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;
102 $ENV{"TMPDIR"} ||= "/tmp";
103 unless (defined $ENV{"CRUNCH_TMP"}) {
104 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
105 if ($ENV{"USER"} ne "crunch" && $< != 0) {
106 # use a tmp dir unique for my uid
107 $ENV{"CRUNCH_TMP"} .= "-$<";
111 # Create the tmp directory if it does not exist
112 if ( ! -d $ENV{"CRUNCH_TMP"} ) {
113 make_path $ENV{"CRUNCH_TMP"} or die "Failed to create temporary working directory: " . $ENV{"CRUNCH_TMP"};
116 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
117 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
118 $ENV{"CRUNCH_WORK"} = $ENV{"JOB_WORK"}; # deprecated
119 mkdir ($ENV{"JOB_WORK"});
127 GetOptions('force-unlock' => \$force_unlock,
128 'git-dir=s' => \$git_dir,
129 'job=s' => \$jobspec,
130 'job-api-token=s' => \$job_api_token,
131 'no-clear-tmp' => \$no_clear_tmp,
132 'resume-stash=s' => \$resume_stash,
135 if (defined $job_api_token) {
136 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
139 my $have_slurm = exists $ENV{SLURM_JOBID} && exists $ENV{SLURM_NODELIST};
145 $main::ENV{CRUNCH_DEBUG} = 1;
149 $main::ENV{CRUNCH_DEBUG} = 0;
154 my $arv = Arvados->new('apiVersion' => 'v1');
162 my $User = api_call("users/current");
164 if ($jobspec =~ /^[-a-z\d]+$/)
166 # $jobspec is an Arvados UUID, not a JSON job specification
167 $Job = api_call("jobs/get", uuid => $jobspec);
168 if (!$force_unlock) {
169 # Claim this job, and make sure nobody else does
170 eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
172 Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
179 $Job = JSON::decode_json($jobspec);
183 map { croak ("No $_ specified") unless $Job->{$_} }
184 qw(script script_version script_parameters);
187 $Job->{'is_locked_by_uuid'} = $User->{'uuid'};
188 $Job->{'started_at'} = gmtime;
189 $Job->{'state'} = 'Running';
191 $Job = api_call("jobs/create", job => $Job);
193 $job_id = $Job->{'uuid'};
195 my $keep_logfile = $job_id . '.log.txt';
196 log_writer_start($keep_logfile);
198 $Job->{'runtime_constraints'} ||= {};
199 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
200 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
202 my $gem_versions = `gem list --quiet arvados-cli 2>/dev/null`;
204 $gem_versions =~ s/^arvados-cli \(/ with arvados-cli Gem version(s) /;
205 chomp($gem_versions);
206 chop($gem_versions); # Closing parentheses
211 "running from " . ((-e $0) ? realpath($0) : "stdin") . $gem_versions);
213 Log (undef, "check slurm allocation");
216 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
220 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
221 push @sinfo, "$localcpus localhost";
223 if (exists $ENV{SLURM_NODELIST})
225 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
229 my ($ncpus, $slurm_nodelist) = split;
230 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
233 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
236 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
239 foreach (split (",", $ranges))
252 push @nodelist, map {
254 $n =~ s/\[[-,\d]+\]/$_/;
261 push @nodelist, $nodelist;
264 foreach my $nodename (@nodelist)
266 Log (undef, "node $nodename - $ncpus slots");
267 my $node = { name => $nodename,
271 foreach my $cpu (1..$ncpus)
273 push @slot, { node => $node,
277 push @node, @nodelist;
282 # Ensure that we get one jobstep running on each allocated node before
283 # we start overloading nodes with concurrent steps
285 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
288 $Job->update_attributes(
289 'tasks_summary' => { 'failed' => 0,
294 Log (undef, "start");
295 $SIG{'INT'} = sub { $main::please_freeze = 1; };
296 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
297 $SIG{'TERM'} = \&croak;
298 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
299 $SIG{'ALRM'} = sub { $main::please_info = 1; };
300 $SIG{'CONT'} = sub { $main::please_continue = 1; };
301 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
303 $main::please_freeze = 0;
304 $main::please_info = 0;
305 $main::please_continue = 0;
306 $main::please_refresh = 0;
307 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
309 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
310 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
311 $ENV{"JOB_UUID"} = $job_id;
314 my @jobstep_todo = ();
315 my @jobstep_done = ();
316 my @jobstep_tomerge = ();
317 my $jobstep_tomerge_level = 0;
319 my $squeue_kill_checked;
320 my $latest_refresh = scalar time;
324 if (defined $Job->{thawedfromkey})
326 thaw ($Job->{thawedfromkey});
330 my $first_task = api_call("job_tasks/create", job_task => {
331 'job_uuid' => $Job->{'uuid'},
336 push @jobstep, { 'level' => 0,
338 'arvados_task' => $first_task,
340 push @jobstep_todo, 0;
346 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
349 my $build_script = handle_readall(\*DATA);
350 my $nodelist = join(",", @node);
351 my $git_tar_count = 0;
353 if (!defined $no_clear_tmp) {
354 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
355 Log (undef, "Clean work dirs");
357 my $cleanpid = fork();
360 # Find FUSE mounts that look like Keep mounts (the mount path has the
361 # word "keep") and unmount them. Then clean up work directories.
362 # TODO: When #5036 is done and widely deployed, we can get rid of the
363 # regular expression and just unmount everything with type fuse.keep.
364 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
365 ['bash', '-ec', 'mount -t fuse,fuse.keep | awk \'($3 ~ /\ykeep\y/){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']);
370 last if $cleanpid == waitpid (-1, WNOHANG);
371 freeze_if_want_freeze ($cleanpid);
372 select (undef, undef, undef, 0.1);
374 Log (undef, "Cleanup command exited ".exit_status_s($?));
377 # If this job requires a Docker image, install that.
378 my $docker_bin = "/usr/bin/docker.io";
379 my ($docker_locator, $docker_stream, $docker_hash);
380 if ($docker_locator = $Job->{docker_image_locator}) {
381 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
384 croak("No Docker image hash found from locator $docker_locator");
386 $docker_stream =~ s/^\.//;
387 my $docker_install_script = qq{
388 if ! $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
389 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
392 my $docker_pid = fork();
393 if ($docker_pid == 0)
395 srun (["srun", "--nodelist=" . join(',', @node)],
396 ["/bin/sh", "-ec", $docker_install_script]);
401 last if $docker_pid == waitpid (-1, WNOHANG);
402 freeze_if_want_freeze ($docker_pid);
403 select (undef, undef, undef, 0.1);
407 croak("Installing Docker image from $docker_locator exited "
411 if ($Job->{arvados_sdk_version}) {
412 # The job also specifies an Arvados SDK version. Add the SDKs to the
413 # tar file for the build script to install.
414 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
415 $Job->{arvados_sdk_version}));
416 add_git_archive("git", "--git-dir=$git_dir", "archive",
417 "--prefix=.arvados.sdk/",
418 $Job->{arvados_sdk_version}, "sdk");
422 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
423 # If script_version looks like an absolute path, *and* the --git-dir
424 # argument was not given -- which implies we were not invoked by
425 # crunch-dispatch -- we will use the given path as a working
426 # directory instead of resolving script_version to a git commit (or
427 # doing anything else with git).
428 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
429 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
432 # Resolve the given script_version to a git commit sha1. Also, if
433 # the repository is remote, clone it into our local filesystem: this
434 # ensures "git archive" will work, and is necessary to reliably
435 # resolve a symbolic script_version like "master^".
436 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
438 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
440 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
442 # If we're running under crunch-dispatch, it will have already
443 # pulled the appropriate source tree into its own repository, and
444 # given us that repo's path as $git_dir.
446 # If we're running a "local" job, we might have to fetch content
447 # from a remote repository.
449 # (Currently crunch-dispatch gives a local path with --git-dir, but
450 # we might as well accept URLs there too in case it changes its
452 my $repo = $git_dir || $Job->{'repository'};
454 # Repository can be remote or local. If remote, we'll need to fetch it
455 # to a local dir before doing `git log` et al.
458 if ($repo =~ m{://|^[^/]*:}) {
459 # $repo is a git url we can clone, like git:// or https:// or
460 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
461 # not recognized here because distinguishing that from a local
462 # path is too fragile. If you really need something strange here,
463 # use the ssh:// form.
464 $repo_location = 'remote';
465 } elsif ($repo =~ m{^\.*/}) {
466 # $repo is a local path to a git index. We'll also resolve ../foo
467 # to ../foo/.git if the latter is a directory. To help
468 # disambiguate local paths from named hosted repositories, this
469 # form must be given as ./ or ../ if it's a relative path.
470 if (-d "$repo/.git") {
471 $repo = "$repo/.git";
473 $repo_location = 'local';
475 # $repo is none of the above. It must be the name of a hosted
477 my $arv_repo_list = api_call("repositories/list",
478 'filters' => [['name','=',$repo]]);
479 my @repos_found = @{$arv_repo_list->{'items'}};
480 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
482 Log(undef, "Repository '$repo' -> "
483 . join(", ", map { $_->{'uuid'} } @repos_found));
486 croak("Error: Found $n_found repositories with name '$repo'.");
488 $repo = $repos_found[0]->{'fetch_url'};
489 $repo_location = 'remote';
491 Log(undef, "Using $repo_location repository '$repo'");
492 $ENV{"CRUNCH_SRC_URL"} = $repo;
494 # Resolve given script_version (we'll call that $treeish here) to a
495 # commit sha1 ($commit).
496 my $treeish = $Job->{'script_version'};
498 if ($repo_location eq 'remote') {
499 # We minimize excess object-fetching by re-using the same bare
500 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
501 # just keep adding remotes to it as needed.
502 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
503 my $gitcmd = "git --git-dir=\Q$local_repo\E";
505 # Set up our local repo for caching remote objects, making
507 if (!-d $local_repo) {
508 make_path($local_repo) or croak("Error: could not create $local_repo");
510 # This works (exits 0 and doesn't delete fetched objects) even
511 # if $local_repo is already initialized:
512 `$gitcmd init --bare`;
514 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
517 # If $treeish looks like a hash (or abbrev hash) we look it up in
518 # our local cache first, since that's cheaper. (We don't want to
519 # do that with tags/branches though -- those change over time, so
520 # they should always be resolved by the remote repo.)
521 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
522 # Hide stderr because it's normal for this to fail:
523 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
525 # Careful not to resolve a branch named abcdeff to commit 1234567:
526 $sha1 =~ /^$treeish/ &&
527 $sha1 =~ /^([0-9a-f]{40})$/s) {
529 Log(undef, "Commit $commit already present in $local_repo");
533 if (!defined $commit) {
534 # If $treeish isn't just a hash or abbrev hash, or isn't here
535 # yet, we need to fetch the remote to resolve it correctly.
537 # First, remove all local heads. This prevents a name that does
538 # not exist on the remote from resolving to (or colliding with)
539 # a previously fetched branch or tag (possibly from a different
541 remove_tree("$local_repo/refs/heads", {keep_root => 1});
543 Log(undef, "Fetching objects from $repo to $local_repo");
544 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
546 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
550 # Now that the data is all here, we will use our local repo for
551 # the rest of our git activities.
555 my $gitcmd = "git --git-dir=\Q$repo\E";
556 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
557 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
558 croak("`$gitcmd rev-list` exited "
560 .", '$treeish' not found. Giving up.");
563 Log(undef, "Version $treeish is commit $commit");
565 if ($commit ne $Job->{'script_version'}) {
566 # Record the real commit id in the database, frozentokey, logs,
567 # etc. -- instead of an abbreviation or a branch name which can
568 # become ambiguous or point to a different commit in the future.
569 if (!$Job->update_attributes('script_version' => $commit)) {
570 croak("Error: failed to update job's script_version attribute");
574 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
575 add_git_archive("$gitcmd archive ''\Q$commit\E");
578 my $git_archive = combined_git_archive();
579 if (!defined $git_archive) {
580 Log(undef, "Skip install phase (no git archive)");
582 Log(undef, "Warning: This probably means workers have no source tree!");
586 Log(undef, "Run install script on all workers");
588 my @srunargs = ("srun",
589 "--nodelist=$nodelist",
590 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
591 my @execargs = ("sh", "-c",
592 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
594 my $installpid = fork();
595 if ($installpid == 0)
597 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
602 last if $installpid == waitpid (-1, WNOHANG);
603 freeze_if_want_freeze ($installpid);
604 select (undef, undef, undef, 0.1);
606 my $install_exited = $?;
607 Log (undef, "Install script exited ".exit_status_s($install_exited));
608 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
609 unlink($tar_filename);
611 exit (1) if $install_exited != 0;
614 foreach (qw (script script_version script_parameters runtime_constraints))
618 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
620 foreach (split (/\n/, $Job->{knobs}))
622 Log (undef, "knob " . $_);
627 $main::success = undef;
633 my $thisround_succeeded = 0;
634 my $thisround_failed = 0;
635 my $thisround_failed_multiple = 0;
637 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
638 or $a <=> $b } @jobstep_todo;
639 my $level = $jobstep[$jobstep_todo[0]]->{level};
640 Log (undef, "start level $level");
645 my @freeslot = (0..$#slot);
648 my $progress_is_dirty = 1;
649 my $progress_stats_updated = 0;
651 update_progress_stats();
656 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
658 my $id = $jobstep_todo[$todo_ptr];
659 my $Jobstep = $jobstep[$id];
660 if ($Jobstep->{level} != $level)
665 pipe $reader{$id}, "writer" or croak ($!);
666 my $flags = fcntl ($reader{$id}, F_GETFL, 0) or croak ($!);
667 fcntl ($reader{$id}, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
669 my $childslot = $freeslot[0];
670 my $childnode = $slot[$childslot]->{node};
671 my $childslotname = join (".",
672 $slot[$childslot]->{node}->{name},
673 $slot[$childslot]->{cpu});
675 my $childpid = fork();
678 $SIG{'INT'} = 'DEFAULT';
679 $SIG{'QUIT'} = 'DEFAULT';
680 $SIG{'TERM'} = 'DEFAULT';
682 foreach (values (%reader))
686 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
687 open(STDOUT,">&writer");
688 open(STDERR,">&writer");
693 delete $ENV{"GNUPGHOME"};
694 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
695 $ENV{"TASK_QSEQUENCE"} = $id;
696 $ENV{"TASK_SEQUENCE"} = $level;
697 $ENV{"JOB_SCRIPT"} = $Job->{script};
698 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
699 $param =~ tr/a-z/A-Z/;
700 $ENV{"JOB_PARAMETER_$param"} = $value;
702 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
703 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
704 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
705 $ENV{"HOME"} = $ENV{"TASK_WORK"};
706 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
707 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
708 $ENV{"CRUNCH_NODE_SLOTS"} = $slot[$childslot]->{node}->{ncpus};
709 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
715 "--nodelist=".$childnode->{name},
716 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
717 "--job-name=$job_id.$id.$$",
720 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
721 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
722 ."&& cd $ENV{CRUNCH_TMP} "
723 ."&& MEM=\$(cat /proc/meminfo | grep MemTotal | sed 's/\\s\\s*/ /g' |cut -d' ' -f2) "
724 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) ";
725 $command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
728 my $cidfile = "$ENV{CRUNCH_TMP}/$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}.cid";
729 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
730 $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy --memory=\${MEMLIMIT}k ";
732 # Dynamically configure the container to use the host system as its
733 # DNS server. Get the host's global addresses from the ip command,
734 # and turn them into docker --dns options using gawk.
736 q{$(ip -o address show scope global |
737 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
739 # The source tree and $destdir directory (which we have
740 # installed on the worker host) are available in the container,
741 # under the same path.
742 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
743 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
745 # Currently, we make arv-mount's mount point appear at /keep
746 # inside the container (instead of using the same path as the
747 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
748 # crunch scripts and utilities must not rely on this. They must
749 # use $TASK_KEEPMOUNT.
750 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
751 $ENV{TASK_KEEPMOUNT} = "/keep";
753 # TASK_WORK is almost exactly like a docker data volume: it
754 # starts out empty, is writable, and persists until no
755 # containers use it any more. We don't use --volumes-from to
756 # share it with other containers: it is only accessible to this
757 # task, and it goes away when this task stops.
759 # However, a docker data volume is writable only by root unless
760 # the mount point already happens to exist in the container with
761 # different permissions. Therefore, we [1] assume /tmp already
762 # exists in the image and is writable by the crunch user; [2]
763 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
764 # writable if they are created by docker while setting up the
765 # other --volumes); and [3] create $TASK_WORK inside the
766 # container using $build_script.
767 $command .= "--volume=/tmp ";
768 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
769 $ENV{"HOME"} = $ENV{"TASK_WORK"};
770 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
772 # TODO: Share a single JOB_WORK volume across all task
773 # containers on a given worker node, and delete it when the job
774 # ends (and, in case that doesn't work, when the next job
777 # For now, use the same approach as TASK_WORK above.
778 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
780 while (my ($env_key, $env_val) = each %ENV)
782 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
783 $command .= "--env=\Q$env_key=$env_val\E ";
786 $command .= "--env=\QHOME=$ENV{HOME}\E ";
787 $command .= "\Q$docker_hash\E ";
788 $command .= "stdbuf --output=0 --error=0 ";
789 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
792 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
793 $command .= "stdbuf --output=0 --error=0 ";
794 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
797 my @execargs = ('bash', '-c', $command);
798 srun (\@srunargs, \@execargs, undef, $build_script);
799 # exec() failed, we assume nothing happened.
800 die "srun() failed on build script\n";
803 if (!defined $childpid)
810 $proc{$childpid} = { jobstep => $id,
813 jobstepname => "$job_id.$id.$childpid",
815 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
816 $slot[$childslot]->{pid} = $childpid;
818 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
819 Log ($id, "child $childpid started on $childslotname");
820 $Jobstep->{starttime} = time;
821 $Jobstep->{node} = $childnode->{name};
822 $Jobstep->{slotindex} = $childslot;
823 delete $Jobstep->{stderr};
824 delete $Jobstep->{finishtime};
826 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
827 $Jobstep->{'arvados_task'}->save;
829 splice @jobstep_todo, $todo_ptr, 1;
832 $progress_is_dirty = 1;
836 (@slot > @freeslot && $todo_ptr+1 > $#jobstep_todo))
838 last THISROUND if $main::please_freeze || defined($main::success);
839 if ($main::please_info)
841 $main::please_info = 0;
843 create_output_collection();
845 update_progress_stats();
852 check_refresh_wanted();
854 update_progress_stats();
855 select (undef, undef, undef, 0.1);
857 elsif (time - $progress_stats_updated >= 30)
859 update_progress_stats();
861 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
862 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
864 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
865 .($thisround_failed+$thisround_succeeded)
866 .") -- giving up on this round";
867 Log (undef, $message);
871 # move slots from freeslot to holdslot (or back to freeslot) if necessary
872 for (my $i=$#freeslot; $i>=0; $i--) {
873 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
874 push @holdslot, (splice @freeslot, $i, 1);
877 for (my $i=$#holdslot; $i>=0; $i--) {
878 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
879 push @freeslot, (splice @holdslot, $i, 1);
883 # give up if no nodes are succeeding
884 if (!grep { $_->{node}->{losing_streak} == 0 &&
885 $_->{node}->{hold_count} < 4 } @slot) {
886 my $message = "Every node has failed -- giving up on this round";
887 Log (undef, $message);
894 push @freeslot, splice @holdslot;
895 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
898 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
901 if ($main::please_continue) {
902 $main::please_continue = 0;
905 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
909 check_refresh_wanted();
911 update_progress_stats();
912 select (undef, undef, undef, 0.1);
913 killem (keys %proc) if $main::please_freeze;
917 update_progress_stats();
918 freeze_if_want_freeze();
921 if (!defined $main::success)
924 $thisround_succeeded == 0 &&
925 ($thisround_failed == 0 || $thisround_failed > 4))
927 my $message = "stop because $thisround_failed tasks failed and none succeeded";
928 Log (undef, $message);
937 goto ONELEVEL if !defined $main::success;
940 release_allocation();
942 my $collated_output = &create_output_collection();
944 if (!$collated_output) {
945 Log (undef, "Failed to write output collection");
948 Log(undef, "job output $collated_output");
949 $Job->update_attributes('output' => $collated_output);
952 Log (undef, "finish");
957 if ($collated_output && $main::success) {
958 $final_state = 'Complete';
960 $final_state = 'Failed';
962 $Job->update_attributes('state' => $final_state);
964 exit (($final_state eq 'Complete') ? 0 : 1);
968 sub update_progress_stats
970 $progress_stats_updated = time;
971 return if !$progress_is_dirty;
972 my ($todo, $done, $running) = (scalar @jobstep_todo,
973 scalar @jobstep_done,
974 scalar @slot - scalar @freeslot - scalar @holdslot);
975 $Job->{'tasks_summary'} ||= {};
976 $Job->{'tasks_summary'}->{'todo'} = $todo;
977 $Job->{'tasks_summary'}->{'done'} = $done;
978 $Job->{'tasks_summary'}->{'running'} = $running;
979 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
980 Log (undef, "status: $done done, $running running, $todo todo");
981 $progress_is_dirty = 0;
988 my $pid = waitpid (-1, WNOHANG);
989 return 0 if $pid <= 0;
991 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
993 . $slot[$proc{$pid}->{slot}]->{cpu});
994 my $jobstepid = $proc{$pid}->{jobstep};
995 my $elapsed = time - $proc{$pid}->{time};
996 my $Jobstep = $jobstep[$jobstepid];
998 my $childstatus = $?;
999 my $exitvalue = $childstatus >> 8;
1000 my $exitinfo = "exit ".exit_status_s($childstatus);
1001 $Jobstep->{'arvados_task'}->reload;
1002 my $task_success = $Jobstep->{'arvados_task'}->{success};
1004 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1006 if (!defined $task_success) {
1007 # task did not indicate one way or the other --> fail
1008 $Jobstep->{'arvados_task'}->{success} = 0;
1009 $Jobstep->{'arvados_task'}->save;
1016 $temporary_fail ||= $Jobstep->{node_fail};
1017 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1019 ++$thisround_failed;
1020 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1022 # Check for signs of a failed or misconfigured node
1023 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1024 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1025 # Don't count this against jobstep failure thresholds if this
1026 # node is already suspected faulty and srun exited quickly
1027 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1029 Log ($jobstepid, "blaming failure on suspect node " .
1030 $slot[$proc{$pid}->{slot}]->{node}->{name});
1031 $temporary_fail ||= 1;
1033 ban_node_by_slot($proc{$pid}->{slot});
1036 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1037 ++$Jobstep->{'failures'},
1038 $temporary_fail ? 'temporary ' : 'permanent',
1041 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1042 # Give up on this task, and the whole job
1045 # Put this task back on the todo queue
1046 push @jobstep_todo, $jobstepid;
1047 $Job->{'tasks_summary'}->{'failed'}++;
1051 ++$thisround_succeeded;
1052 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1053 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1054 push @jobstep_done, $jobstepid;
1055 Log ($jobstepid, "success in $elapsed seconds");
1057 $Jobstep->{exitcode} = $childstatus;
1058 $Jobstep->{finishtime} = time;
1059 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1060 $Jobstep->{'arvados_task'}->save;
1061 process_stderr ($jobstepid, $task_success);
1062 Log ($jobstepid, sprintf("task output (%d bytes): %s",
1063 length($Jobstep->{'arvados_task'}->{output}),
1064 $Jobstep->{'arvados_task'}->{output}));
1066 close $reader{$jobstepid};
1067 delete $reader{$jobstepid};
1068 delete $slot[$proc{$pid}->{slot}]->{pid};
1069 push @freeslot, $proc{$pid}->{slot};
1072 if ($task_success) {
1074 my $newtask_list = [];
1075 my $newtask_results;
1077 $newtask_results = api_call(
1080 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1082 'order' => 'qsequence',
1083 'offset' => scalar(@$newtask_list),
1085 push(@$newtask_list, @{$newtask_results->{items}});
1086 } while (@{$newtask_results->{items}});
1087 foreach my $arvados_task (@$newtask_list) {
1089 'level' => $arvados_task->{'sequence'},
1091 'arvados_task' => $arvados_task
1093 push @jobstep, $jobstep;
1094 push @jobstep_todo, $#jobstep;
1098 $progress_is_dirty = 1;
1102 sub check_refresh_wanted
1104 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1105 if (@stat && $stat[9] > $latest_refresh) {
1106 $latest_refresh = scalar time;
1107 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1108 for my $attr ('cancelled_at',
1109 'cancelled_by_user_uuid',
1110 'cancelled_by_client_uuid',
1112 $Job->{$attr} = $Job2->{$attr};
1114 if ($Job->{'state'} ne "Running") {
1115 if ($Job->{'state'} eq "Cancelled") {
1116 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1118 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1121 $main::please_freeze = 1;
1128 # return if the kill list was checked <4 seconds ago
1129 if (defined $squeue_kill_checked && $squeue_kill_checked > time - 4)
1133 $squeue_kill_checked = time;
1135 # use killem() on procs whose killtime is reached
1138 if (exists $proc{$_}->{killtime}
1139 && $proc{$_}->{killtime} <= time)
1145 # return if the squeue was checked <60 seconds ago
1146 if (defined $squeue_checked && $squeue_checked > time - 60)
1150 $squeue_checked = time;
1154 # here is an opportunity to check for mysterious problems with local procs
1158 # get a list of steps still running
1159 my @squeue = `squeue -s -h -o '%i %j' && echo ok`;
1161 if ($squeue[-1] ne "ok")
1167 # which of my jobsteps are running, according to squeue?
1171 if (/^(\d+)\.(\d+) (\S+)/)
1173 if ($1 eq $ENV{SLURM_JOBID})
1180 # which of my active child procs (>60s old) were not mentioned by squeue?
1181 foreach (keys %proc)
1183 if ($proc{$_}->{time} < time - 60
1184 && !exists $ok{$proc{$_}->{jobstepname}}
1185 && !exists $proc{$_}->{killtime})
1187 # kill this proc if it hasn't exited in 30 seconds
1188 $proc{$_}->{killtime} = time + 30;
1194 sub release_allocation
1198 Log (undef, "release job allocation");
1199 system "scancel $ENV{SLURM_JOBID}";
1207 foreach my $job (keys %reader)
1210 while (0 < sysread ($reader{$job}, $buf, 8192))
1212 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1213 $jobstep[$job]->{stderr} .= $buf;
1214 preprocess_stderr ($job);
1215 if (length ($jobstep[$job]->{stderr}) > 16384)
1217 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1226 sub preprocess_stderr
1230 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1232 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1233 Log ($job, "stderr $line");
1234 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1236 $main::please_freeze = 1;
1238 elsif ($line =~ /srun: error: (Node failure on|Unable to create job step|.*: Communication connection failure)/) {
1239 $jobstep[$job]->{node_fail} = 1;
1240 ban_node_by_slot($jobstep[$job]->{slotindex});
1249 my $task_success = shift;
1250 preprocess_stderr ($job);
1253 Log ($job, "stderr $_");
1254 } split ("\n", $jobstep[$job]->{stderr});
1261 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1262 Log(undef, "fetch_block run error from arv-get $hash: $!");
1265 my $output_block = "";
1268 my $bytes = sysread($keep, $buf, 1024 * 1024);
1269 if (!defined $bytes) {
1270 Log(undef, "fetch_block read error from arv-get: $!");
1271 $output_block = undef;
1273 } elsif ($bytes == 0) {
1274 # sysread returns 0 at the end of the pipe.
1277 # some bytes were read into buf.
1278 $output_block .= $buf;
1283 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1284 $output_block = undef;
1286 return $output_block;
1289 # Create a collection by concatenating the output of all tasks (each
1290 # task's output is either a manifest fragment, a locator for a
1291 # manifest fragment stored in Keep, or nothing at all). Return the
1292 # portable_data_hash of the new collection.
1293 sub create_output_collection
1295 Log (undef, "collate");
1297 my ($child_out, $child_in);
1298 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1301 print (arvados.api("v1").collections().
1302 create(body={"manifest_text": sys.stdin.read()}).
1303 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1307 my $manifest_size = 0;
1311 my $output = $_->{'arvados_task'}->{output};
1312 next if (!defined($output));
1314 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1315 $next_write = fetch_block($output);
1317 $next_write = $output;
1319 if (defined($next_write)) {
1320 if (!defined(syswrite($child_in, $next_write))) {
1321 # There's been an error writing. Stop the loop.
1322 # We'll log details about the exit code later.
1325 $manifest_size += length($next_write);
1328 my $uuid = $_->{'arvados_task'}->{'uuid'};
1329 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1334 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1337 my $s = IO::Select->new($child_out);
1338 if ($s->can_read(120)) {
1339 sysread($child_out, $joboutput, 1024 * 1024);
1342 Log(undef, "output collection creation exited " . exit_status_s($?));
1348 Log (undef, "timed out while creating output collection");
1349 foreach my $signal (2, 2, 2, 15, 15, 9) {
1350 kill($signal, $pid);
1351 last if waitpid($pid, WNOHANG) == -1;
1365 my $sig = 2; # SIGINT first
1366 if (exists $proc{$_}->{"sent_$sig"} &&
1367 time - $proc{$_}->{"sent_$sig"} > 4)
1369 $sig = 15; # SIGTERM if SIGINT doesn't work
1371 if (exists $proc{$_}->{"sent_$sig"} &&
1372 time - $proc{$_}->{"sent_$sig"} > 4)
1374 $sig = 9; # SIGKILL if SIGTERM doesn't work
1376 if (!exists $proc{$_}->{"sent_$sig"})
1378 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1380 select (undef, undef, undef, 0.1);
1383 kill $sig, $_; # srun wants two SIGINT to really interrupt
1385 $proc{$_}->{"sent_$sig"} = time;
1386 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1396 vec($bits,fileno($_),1) = 1;
1402 # Send log output to Keep via arv-put.
1404 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1405 # $log_pipe_pid is the pid of the arv-put subprocess.
1407 # The only functions that should access these variables directly are:
1409 # log_writer_start($logfilename)
1410 # Starts an arv-put pipe, reading data on stdin and writing it to
1411 # a $logfilename file in an output collection.
1413 # log_writer_send($txt)
1414 # Writes $txt to the output log collection.
1416 # log_writer_finish()
1417 # Closes the arv-put pipe and returns the output that it produces.
1419 # log_writer_is_active()
1420 # Returns a true value if there is currently a live arv-put
1421 # process, false otherwise.
1423 my ($log_pipe_in, $log_pipe_out, $log_pipe_pid);
1425 sub log_writer_start($)
1427 my $logfilename = shift;
1428 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1430 '--portable-data-hash',
1431 '--project-uuid', $Job->{owner_uuid},
1433 '--name', $logfilename,
1434 '--filename', $logfilename,
1438 sub log_writer_send($)
1441 print $log_pipe_in $txt;
1444 sub log_writer_finish()
1446 return unless $log_pipe_pid;
1448 close($log_pipe_in);
1451 my $s = IO::Select->new($log_pipe_out);
1452 if ($s->can_read(120)) {
1453 sysread($log_pipe_out, $arv_put_output, 1024);
1454 chomp($arv_put_output);
1456 Log (undef, "timed out reading from 'arv-put'");
1459 waitpid($log_pipe_pid, 0);
1460 $log_pipe_pid = $log_pipe_in = $log_pipe_out = undef;
1462 Log("log_writer_finish: arv-put exited ".exit_status_s($?))
1465 return $arv_put_output;
1468 sub log_writer_is_active() {
1469 return $log_pipe_pid;
1472 sub Log # ($jobstep_id, $logmessage)
1474 if ($_[1] =~ /\n/) {
1475 for my $line (split (/\n/, $_[1])) {
1480 my $fh = select STDERR; $|=1; select $fh;
1481 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1482 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1485 if (log_writer_is_active() || -t STDERR) {
1486 my @gmtime = gmtime;
1487 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1488 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1490 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1492 if (log_writer_is_active()) {
1493 log_writer_send($datetime . " " . $message);
1500 my ($package, $file, $line) = caller;
1501 my $message = "@_ at $file line $line\n";
1502 Log (undef, $message);
1503 freeze() if @jobstep_todo;
1504 create_output_collection() if @jobstep_todo;
1514 if ($Job->{'state'} eq 'Cancelled') {
1515 $Job->update_attributes('finished_at' => scalar gmtime);
1517 $Job->update_attributes('state' => 'Failed');
1524 my $justcheckpoint = shift; # false if this will be the last meta saved
1525 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1526 return unless log_writer_is_active();
1528 my $loglocator = log_writer_finish();
1529 Log (undef, "log manifest is $loglocator");
1530 $Job->{'log'} = $loglocator;
1531 $Job->update_attributes('log', $loglocator);
1535 sub freeze_if_want_freeze
1537 if ($main::please_freeze)
1539 release_allocation();
1542 # kill some srun procs before freeze+stop
1543 map { $proc{$_} = {} } @_;
1546 killem (keys %proc);
1547 select (undef, undef, undef, 0.1);
1549 while (($died = waitpid (-1, WNOHANG)) > 0)
1551 delete $proc{$died};
1556 create_output_collection();
1566 Log (undef, "Freeze not implemented");
1573 croak ("Thaw not implemented");
1589 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1596 my $srunargs = shift;
1597 my $execargs = shift;
1598 my $opts = shift || {};
1600 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1602 $Data::Dumper::Terse = 1;
1603 $Data::Dumper::Indent = 0;
1604 my $show_cmd = Dumper($args);
1605 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1606 $show_cmd =~ s/\n/ /g;
1607 warn "starting: $show_cmd\n";
1609 if (defined $stdin) {
1610 my $child = open STDIN, "-|";
1611 defined $child or die "no fork: $!";
1613 print $stdin or die $!;
1614 close STDOUT or die $!;
1619 return system (@$args) if $opts->{fork};
1622 warn "ENV size is ".length(join(" ",%ENV));
1623 die "exec failed: $!: @$args";
1627 sub ban_node_by_slot {
1628 # Don't start any new jobsteps on this node for 60 seconds
1630 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1631 $slot[$slotid]->{node}->{hold_count}++;
1632 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1637 my ($lockfile, $error_message) = @_;
1638 open L, ">", $lockfile or croak("$lockfile: $!");
1639 if (!flock L, LOCK_EX|LOCK_NB) {
1640 croak("Can't lock $lockfile: $error_message\n");
1644 sub find_docker_image {
1645 # Given a Keep locator, check to see if it contains a Docker image.
1646 # If so, return its stream name and Docker hash.
1647 # If not, return undef for both values.
1648 my $locator = shift;
1649 my ($streamname, $filename);
1650 my $image = api_call("collections/get", uuid => $locator);
1652 foreach my $line (split(/\n/, $image->{manifest_text})) {
1653 my @tokens = split(/\s+/, $line);
1655 $streamname = shift(@tokens);
1656 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1657 if (defined($filename)) {
1658 return (undef, undef); # More than one file in the Collection.
1660 $filename = (split(/:/, $filedata, 3))[2];
1665 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1666 return ($streamname, $1);
1668 return (undef, undef);
1673 # Calculate the number of times an operation should be retried,
1674 # assuming exponential backoff, and that we're willing to retry as
1675 # long as tasks have been running. Enforce a minimum of 3 retries.
1676 my ($starttime, $endtime, $timediff, $retries);
1678 $starttime = $jobstep[0]->{starttime};
1679 $endtime = $jobstep[-1]->{finishtime};
1681 if (!defined($starttime)) {
1683 } elsif (!defined($endtime)) {
1684 $timediff = time - $starttime;
1686 $timediff = ($endtime - $starttime) - (time - $endtime);
1688 if ($timediff > 0) {
1689 $retries = int(log($timediff) / log(2));
1691 $retries = 1; # Use the minimum.
1693 return ($retries > 3) ? $retries : 3;
1697 # Pass in two function references.
1698 # This method will be called with the remaining arguments.
1699 # If it dies, retry it with exponential backoff until it succeeds,
1700 # or until the current retry_count is exhausted. After each failure
1701 # that can be retried, the second function will be called with
1702 # the current try count (0-based), next try time, and error message.
1703 my $operation = shift;
1704 my $retry_callback = shift;
1705 my $retries = retry_count();
1706 foreach my $try_count (0..$retries) {
1707 my $next_try = time + (2 ** $try_count);
1708 my $result = eval { $operation->(@_); };
1711 } elsif ($try_count < $retries) {
1712 $retry_callback->($try_count, $next_try, $@);
1713 my $sleep_time = $next_try - time;
1714 sleep($sleep_time) if ($sleep_time > 0);
1717 # Ensure the error message ends in a newline, so Perl doesn't add
1718 # retry_op's line number to it.
1724 # Pass in a /-separated API method name, and arguments for it.
1725 # This function will call that method, retrying as needed until
1726 # the current retry_count is exhausted, with a log on the first failure.
1727 my $method_name = shift;
1728 my $log_api_retry = sub {
1729 my ($try_count, $next_try_at, $errmsg) = @_;
1730 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
1731 $errmsg =~ s/\s/ /g;
1732 $errmsg =~ s/\s+$//;
1734 if ($next_try_at < time) {
1735 $retry_msg = "Retrying.";
1737 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
1738 $retry_msg = "Retrying at $next_try_fmt.";
1740 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
1743 foreach my $key (split(/\//, $method_name)) {
1744 $method = $method->{$key};
1746 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
1750 # Given a $?, return a human-readable exit code string like "0" or
1751 # "1" or "0 with signal 1" or "1 with signal 11".
1752 my $exitcode = shift;
1753 my $s = $exitcode >> 8;
1754 if ($exitcode & 0x7f) {
1755 $s .= " with signal " . ($exitcode & 0x7f);
1757 if ($exitcode & 0x80) {
1758 $s .= " with core dump";
1763 sub handle_readall {
1764 # Pass in a glob reference to a file handle.
1765 # Read all its contents and return them as a string.
1766 my $fh_glob_ref = shift;
1768 return <$fh_glob_ref>;
1771 sub tar_filename_n {
1773 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
1776 sub add_git_archive {
1777 # Pass in a git archive command as a string or list, a la system().
1778 # This method will save its output to be included in the archive sent to the
1782 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
1783 croak("Failed to save git archive: $!");
1785 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
1787 waitpid($git_pid, 0);
1790 croak("Failed to save git archive: git exited " . exit_status_s($?));
1794 sub combined_git_archive {
1795 # Combine all saved tar archives into a single archive, then return its
1796 # contents in a string. Return undef if no archives have been saved.
1797 if ($git_tar_count < 1) {
1800 my $base_tar_name = tar_filename_n(1);
1801 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
1802 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
1803 if ($tar_exit != 0) {
1804 croak("Error preparing build archive: tar -A exited " .
1805 exit_status_s($tar_exit));
1808 if (!open(GIT_TAR, "<", $base_tar_name)) {
1809 croak("Could not open build archive: $!");
1811 my $tar_contents = handle_readall(\*GIT_TAR);
1813 return $tar_contents;
1819 # This is crunch-job's internal dispatch script. crunch-job running on the API
1820 # server invokes this script on individual compute nodes, or localhost if we're
1821 # running a job locally. It gets called in two modes:
1823 # * No arguments: Installation mode. Read a tar archive from the DATA
1824 # file handle; it includes the Crunch script's source code, and
1825 # maybe SDKs as well. Those should be installed in the proper
1826 # locations. This runs outside of any Docker container, so don't try to
1827 # introspect Crunch's runtime environment.
1829 # * With arguments: Crunch script run mode. This script should set up the
1830 # environment, then run the command specified in the arguments. This runs
1831 # inside any Docker container.
1834 use File::Path qw( make_path remove_tree );
1835 use POSIX qw(getcwd);
1837 use constant TASK_TEMPFAIL => 111;
1839 # Map SDK subdirectories to the path environments they belong to.
1840 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
1842 my $destdir = $ENV{"CRUNCH_SRC"};
1843 my $commit = $ENV{"CRUNCH_SRC_COMMIT"};
1844 my $repo = $ENV{"CRUNCH_SRC_URL"};
1845 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
1846 my $job_work = $ENV{"JOB_WORK"};
1847 my $task_work = $ENV{"TASK_WORK"};
1849 for my $dir ($destdir, $job_work, $task_work) {
1852 -e $dir or die "Failed to create temporary directory ($dir): $!";
1857 remove_tree($task_work, {keep_root => 1});
1860 open(STDOUT_ORIG, ">&", STDOUT);
1861 open(STDERR_ORIG, ">&", STDERR);
1862 open(STDOUT, ">>", "$destdir.log");
1863 open(STDERR, ">&", STDOUT);
1865 ### Crunch script run mode
1867 # We want to do routine logging during task 0 only. This gives the user
1868 # the information they need, but avoids repeating the information for every
1871 if ($ENV{TASK_SEQUENCE} eq "0") {
1874 printf STDERR_ORIG "[Crunch] $msg\n", @_;
1880 my $python_src = "$install_dir/python";
1881 my $venv_dir = "$job_work/.arvados.venv";
1882 my $venv_built = -e "$venv_dir/bin/activate";
1883 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
1884 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
1885 "--python=python2.7", $venv_dir);
1886 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
1888 $Log->("Built Python SDK virtualenv");
1891 my $pip_bin = "pip";
1893 $Log->("Running in Python SDK virtualenv");
1894 $pip_bin = "$venv_dir/bin/pip";
1895 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
1896 @ARGV = ("/bin/sh", "-ec",
1897 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
1898 } elsif (-d $python_src) {
1899 $Log->("Warning: virtualenv not found inside Docker container default " .
1900 "\$PATH. Can't install Python SDK.");
1903 my $pkgs = `(\Q$pip_bin\E freeze 2>/dev/null | grep arvados) || dpkg-query --show '*arvados*'`;
1905 $Log->("Using Arvados SDK:");
1906 foreach my $line (split /\n/, $pkgs) {
1910 $Log->("Arvados SDK packages not found");
1913 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
1914 my $sdk_path = "$install_dir/$sdk_dir";
1916 if ($ENV{$sdk_envkey}) {
1917 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
1919 $ENV{$sdk_envkey} = $sdk_path;
1921 $Log->("Arvados SDK added to %s", $sdk_envkey);
1927 open(STDOUT, ">&", STDOUT_ORIG);
1928 open(STDERR, ">&", STDERR_ORIG);
1930 die "Cannot exec `@ARGV`: $!";
1933 ### Installation mode
1934 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
1936 if (readlink ("$destdir.commit") eq $commit && -d $destdir) {
1937 # This version already installed -> nothing to do.
1941 unlink "$destdir.commit";
1944 if (!open(TARX, "|-", "tar", "-xC", $destdir)) {
1945 die "Error launching 'tar -xC $destdir': $!";
1947 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
1948 # get SIGPIPE. We must feed it data incrementally.
1950 while (read(DATA, $tar_input, 65536)) {
1951 print TARX $tar_input;
1954 die "'tar -xC $destdir' exited $?: $!";
1959 my $sdk_root = "$destdir/.arvados.sdk/sdk";
1961 foreach my $sdk_lang (("python",
1962 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
1963 if (-d "$sdk_root/$sdk_lang") {
1964 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
1965 die "Failed to install $sdk_lang SDK: $!";
1971 my $python_dir = "$install_dir/python";
1972 if ((-d $python_dir) and can_run("python2.7") and
1973 (system("python2.7", "$python_dir/setup.py", "--quiet", "egg_info") != 0)) {
1974 # egg_info failed, probably when it asked git for a build tag.
1975 # Specify no build tag.
1976 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
1977 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
1981 if (-e "$destdir/crunch_scripts/install") {
1982 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
1983 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
1985 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
1986 } elsif (-e "./install.sh") {
1987 shell_or_die (undef, "./install.sh", $install_dir);
1991 unlink "$destdir.commit.new";
1992 symlink ($commit, "$destdir.commit.new") or die "$destdir.commit.new: $!";
1993 rename ("$destdir.commit.new", "$destdir.commit") or die "$destdir.commit: $!";
1999 my $command_name = shift;
2000 open(my $which, "-|", "which", $command_name);
2001 while (<$which>) { }
2008 my $exitcode = shift;
2010 if ($ENV{"DEBUG"}) {
2011 print STDERR "@_\n";
2013 if (system (@_) != 0) {
2016 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2017 open STDERR, ">&STDERR_ORIG";
2018 system ("cat $destdir.log >&2");
2019 warn "@_ failed ($err): $exitstatus";
2020 if (defined($exitcode)) {
2024 exit (($code >> 8) || 1);