_slot($jobstep[$jobstepidx]->{slotindex});
1485 elsif ($line =~ /\bKeep(Read|Write|Request)Error:/) {
1486 $jobstep[$jobstepidx]->{tempfail} = 1;
1492 sub process_stderr_final
1494 my $jobstepidx = shift;
1495 preprocess_stderr ($jobstepidx);
1498 Log ($jobstepidx, "stderr $_");
1499 } split ("\n", $jobstep[$jobstepidx]->{stderr});
1500 $jobstep[$jobstepidx]->{stderr} = '';
1507 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1508 Log(undef, "fetch_block run error from arv-get $hash: $!");
1511 my $output_block = "";
1514 my $bytes = sysread($keep, $buf, 1024 * 1024);
1515 if (!defined $bytes) {
1516 Log(undef, "fetch_block read error from arv-get: $!");
1517 $output_block = undef;
1519 } elsif ($bytes == 0) {
1520 # sysread returns 0 at the end of the pipe.
1523 # some bytes were read into buf.
1524 $output_block .= $buf;
1529 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1530 $output_block = undef;
1532 return $output_block;
1535 # Create a collection by concatenating the output of all tasks (each
1536 # task's output is either a manifest fragment, a locator for a
1537 # manifest fragment stored in Keep, or nothing at all). Return the
1538 # portable_data_hash of the new collection.
1539 sub create_output_collection
1541 Log (undef, "collate");
1543 my ($child_out, $child_in);
1544 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1547 print (arvados.api("v1").collections().
1548 create(body={"manifest_text": sys.stdin.read()}).
1549 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1553 my $manifest_size = 0;
1557 my $output = $_->{'arvados_task'}->{output};
1558 next if (!defined($output));
1560 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1561 $next_write = fetch_block($output);
1563 $next_write = $output;
1565 if (defined($next_write)) {
1566 if (!defined(syswrite($child_in, $next_write))) {
1567 # There's been an error writing. Stop the loop.
1568 # We'll log details about the exit code later.
1571 $manifest_size += length($next_write);
1574 my $uuid = $_->{'arvados_task'}->{'uuid'};
1575 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1580 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1583 my $s = IO::Select->new($child_out);
1584 if ($s->can_read(120)) {
1585 sysread($child_out, $joboutput, 1024 * 1024);
1588 Log(undef, "output collection creation exited " . exit_status_s($?));
1594 Log (undef, "timed out while creating output collection");
1595 foreach my $signal (2, 2, 2, 15, 15, 9) {
1596 kill($signal, $pid);
1597 last if waitpid($pid, WNOHANG) == -1;
1606 # Calls create_output_collection, logs the result, and returns it.
1607 # If that was successful, save that as the output in the job record.
1608 sub save_output_collection {
1609 my $collated_output = create_output_collection();
1611 if (!$collated_output) {
1612 Log(undef, "Failed to write output collection");
1615 Log(undef, "job output $collated_output");
1616 $Job->update_attributes('output' => $collated_output);
1618 return $collated_output;
1625 my $sig = 2; # SIGINT first
1626 if (exists $proc{$_}->{"sent_$sig"} &&
1627 time - $proc{$_}->{"sent_$sig"} > 4)
1629 $sig = 15; # SIGTERM if SIGINT doesn't work
1631 if (exists $proc{$_}->{"sent_$sig"} &&
1632 time - $proc{$_}->{"sent_$sig"} > 4)
1634 $sig = 9; # SIGKILL if SIGTERM doesn't work
1636 if (!exists $proc{$_}->{"sent_$sig"})
1638 Log ($proc{$_}->{jobstepidx}, "sending 2x signal $sig to pid $_");
1640 select (undef, undef, undef, 0.1);
1643 kill $sig, $_; # srun wants two SIGINT to really interrupt
1645 $proc{$_}->{"sent_$sig"} = time;
1646 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1656 vec($bits,fileno($_),1) = 1;
1662 # Send log output to Keep via arv-put.
1664 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1665 # $log_pipe_out_buf is a string containing all output read from arv-put so far.
1666 # $log_pipe_out_select is an IO::Select object around $log_pipe_out.
1667 # $log_pipe_pid is the pid of the arv-put subprocess.
1669 # The only functions that should access these variables directly are:
1671 # log_writer_start($logfilename)
1672 # Starts an arv-put pipe, reading data on stdin and writing it to
1673 # a $logfilename file in an output collection.
1675 # log_writer_read_output([$timeout])
1676 # Read output from $log_pipe_out and append it to $log_pipe_out_buf.
1677 # Passes $timeout to the select() call, with a default of 0.01.
1678 # Returns the result of the last read() call on $log_pipe_out, or
1679 # -1 if read() wasn't called because select() timed out.
1680 # Only other log_writer_* functions should need to call this.
1682 # log_writer_send($txt)
1683 # Writes $txt to the output log collection.
1685 # log_writer_finish()
1686 # Closes the arv-put pipe and returns the output that it produces.
1688 # log_writer_is_active()
1689 # Returns a true value if there is currently a live arv-put
1690 # process, false otherwise.
1692 my ($log_pipe_in, $log_pipe_out, $log_pipe_out_buf, $log_pipe_out_select,
1695 sub log_writer_start($)
1697 my $logfilename = shift;
1698 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1702 '--filename', $logfilename,
1704 $log_pipe_out_buf = "";
1705 $log_pipe_out_select = IO::Select->new($log_pipe_out);
1708 sub log_writer_read_output {
1709 my $timeout = shift || 0.01;
1711 while ($read && $log_pipe_out_select->can_read($timeout)) {
1712 $read = read($log_pipe_out, $log_pipe_out_buf, 65536,
1713 length($log_pipe_out_buf));
1715 if (!defined($read)) {
1716 Log(undef, "error reading log manifest from arv-put: $!");
1721 sub log_writer_send($)
1724 print $log_pipe_in $txt;
1725 log_writer_read_output();
1728 sub log_writer_finish()
1730 return unless $log_pipe_pid;
1732 close($log_pipe_in);
1734 my $logger_failed = 0;
1735 my $read_result = log_writer_read_output(120);
1736 if ($read_result == -1) {
1737 $logger_failed = -1;
1738 Log (undef, "timed out reading from 'arv-put'");
1739 } elsif ($read_result != 0) {
1740 $logger_failed = -2;
1741 Log(undef, "failed to read arv-put log manifest to EOF");
1744 waitpid($log_pipe_pid, 0);
1746 $logger_failed ||= $?;
1747 Log(undef, "log_writer_finish: arv-put exited " . exit_status_s($?))
1750 close($log_pipe_out);
1751 my $arv_put_output = $logger_failed ? undef : $log_pipe_out_buf;
1752 $log_pipe_pid = $log_pipe_in = $log_pipe_out = $log_pipe_out_buf =
1753 $log_pipe_out_select = undef;
1755 return $arv_put_output;
1758 sub log_writer_is_active() {
1759 return $log_pipe_pid;
1762 sub Log # ($jobstepidx, $logmessage)
1764 my ($jobstepidx, $logmessage) = @_;
1765 if ($logmessage =~ /\n/) {
1766 for my $line (split (/\n/, $_[1])) {
1767 Log ($jobstepidx, $line);
1771 my $fh = select STDERR; $|=1; select $fh;
1773 if (defined($jobstepidx) && exists($jobstep[$jobstepidx]->{arvados_task})) {
1774 $task_qseq = $jobstepidx;
1776 my $message = sprintf ("%s %d %s %s", $job_id, $$, $task_qseq, $logmessage);
1777 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1780 if (log_writer_is_active() || -t STDERR) {
1781 my @gmtime = gmtime;
1782 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1783 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1785 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1787 if (log_writer_is_active()) {
1788 log_writer_send($datetime . " " . $message);
1795 my ($package, $file, $line) = caller;
1796 my $message = "@_ at $file line $line\n";
1797 Log (undef, $message);
1798 freeze() if @jobstep_todo;
1799 create_output_collection() if @jobstep_todo;
1809 if ($Job->{'state'} eq 'Cancelled') {
1810 $Job->update_attributes('finished_at' => scalar gmtime);
1812 $Job->update_attributes('state' => 'Failed');
1819 my $justcheckpoint = shift; # false if this will be the last meta saved
1820 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1821 return unless log_writer_is_active();
1822 my $log_manifest = log_writer_finish();
1823 return unless defined($log_manifest);
1826 my $prev_log_coll = api_call("collections/get", uuid => $Job->{log});
1827 $log_manifest = $prev_log_coll->{manifest_text} . $log_manifest;
1830 my $log_coll = api_call(
1831 "collections/create", ensure_unique_name => 1, collection => {
1832 manifest_text => $log_manifest,
1833 owner_uuid => $Job->{owner_uuid},
1834 name => sprintf("Log from %s job %s", $Job->{script}, $Job->{uuid}),
1836 Log(undef, "log collection is " . $log_coll->{portable_data_hash});
1837 $Job->update_attributes('log' => $log_coll->{portable_data_hash});
1841 sub freeze_if_want_freeze
1843 if ($main::please_freeze)
1845 release_allocation();
1848 # kill some srun procs before freeze+stop
1849 map { $proc{$_} = {} } @_;
1852 killem (keys %proc);
1853 select (undef, undef, undef, 0.1);
1855 while (($died = waitpid (-1, WNOHANG)) > 0)
1857 delete $proc{$died};
1862 create_output_collection();
1872 Log (undef, "Freeze not implemented");
1879 croak ("Thaw not implemented");
1895 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1902 my $srunargs = shift;
1903 my $execargs = shift;
1904 my $opts = shift || {};
1907 my $label = exists $opts->{label} ? $opts->{label} : "@$execargs";
1908 Log (undef, "$label: start");
1910 my ($stderr_r, $stderr_w);
1911 pipe $stderr_r, $stderr_w or croak("pipe() failed: $!");
1913 my ($stdout_r, $stdout_w);
1914 pipe $stdout_r, $stdout_w or croak("pipe() failed: $!");
1916 my $srunpid = fork();
1921 fcntl($stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
1922 fcntl($stdout_w, F_SETFL, 0) or croak($!);
1923 open(STDERR, ">&", $stderr_w);
1924 open(STDOUT, ">&", $stdout_w);
1925 srun ($srunargs, $execargs, $opts, $stdin);
1931 set_nonblocking($stderr_r);
1932 set_nonblocking($stdout_r);
1934 # Add entries to @jobstep and %proc so check_squeue() and
1935 # freeze_if_want_freeze() can treat it like a job task process.
1939 stderr_captured => '',
1940 stdout_r => $stdout_r,
1941 stdout_captured => '',
1943 my $jobstepidx = $#jobstep;
1945 jobstepidx => $jobstepidx,
1947 $reader{$jobstepidx} = $stderr_r;
1949 while ($srunpid != waitpid ($srunpid, WNOHANG)) {
1950 my $busy = readfrompipes();
1951 if (!$busy || ($latest_refresh + 2 < scalar time)) {
1952 check_refresh_wanted();
1956 select(undef, undef, undef, 0.1);
1958 killem(keys %proc) if $main::please_freeze;
1962 1 while readfrompipes();
1963 process_stderr_final ($jobstepidx);
1965 Log (undef, "$label: exit ".exit_status_s($exited));
1969 delete $proc{$srunpid};
1970 delete $reader{$jobstepidx};
1972 my $j = pop @jobstep;
1973 return ($exited, $j->{stdout_captured}, $j->{stderr_captured});
1979 my $srunargs = shift;
1980 my $execargs = shift;
1981 my $opts = shift || {};
1983 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1985 $Data::Dumper::Terse = 1;
1986 $Data::Dumper::Indent = 0;
1987 my $show_cmd = Dumper($args);
1988 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1989 $show_cmd =~ s/\n/ /g;
1990 if ($opts->{fork}) {
1991 Log(undef, "starting: $show_cmd");
1993 # This is a child process: parent is in charge of reading our
1994 # stderr and copying it to Log() if needed.
1995 warn "starting: $show_cmd\n";
1998 if (defined $stdin) {
1999 my $child = open STDIN, "-|";
2000 defined $child or die "no fork: $!";
2002 print $stdin or die $!;
2003 close STDOUT or die $!;
2008 return system (@$args) if $opts->{fork};
2011 warn "ENV size is ".length(join(" ",%ENV));
2012 die "exec failed: $!: @$args";
2016 sub ban_node_by_slot {
2017 # Don't start any new jobsteps on this node for 60 seconds
2019 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
2020 $slot[$slotid]->{node}->{hold_count}++;
2021 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
2026 my ($lockfile, $error_message) = @_;
2027 open L, ">", $lockfile or croak("$lockfile: $!");
2028 if (!flock L, LOCK_EX|LOCK_NB) {
2029 croak("Can't lock $lockfile: $error_message\n");
2033 sub find_docker_image {
2034 # Given a Keep locator, check to see if it contains a Docker image.
2035 # If so, return its stream name and Docker hash.
2036 # If not, return undef for both values.
2037 my $locator = shift;
2038 my ($streamname, $filename);
2039 my $image = api_call("collections/get", uuid => $locator);
2041 foreach my $line (split(/\n/, $image->{manifest_text})) {
2042 my @tokens = split(/\s+/, $line);
2044 $streamname = shift(@tokens);
2045 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
2046 if (defined($filename)) {
2047 return (undef, undef); # More than one file in the Collection.
2049 $filename = (split(/:/, $filedata, 3))[2];
2054 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
2055 return ($streamname, $1);
2057 return (undef, undef);
2062 # Calculate the number of times an operation should be retried,
2063 # assuming exponential backoff, and that we're willing to retry as
2064 # long as tasks have been running. Enforce a minimum of 3 retries.
2065 my ($starttime, $endtime, $timediff, $retries);
2067 $starttime = $jobstep[0]->{starttime};
2068 $endtime = $jobstep[-1]->{finishtime};
2070 if (!defined($starttime)) {
2072 } elsif (!defined($endtime)) {
2073 $timediff = time - $starttime;
2075 $timediff = ($endtime - $starttime) - (time - $endtime);
2077 if ($timediff > 0) {
2078 $retries = int(log($timediff) / log(2));
2080 $retries = 1; # Use the minimum.
2082 return ($retries > 3) ? $retries : 3;
2086 # Pass in two function references.
2087 # This method will be called with the remaining arguments.
2088 # If it dies, retry it with exponential backoff until it succeeds,
2089 # or until the current retry_count is exhausted. After each failure
2090 # that can be retried, the second function will be called with
2091 # the current try count (0-based), next try time, and error message.
2092 my $operation = shift;
2093 my $retry_callback = shift;
2094 my $retries = retry_count();
2095 foreach my $try_count (0..$retries) {
2096 my $next_try = time + (2 ** $try_count);
2097 my $result = eval { $operation->(@_); };
2100 } elsif ($try_count < $retries) {
2101 $retry_callback->($try_count, $next_try, $@);
2102 my $sleep_time = $next_try - time;
2103 sleep($sleep_time) if ($sleep_time > 0);
2106 # Ensure the error message ends in a newline, so Perl doesn't add
2107 # retry_op's line number to it.
2113 # Pass in a /-separated API method name, and arguments for it.
2114 # This function will call that method, retrying as needed until
2115 # the current retry_count is exhausted, with a log on the first failure.
2116 my $method_name = shift;
2117 my $log_api_retry = sub {
2118 my ($try_count, $next_try_at, $errmsg) = @_;
2119 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
2120 $errmsg =~ s/\s/ /g;
2121 $errmsg =~ s/\s+$//;
2123 if ($next_try_at < time) {
2124 $retry_msg = "Retrying.";
2126 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
2127 $retry_msg = "Retrying at $next_try_fmt.";
2129 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
2132 foreach my $key (split(/\//, $method_name)) {
2133 $method = $method->{$key};
2135 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
2139 # Given a $?, return a human-readable exit code string like "0" or
2140 # "1" or "0 with signal 1" or "1 with signal 11".
2141 my $exitcode = shift;
2142 my $s = $exitcode >> 8;
2143 if ($exitcode & 0x7f) {
2144 $s .= " with signal " . ($exitcode & 0x7f);
2146 if ($exitcode & 0x80) {
2147 $s .= " with core dump";
2152 sub handle_readall {
2153 # Pass in a glob reference to a file handle.
2154 # Read all its contents and return them as a string.
2155 my $fh_glob_ref = shift;
2157 return <$fh_glob_ref>;
2160 sub tar_filename_n {
2162 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
2165 sub add_git_archive {
2166 # Pass in a git archive command as a string or list, a la system().
2167 # This method will save its output to be included in the archive sent to the
2171 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
2172 croak("Failed to save git archive: $!");
2174 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
2176 waitpid($git_pid, 0);
2179 croak("Failed to save git archive: git exited " . exit_status_s($?));
2183 sub combined_git_archive {
2184 # Combine all saved tar archives into a single archive, then return its
2185 # contents in a string. Return undef if no archives have been saved.
2186 if ($git_tar_count < 1) {
2189 my $base_tar_name = tar_filename_n(1);
2190 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
2191 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
2192 if ($tar_exit != 0) {
2193 croak("Error preparing build archive: tar -A exited " .
2194 exit_status_s($tar_exit));
2197 if (!open(GIT_TAR, "<", $base_tar_name)) {
2198 croak("Could not open build archive: $!");
2200 my $tar_contents = handle_readall(\*GIT_TAR);
2202 return $tar_contents;
2205 sub set_nonblocking {
2207 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
2208 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
2214 # This is crunch-job's internal dispatch script. crunch-job running on the API
2215 # server invokes this script on individual compute nodes, or localhost if we're
2216 # running a job locally. It gets called in two modes:
2218 # * No arguments: Installation mode. Read a tar archive from the DATA
2219 # file handle; it includes the Crunch script's source code, and
2220 # maybe SDKs as well. Those should be installed in the proper
2221 # locations. This runs outside of any Docker container, so don't try to
2222 # introspect Crunch's runtime environment.
2224 # * With arguments: Crunch script run mode. This script should set up the
2225 # environment, then run the command specified in the arguments. This runs
2226 # inside any Docker container.
2229 use File::Path qw( make_path remove_tree );
2230 use POSIX qw(getcwd);
2232 use constant TASK_TEMPFAIL => 111;
2234 # Map SDK subdirectories to the path environments they belong to.
2235 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2237 my $destdir = $ENV{"CRUNCH_SRC"};
2238 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2239 my $repo = $ENV{"CRUNCH_SRC_URL"};
2240 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2241 my $job_work = $ENV{"JOB_WORK"};
2242 my $task_work = $ENV{"TASK_WORK"};
2244 open(STDOUT_ORIG, ">&", STDOUT);
2245 open(STDERR_ORIG, ">&", STDERR);
2247 for my $dir ($destdir, $job_work, $task_work) {
2250 -e $dir or die "Failed to create temporary directory ($dir): $!";
2255 remove_tree($task_work, {keep_root => 1});
2258 ### Crunch script run mode
2260 # We want to do routine logging during task 0 only. This gives the user
2261 # the information they need, but avoids repeating the information for every
2264 if ($ENV{TASK_SEQUENCE} eq "0") {
2267 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2273 my $python_src = "$install_dir/python";
2274 my $venv_dir = "$job_work/.arvados.venv";
2275 my $venv_built = -e "$venv_dir/bin/activate";
2276 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2277 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2278 "--python=python2.7", $venv_dir);
2279 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2281 $Log->("Built Python SDK virtualenv");
2284 my @pysdk_version_cmd = ("python", "-c",
2285 "from pkg_resources import get_distribution as get; print get('arvados-python-client').version");
2287 $Log->("Running in Python SDK virtualenv");
2288 @pysdk_version_cmd = ();
2289 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2290 @ARGV = ("/bin/sh", "-ec",
2291 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2292 } elsif (-d $python_src) {
2293 $Log->("Warning: virtualenv not found inside Docker container default " .
2294 "\$PATH. Can't install Python SDK.");
2297 if (@pysdk_version_cmd) {
2298 open(my $pysdk_version_pipe, "-|", @pysdk_version_cmd);
2299 my $pysdk_version = <$pysdk_version_pipe>;
2300 close($pysdk_version_pipe);
2302 chomp($pysdk_version);
2303 $Log->("Using Arvados SDK version $pysdk_version");
2305 # A lot could've gone wrong here, but pretty much all of it means that
2306 # Python won't be able to load the Arvados SDK.
2307 $Log->("Warning: Arvados SDK not found");
2311 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2312 my $sdk_path = "$install_dir/$sdk_dir";
2314 if ($ENV{$sdk_envkey}) {
2315 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2317 $ENV{$sdk_envkey} = $sdk_path;
2319 $Log->("Arvados SDK added to %s", $sdk_envkey);
2324 die "Cannot exec `@ARGV`: $!";
2327 ### Installation mode
2328 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2330 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2331 # This exact git archive (source + arvados sdk) is already installed
2332 # here, so there's no need to reinstall it.
2334 # We must consume our DATA section, though: otherwise the process
2335 # feeding it to us will get SIGPIPE.
2337 while (read(DATA, $buf, 65536)) { }
2342 unlink "$destdir.archive_hash";
2346 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2347 local $SIG{PIPE} = "IGNORE";
2348 warn "Extracting archive: $archive_hash\n";
2349 # --ignore-zeros is necessary sometimes: depending on how much NUL
2350 # padding tar -A put on our combined archive (which in turn depends
2351 # on the length of the component archives) tar without
2352 # --ignore-zeros will exit before consuming stdin and cause close()
2353 # to fail on the resulting SIGPIPE.
2354 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2355 die "Error launching 'tar -xC $destdir': $!";
2357 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2358 # get SIGPIPE. We must feed it data incrementally.
2360 while (read(DATA, $tar_input, 65536)) {
2361 print TARX $tar_input;
2364 die "'tar -xC $destdir' exited $?: $!";
2370 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2372 foreach my $sdk_lang (("python",
2373 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2374 if (-d "$sdk_root/$sdk_lang") {
2375 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2376 die "Failed to install $sdk_lang SDK: $!";
2382 my $python_dir = "$install_dir/python";
2383 if ((-d $python_dir) and can_run("python2.7")) {
2384 open(my $egg_info_pipe, "-|",
2385 "python2.7 \Q$python_dir/setup.py\E egg_info 2>&1 >/dev/null");
2386 my @egg_info_errors = <$egg_info_pipe>;
2387 close($egg_info_pipe);
2390 if (@egg_info_errors and (($egg_info_errors[-1] =~ /\bgit\b/) or ($egg_info_errors[-1] =~ /\[Errno 2\]/))) {
2391 # egg_info apparently failed because it couldn't ask git for a build tag.
2392 # Specify no build tag.
2393 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2394 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2397 my $egg_info_exit = $? >> 8;
2398 foreach my $errline (@egg_info_errors) {
2401 warn "python setup.py egg_info failed: exit $egg_info_exit";
2402 exit ($egg_info_exit || 1);
2407 # Hide messages from the install script (unless it fails: shell_or_die
2408 # will show $destdir.log in that case).
2409 open(STDOUT, ">>", "$destdir.log");
2410 open(STDERR, ">&", STDOUT);
2412 if (-e "$destdir/crunch_scripts/install") {
2413 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2414 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2416 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2417 } elsif (-e "./install.sh") {
2418 shell_or_die (undef, "./install.sh", $install_dir);
2421 if ($archive_hash) {
2422 unlink "$destdir.archive_hash.new";
2423 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2424 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2430 my $command_name = shift;
2431 open(my $which, "-|", "which", $command_name);
2432 while (<$which>) { }
2439 my $exitcode = shift;
2441 if ($ENV{"DEBUG"}) {
2442 print STDERR "@_\n";
2444 if (system (@_) != 0) {
2447 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2448 open STDERR, ">&STDERR_ORIG";
2449 system ("cat $destdir.log >&2");
2450 warn "@_ failed ($err): $exitstatus";
2451 if (defined($exitcode)) {
2455 exit (($code >> 8) || 1);