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, $docker_limitmem);
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 # Determine whether this version of Docker supports memory+swap limits.
412 srun(["srun", "--nodelist=" . $node[0]],
413 ["/bin/sh", "-ec", "$docker_bin run --help | grep -qe --memory-swap="],
415 $docker_limitmem = ($? == 0);
417 if ($Job->{arvados_sdk_version}) {
418 # The job also specifies an Arvados SDK version. Add the SDKs to the
419 # tar file for the build script to install.
420 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
421 $Job->{arvados_sdk_version}));
422 add_git_archive("git", "--git-dir=$git_dir", "archive",
423 "--prefix=.arvados.sdk/",
424 $Job->{arvados_sdk_version}, "sdk");
428 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
429 # If script_version looks like an absolute path, *and* the --git-dir
430 # argument was not given -- which implies we were not invoked by
431 # crunch-dispatch -- we will use the given path as a working
432 # directory instead of resolving script_version to a git commit (or
433 # doing anything else with git).
434 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
435 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
438 # Resolve the given script_version to a git commit sha1. Also, if
439 # the repository is remote, clone it into our local filesystem: this
440 # ensures "git archive" will work, and is necessary to reliably
441 # resolve a symbolic script_version like "master^".
442 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
444 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
446 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
448 # If we're running under crunch-dispatch, it will have already
449 # pulled the appropriate source tree into its own repository, and
450 # given us that repo's path as $git_dir.
452 # If we're running a "local" job, we might have to fetch content
453 # from a remote repository.
455 # (Currently crunch-dispatch gives a local path with --git-dir, but
456 # we might as well accept URLs there too in case it changes its
458 my $repo = $git_dir || $Job->{'repository'};
460 # Repository can be remote or local. If remote, we'll need to fetch it
461 # to a local dir before doing `git log` et al.
464 if ($repo =~ m{://|^[^/]*:}) {
465 # $repo is a git url we can clone, like git:// or https:// or
466 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
467 # not recognized here because distinguishing that from a local
468 # path is too fragile. If you really need something strange here,
469 # use the ssh:// form.
470 $repo_location = 'remote';
471 } elsif ($repo =~ m{^\.*/}) {
472 # $repo is a local path to a git index. We'll also resolve ../foo
473 # to ../foo/.git if the latter is a directory. To help
474 # disambiguate local paths from named hosted repositories, this
475 # form must be given as ./ or ../ if it's a relative path.
476 if (-d "$repo/.git") {
477 $repo = "$repo/.git";
479 $repo_location = 'local';
481 # $repo is none of the above. It must be the name of a hosted
483 my $arv_repo_list = api_call("repositories/list",
484 'filters' => [['name','=',$repo]]);
485 my @repos_found = @{$arv_repo_list->{'items'}};
486 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
488 Log(undef, "Repository '$repo' -> "
489 . join(", ", map { $_->{'uuid'} } @repos_found));
492 croak("Error: Found $n_found repositories with name '$repo'.");
494 $repo = $repos_found[0]->{'fetch_url'};
495 $repo_location = 'remote';
497 Log(undef, "Using $repo_location repository '$repo'");
498 $ENV{"CRUNCH_SRC_URL"} = $repo;
500 # Resolve given script_version (we'll call that $treeish here) to a
501 # commit sha1 ($commit).
502 my $treeish = $Job->{'script_version'};
504 if ($repo_location eq 'remote') {
505 # We minimize excess object-fetching by re-using the same bare
506 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
507 # just keep adding remotes to it as needed.
508 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
509 my $gitcmd = "git --git-dir=\Q$local_repo\E";
511 # Set up our local repo for caching remote objects, making
513 if (!-d $local_repo) {
514 make_path($local_repo) or croak("Error: could not create $local_repo");
516 # This works (exits 0 and doesn't delete fetched objects) even
517 # if $local_repo is already initialized:
518 `$gitcmd init --bare`;
520 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
523 # If $treeish looks like a hash (or abbrev hash) we look it up in
524 # our local cache first, since that's cheaper. (We don't want to
525 # do that with tags/branches though -- those change over time, so
526 # they should always be resolved by the remote repo.)
527 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
528 # Hide stderr because it's normal for this to fail:
529 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
531 # Careful not to resolve a branch named abcdeff to commit 1234567:
532 $sha1 =~ /^$treeish/ &&
533 $sha1 =~ /^([0-9a-f]{40})$/s) {
535 Log(undef, "Commit $commit already present in $local_repo");
539 if (!defined $commit) {
540 # If $treeish isn't just a hash or abbrev hash, or isn't here
541 # yet, we need to fetch the remote to resolve it correctly.
543 # First, remove all local heads. This prevents a name that does
544 # not exist on the remote from resolving to (or colliding with)
545 # a previously fetched branch or tag (possibly from a different
547 remove_tree("$local_repo/refs/heads", {keep_root => 1});
549 Log(undef, "Fetching objects from $repo to $local_repo");
550 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
552 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
556 # Now that the data is all here, we will use our local repo for
557 # the rest of our git activities.
561 my $gitcmd = "git --git-dir=\Q$repo\E";
562 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
563 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
564 croak("`$gitcmd rev-list` exited "
566 .", '$treeish' not found. Giving up.");
569 Log(undef, "Version $treeish is commit $commit");
571 if ($commit ne $Job->{'script_version'}) {
572 # Record the real commit id in the database, frozentokey, logs,
573 # etc. -- instead of an abbreviation or a branch name which can
574 # become ambiguous or point to a different commit in the future.
575 if (!$Job->update_attributes('script_version' => $commit)) {
576 croak("Error: failed to update job's script_version attribute");
580 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
581 add_git_archive("$gitcmd archive ''\Q$commit\E");
584 my $git_archive = combined_git_archive();
585 if (!defined $git_archive) {
586 Log(undef, "Skip install phase (no git archive)");
588 Log(undef, "Warning: This probably means workers have no source tree!");
592 Log(undef, "Run install script on all workers");
594 my @srunargs = ("srun",
595 "--nodelist=$nodelist",
596 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
597 my @execargs = ("sh", "-c",
598 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
600 my $installpid = fork();
601 if ($installpid == 0)
603 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
608 last if $installpid == waitpid (-1, WNOHANG);
609 freeze_if_want_freeze ($installpid);
610 select (undef, undef, undef, 0.1);
612 my $install_exited = $?;
613 Log (undef, "Install script exited ".exit_status_s($install_exited));
614 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
615 unlink($tar_filename);
617 exit (1) if $install_exited != 0;
620 foreach (qw (script script_version script_parameters runtime_constraints))
624 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
626 foreach (split (/\n/, $Job->{knobs}))
628 Log (undef, "knob " . $_);
633 $main::success = undef;
639 my $thisround_succeeded = 0;
640 my $thisround_failed = 0;
641 my $thisround_failed_multiple = 0;
643 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
644 or $a <=> $b } @jobstep_todo;
645 my $level = $jobstep[$jobstep_todo[0]]->{level};
646 Log (undef, "start level $level");
651 my @freeslot = (0..$#slot);
654 my $progress_is_dirty = 1;
655 my $progress_stats_updated = 0;
657 update_progress_stats();
660 my $tasks_this_level = 0;
661 foreach my $id (@jobstep_todo) {
662 $tasks_this_level++ if ($jobstep[$id]->{level} == $level);
667 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
669 my $id = $jobstep_todo[$todo_ptr];
670 my $Jobstep = $jobstep[$id];
671 if ($Jobstep->{level} != $level)
676 pipe $reader{$id}, "writer" or croak ($!);
677 my $flags = fcntl ($reader{$id}, F_GETFL, 0) or croak ($!);
678 fcntl ($reader{$id}, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
680 my $childslot = $freeslot[0];
681 my $childnode = $slot[$childslot]->{node};
682 my $childslotname = join (".",
683 $slot[$childslot]->{node}->{name},
684 $slot[$childslot]->{cpu});
686 my $childpid = fork();
689 $SIG{'INT'} = 'DEFAULT';
690 $SIG{'QUIT'} = 'DEFAULT';
691 $SIG{'TERM'} = 'DEFAULT';
693 foreach (values (%reader))
697 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
698 open(STDOUT,">&writer");
699 open(STDERR,">&writer");
704 delete $ENV{"GNUPGHOME"};
705 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
706 $ENV{"TASK_QSEQUENCE"} = $id;
707 $ENV{"TASK_SEQUENCE"} = $level;
708 $ENV{"JOB_SCRIPT"} = $Job->{script};
709 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
710 $param =~ tr/a-z/A-Z/;
711 $ENV{"JOB_PARAMETER_$param"} = $value;
713 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
714 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
715 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
716 $ENV{"HOME"} = $ENV{"TASK_WORK"};
717 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
718 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
719 $ENV{"CRUNCH_NODE_SLOTS"} = $slot[$childslot]->{node}->{ncpus};
720 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
724 my $max_node_concurrent_tasks = $ENV{CRUNCH_NODE_SLOTS};
725 if ($tasks_this_level < $max_node_concurrent_tasks) {
726 $max_node_concurrent_tasks = $tasks_this_level;
731 "--nodelist=".$childnode->{name},
732 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
733 "--job-name=$job_id.$id.$$",
736 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
737 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
738 ."&& cd $ENV{CRUNCH_TMP} "
739 # These environment variables get used explicitly later in
740 # $command. No tool is expected to read these values directly.
741 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
742 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
743 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($max_node_concurrent_tasks * 100) )) "
744 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
745 $command .= "&& exec arv-mount --by-id --allow-other $ENV{TASK_KEEPMOUNT} --exec ";
748 my $cidfile = "$ENV{CRUNCH_TMP}/$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}.cid";
749 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
750 $command .= "$docker_bin run --rm=true --attach=stdout --attach=stderr --attach=stdin -i --user=crunch --cidfile=$cidfile --sig-proxy ";
751 # We only set memory limits if Docker lets us limit both memory and swap.
752 # Memory limits alone have been supported longer, but subprocesses tend
753 # to get SIGKILL if they exceed that without any swap limit set.
754 # See #5642 for additional background.
755 if ($docker_limitmem) {
756 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
759 # Dynamically configure the container to use the host system as its
760 # DNS server. Get the host's global addresses from the ip command,
761 # and turn them into docker --dns options using gawk.
763 q{$(ip -o address show scope global |
764 gawk 'match($4, /^([0-9\.:]+)\//, x){print "--dns", x[1]}') };
766 # The source tree and $destdir directory (which we have
767 # installed on the worker host) are available in the container,
768 # under the same path.
769 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
770 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
772 # Currently, we make arv-mount's mount point appear at /keep
773 # inside the container (instead of using the same path as the
774 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
775 # crunch scripts and utilities must not rely on this. They must
776 # use $TASK_KEEPMOUNT.
777 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
778 $ENV{TASK_KEEPMOUNT} = "/keep";
780 # TASK_WORK is almost exactly like a docker data volume: it
781 # starts out empty, is writable, and persists until no
782 # containers use it any more. We don't use --volumes-from to
783 # share it with other containers: it is only accessible to this
784 # task, and it goes away when this task stops.
786 # However, a docker data volume is writable only by root unless
787 # the mount point already happens to exist in the container with
788 # different permissions. Therefore, we [1] assume /tmp already
789 # exists in the image and is writable by the crunch user; [2]
790 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
791 # writable if they are created by docker while setting up the
792 # other --volumes); and [3] create $TASK_WORK inside the
793 # container using $build_script.
794 $command .= "--volume=/tmp ";
795 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
796 $ENV{"HOME"} = $ENV{"TASK_WORK"};
797 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
799 # TODO: Share a single JOB_WORK volume across all task
800 # containers on a given worker node, and delete it when the job
801 # ends (and, in case that doesn't work, when the next job
804 # For now, use the same approach as TASK_WORK above.
805 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
807 while (my ($env_key, $env_val) = each %ENV)
809 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
810 $command .= "--env=\Q$env_key=$env_val\E ";
813 $command .= "--env=\QHOME=$ENV{HOME}\E ";
814 $command .= "\Q$docker_hash\E ";
815 $command .= "stdbuf --output=0 --error=0 ";
816 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
819 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
820 $command .= "stdbuf --output=0 --error=0 ";
821 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
824 my @execargs = ('bash', '-c', $command);
825 srun (\@srunargs, \@execargs, undef, $build_script);
826 # exec() failed, we assume nothing happened.
827 die "srun() failed on build script\n";
830 if (!defined $childpid)
837 $proc{$childpid} = { jobstep => $id,
840 jobstepname => "$job_id.$id.$childpid",
842 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
843 $slot[$childslot]->{pid} = $childpid;
845 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
846 Log ($id, "child $childpid started on $childslotname");
847 $Jobstep->{starttime} = time;
848 $Jobstep->{node} = $childnode->{name};
849 $Jobstep->{slotindex} = $childslot;
850 delete $Jobstep->{stderr};
851 delete $Jobstep->{finishtime};
853 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
854 $Jobstep->{'arvados_task'}->save;
856 splice @jobstep_todo, $todo_ptr, 1;
859 $progress_is_dirty = 1;
863 (@slot > @freeslot && $todo_ptr+1 > $#jobstep_todo))
865 last THISROUND if $main::please_freeze || defined($main::success);
866 if ($main::please_info)
868 $main::please_info = 0;
870 create_output_collection();
872 update_progress_stats();
879 check_refresh_wanted();
881 update_progress_stats();
882 select (undef, undef, undef, 0.1);
884 elsif (time - $progress_stats_updated >= 30)
886 update_progress_stats();
888 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
889 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
891 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
892 .($thisround_failed+$thisround_succeeded)
893 .") -- giving up on this round";
894 Log (undef, $message);
898 # move slots from freeslot to holdslot (or back to freeslot) if necessary
899 for (my $i=$#freeslot; $i>=0; $i--) {
900 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
901 push @holdslot, (splice @freeslot, $i, 1);
904 for (my $i=$#holdslot; $i>=0; $i--) {
905 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
906 push @freeslot, (splice @holdslot, $i, 1);
910 # give up if no nodes are succeeding
911 if (!grep { $_->{node}->{losing_streak} == 0 &&
912 $_->{node}->{hold_count} < 4 } @slot) {
913 my $message = "Every node has failed -- giving up on this round";
914 Log (undef, $message);
921 push @freeslot, splice @holdslot;
922 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
925 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
928 if ($main::please_continue) {
929 $main::please_continue = 0;
932 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
936 check_refresh_wanted();
938 update_progress_stats();
939 select (undef, undef, undef, 0.1);
940 killem (keys %proc) if $main::please_freeze;
944 update_progress_stats();
945 freeze_if_want_freeze();
948 if (!defined $main::success)
951 $thisround_succeeded == 0 &&
952 ($thisround_failed == 0 || $thisround_failed > 4))
954 my $message = "stop because $thisround_failed tasks failed and none succeeded";
955 Log (undef, $message);
964 goto ONELEVEL if !defined $main::success;
967 release_allocation();
969 my $collated_output = &create_output_collection();
971 if (!$collated_output) {
972 Log (undef, "Failed to write output collection");
975 Log(undef, "job output $collated_output");
976 $Job->update_attributes('output' => $collated_output);
979 Log (undef, "finish");
984 if ($collated_output && $main::success) {
985 $final_state = 'Complete';
987 $final_state = 'Failed';
989 $Job->update_attributes('state' => $final_state);
991 exit (($final_state eq 'Complete') ? 0 : 1);
995 sub update_progress_stats
997 $progress_stats_updated = time;
998 return if !$progress_is_dirty;
999 my ($todo, $done, $running) = (scalar @jobstep_todo,
1000 scalar @jobstep_done,
1001 scalar @slot - scalar @freeslot - scalar @holdslot);
1002 $Job->{'tasks_summary'} ||= {};
1003 $Job->{'tasks_summary'}->{'todo'} = $todo;
1004 $Job->{'tasks_summary'}->{'done'} = $done;
1005 $Job->{'tasks_summary'}->{'running'} = $running;
1006 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1007 Log (undef, "status: $done done, $running running, $todo todo");
1008 $progress_is_dirty = 0;
1015 my $pid = waitpid (-1, WNOHANG);
1016 return 0 if $pid <= 0;
1018 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1020 . $slot[$proc{$pid}->{slot}]->{cpu});
1021 my $jobstepid = $proc{$pid}->{jobstep};
1022 my $elapsed = time - $proc{$pid}->{time};
1023 my $Jobstep = $jobstep[$jobstepid];
1025 my $childstatus = $?;
1026 my $exitvalue = $childstatus >> 8;
1027 my $exitinfo = "exit ".exit_status_s($childstatus);
1028 $Jobstep->{'arvados_task'}->reload;
1029 my $task_success = $Jobstep->{'arvados_task'}->{success};
1031 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1033 if (!defined $task_success) {
1034 # task did not indicate one way or the other --> fail
1035 $Jobstep->{'arvados_task'}->{success} = 0;
1036 $Jobstep->{'arvados_task'}->save;
1043 $temporary_fail ||= $Jobstep->{node_fail};
1044 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1046 ++$thisround_failed;
1047 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1049 # Check for signs of a failed or misconfigured node
1050 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1051 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1052 # Don't count this against jobstep failure thresholds if this
1053 # node is already suspected faulty and srun exited quickly
1054 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1056 Log ($jobstepid, "blaming failure on suspect node " .
1057 $slot[$proc{$pid}->{slot}]->{node}->{name});
1058 $temporary_fail ||= 1;
1060 ban_node_by_slot($proc{$pid}->{slot});
1063 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1064 ++$Jobstep->{'failures'},
1065 $temporary_fail ? 'temporary ' : 'permanent',
1068 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1069 # Give up on this task, and the whole job
1072 # Put this task back on the todo queue
1073 push @jobstep_todo, $jobstepid;
1074 $Job->{'tasks_summary'}->{'failed'}++;
1078 ++$thisround_succeeded;
1079 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1080 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1081 push @jobstep_done, $jobstepid;
1082 Log ($jobstepid, "success in $elapsed seconds");
1084 $Jobstep->{exitcode} = $childstatus;
1085 $Jobstep->{finishtime} = time;
1086 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1087 $Jobstep->{'arvados_task'}->save;
1088 process_stderr ($jobstepid, $task_success);
1089 Log ($jobstepid, sprintf("task output (%d bytes): %s",
1090 length($Jobstep->{'arvados_task'}->{output}),
1091 $Jobstep->{'arvados_task'}->{output}));
1093 close $reader{$jobstepid};
1094 delete $reader{$jobstepid};
1095 delete $slot[$proc{$pid}->{slot}]->{pid};
1096 push @freeslot, $proc{$pid}->{slot};
1099 if ($task_success) {
1101 my $newtask_list = [];
1102 my $newtask_results;
1104 $newtask_results = api_call(
1107 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1109 'order' => 'qsequence',
1110 'offset' => scalar(@$newtask_list),
1112 push(@$newtask_list, @{$newtask_results->{items}});
1113 } while (@{$newtask_results->{items}});
1114 foreach my $arvados_task (@$newtask_list) {
1116 'level' => $arvados_task->{'sequence'},
1118 'arvados_task' => $arvados_task
1120 push @jobstep, $jobstep;
1121 push @jobstep_todo, $#jobstep;
1125 $progress_is_dirty = 1;
1129 sub check_refresh_wanted
1131 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1132 if (@stat && $stat[9] > $latest_refresh) {
1133 $latest_refresh = scalar time;
1134 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1135 for my $attr ('cancelled_at',
1136 'cancelled_by_user_uuid',
1137 'cancelled_by_client_uuid',
1139 $Job->{$attr} = $Job2->{$attr};
1141 if ($Job->{'state'} ne "Running") {
1142 if ($Job->{'state'} eq "Cancelled") {
1143 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1145 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1148 $main::please_freeze = 1;
1155 # return if the kill list was checked <4 seconds ago
1156 if (defined $squeue_kill_checked && $squeue_kill_checked > time - 4)
1160 $squeue_kill_checked = time;
1162 # use killem() on procs whose killtime is reached
1165 if (exists $proc{$_}->{killtime}
1166 && $proc{$_}->{killtime} <= time)
1172 # return if the squeue was checked <60 seconds ago
1173 if (defined $squeue_checked && $squeue_checked > time - 60)
1177 $squeue_checked = time;
1181 # here is an opportunity to check for mysterious problems with local procs
1185 # get a list of steps still running
1186 my @squeue = `squeue -s -h -o '%i %j' && echo ok`;
1188 if ($squeue[-1] ne "ok")
1194 # which of my jobsteps are running, according to squeue?
1198 if (/^(\d+)\.(\d+) (\S+)/)
1200 if ($1 eq $ENV{SLURM_JOBID})
1207 # which of my active child procs (>60s old) were not mentioned by squeue?
1208 foreach (keys %proc)
1210 if ($proc{$_}->{time} < time - 60
1211 && !exists $ok{$proc{$_}->{jobstepname}}
1212 && !exists $proc{$_}->{killtime})
1214 # kill this proc if it hasn't exited in 30 seconds
1215 $proc{$_}->{killtime} = time + 30;
1221 sub release_allocation
1225 Log (undef, "release job allocation");
1226 system "scancel $ENV{SLURM_JOBID}";
1234 foreach my $job (keys %reader)
1237 while (0 < sysread ($reader{$job}, $buf, 8192))
1239 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1240 $jobstep[$job]->{stderr} .= $buf;
1241 preprocess_stderr ($job);
1242 if (length ($jobstep[$job]->{stderr}) > 16384)
1244 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1253 sub preprocess_stderr
1257 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1259 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1260 Log ($job, "stderr $line");
1261 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1263 $main::please_freeze = 1;
1265 elsif ($line =~ /(srun: error: (Node failure on|Unable to create job step|.*: Communication connection failure))|arvados.errors.Keep/) {
1266 $jobstep[$job]->{node_fail} = 1;
1267 ban_node_by_slot($jobstep[$job]->{slotindex});
1276 my $task_success = shift;
1277 preprocess_stderr ($job);
1280 Log ($job, "stderr $_");
1281 } split ("\n", $jobstep[$job]->{stderr});
1288 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1289 Log(undef, "fetch_block run error from arv-get $hash: $!");
1292 my $output_block = "";
1295 my $bytes = sysread($keep, $buf, 1024 * 1024);
1296 if (!defined $bytes) {
1297 Log(undef, "fetch_block read error from arv-get: $!");
1298 $output_block = undef;
1300 } elsif ($bytes == 0) {
1301 # sysread returns 0 at the end of the pipe.
1304 # some bytes were read into buf.
1305 $output_block .= $buf;
1310 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1311 $output_block = undef;
1313 return $output_block;
1316 # Create a collection by concatenating the output of all tasks (each
1317 # task's output is either a manifest fragment, a locator for a
1318 # manifest fragment stored in Keep, or nothing at all). Return the
1319 # portable_data_hash of the new collection.
1320 sub create_output_collection
1322 Log (undef, "collate");
1324 my ($child_out, $child_in);
1325 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1328 print (arvados.api("v1").collections().
1329 create(body={"manifest_text": sys.stdin.read()}).
1330 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1334 my $manifest_size = 0;
1338 my $output = $_->{'arvados_task'}->{output};
1339 next if (!defined($output));
1341 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1342 $next_write = fetch_block($output);
1344 $next_write = $output;
1346 if (defined($next_write)) {
1347 if (!defined(syswrite($child_in, $next_write))) {
1348 # There's been an error writing. Stop the loop.
1349 # We'll log details about the exit code later.
1352 $manifest_size += length($next_write);
1355 my $uuid = $_->{'arvados_task'}->{'uuid'};
1356 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1361 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1364 my $s = IO::Select->new($child_out);
1365 if ($s->can_read(120)) {
1366 sysread($child_out, $joboutput, 1024 * 1024);
1369 Log(undef, "output collection creation exited " . exit_status_s($?));
1375 Log (undef, "timed out while creating output collection");
1376 foreach my $signal (2, 2, 2, 15, 15, 9) {
1377 kill($signal, $pid);
1378 last if waitpid($pid, WNOHANG) == -1;
1392 my $sig = 2; # SIGINT first
1393 if (exists $proc{$_}->{"sent_$sig"} &&
1394 time - $proc{$_}->{"sent_$sig"} > 4)
1396 $sig = 15; # SIGTERM if SIGINT doesn't work
1398 if (exists $proc{$_}->{"sent_$sig"} &&
1399 time - $proc{$_}->{"sent_$sig"} > 4)
1401 $sig = 9; # SIGKILL if SIGTERM doesn't work
1403 if (!exists $proc{$_}->{"sent_$sig"})
1405 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1407 select (undef, undef, undef, 0.1);
1410 kill $sig, $_; # srun wants two SIGINT to really interrupt
1412 $proc{$_}->{"sent_$sig"} = time;
1413 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1423 vec($bits,fileno($_),1) = 1;
1429 # Send log output to Keep via arv-put.
1431 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1432 # $log_pipe_pid is the pid of the arv-put subprocess.
1434 # The only functions that should access these variables directly are:
1436 # log_writer_start($logfilename)
1437 # Starts an arv-put pipe, reading data on stdin and writing it to
1438 # a $logfilename file in an output collection.
1440 # log_writer_send($txt)
1441 # Writes $txt to the output log collection.
1443 # log_writer_finish()
1444 # Closes the arv-put pipe and returns the output that it produces.
1446 # log_writer_is_active()
1447 # Returns a true value if there is currently a live arv-put
1448 # process, false otherwise.
1450 my ($log_pipe_in, $log_pipe_out, $log_pipe_pid);
1452 sub log_writer_start($)
1454 my $logfilename = shift;
1455 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1457 '--portable-data-hash',
1458 '--project-uuid', $Job->{owner_uuid},
1460 '--name', $logfilename,
1461 '--filename', $logfilename,
1465 sub log_writer_send($)
1468 print $log_pipe_in $txt;
1471 sub log_writer_finish()
1473 return unless $log_pipe_pid;
1475 close($log_pipe_in);
1478 my $s = IO::Select->new($log_pipe_out);
1479 if ($s->can_read(120)) {
1480 sysread($log_pipe_out, $arv_put_output, 1024);
1481 chomp($arv_put_output);
1483 Log (undef, "timed out reading from 'arv-put'");
1486 waitpid($log_pipe_pid, 0);
1487 $log_pipe_pid = $log_pipe_in = $log_pipe_out = undef;
1489 Log("log_writer_finish: arv-put exited ".exit_status_s($?))
1492 return $arv_put_output;
1495 sub log_writer_is_active() {
1496 return $log_pipe_pid;
1499 sub Log # ($jobstep_id, $logmessage)
1501 if ($_[1] =~ /\n/) {
1502 for my $line (split (/\n/, $_[1])) {
1507 my $fh = select STDERR; $|=1; select $fh;
1508 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1509 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1512 if (log_writer_is_active() || -t STDERR) {
1513 my @gmtime = gmtime;
1514 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1515 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1517 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1519 if (log_writer_is_active()) {
1520 log_writer_send($datetime . " " . $message);
1527 my ($package, $file, $line) = caller;
1528 my $message = "@_ at $file line $line\n";
1529 Log (undef, $message);
1530 freeze() if @jobstep_todo;
1531 create_output_collection() if @jobstep_todo;
1541 if ($Job->{'state'} eq 'Cancelled') {
1542 $Job->update_attributes('finished_at' => scalar gmtime);
1544 $Job->update_attributes('state' => 'Failed');
1551 my $justcheckpoint = shift; # false if this will be the last meta saved
1552 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1553 return unless log_writer_is_active();
1555 my $loglocator = log_writer_finish();
1556 Log (undef, "log manifest is $loglocator");
1557 $Job->{'log'} = $loglocator;
1558 $Job->update_attributes('log', $loglocator);
1562 sub freeze_if_want_freeze
1564 if ($main::please_freeze)
1566 release_allocation();
1569 # kill some srun procs before freeze+stop
1570 map { $proc{$_} = {} } @_;
1573 killem (keys %proc);
1574 select (undef, undef, undef, 0.1);
1576 while (($died = waitpid (-1, WNOHANG)) > 0)
1578 delete $proc{$died};
1583 create_output_collection();
1593 Log (undef, "Freeze not implemented");
1600 croak ("Thaw not implemented");
1616 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1623 my $srunargs = shift;
1624 my $execargs = shift;
1625 my $opts = shift || {};
1627 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1629 $Data::Dumper::Terse = 1;
1630 $Data::Dumper::Indent = 0;
1631 my $show_cmd = Dumper($args);
1632 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1633 $show_cmd =~ s/\n/ /g;
1634 warn "starting: $show_cmd\n";
1636 if (defined $stdin) {
1637 my $child = open STDIN, "-|";
1638 defined $child or die "no fork: $!";
1640 print $stdin or die $!;
1641 close STDOUT or die $!;
1646 return system (@$args) if $opts->{fork};
1649 warn "ENV size is ".length(join(" ",%ENV));
1650 die "exec failed: $!: @$args";
1654 sub ban_node_by_slot {
1655 # Don't start any new jobsteps on this node for 60 seconds
1657 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1658 $slot[$slotid]->{node}->{hold_count}++;
1659 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1664 my ($lockfile, $error_message) = @_;
1665 open L, ">", $lockfile or croak("$lockfile: $!");
1666 if (!flock L, LOCK_EX|LOCK_NB) {
1667 croak("Can't lock $lockfile: $error_message\n");
1671 sub find_docker_image {
1672 # Given a Keep locator, check to see if it contains a Docker image.
1673 # If so, return its stream name and Docker hash.
1674 # If not, return undef for both values.
1675 my $locator = shift;
1676 my ($streamname, $filename);
1677 my $image = api_call("collections/get", uuid => $locator);
1679 foreach my $line (split(/\n/, $image->{manifest_text})) {
1680 my @tokens = split(/\s+/, $line);
1682 $streamname = shift(@tokens);
1683 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1684 if (defined($filename)) {
1685 return (undef, undef); # More than one file in the Collection.
1687 $filename = (split(/:/, $filedata, 3))[2];
1692 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1693 return ($streamname, $1);
1695 return (undef, undef);
1700 # Calculate the number of times an operation should be retried,
1701 # assuming exponential backoff, and that we're willing to retry as
1702 # long as tasks have been running. Enforce a minimum of 3 retries.
1703 my ($starttime, $endtime, $timediff, $retries);
1705 $starttime = $jobstep[0]->{starttime};
1706 $endtime = $jobstep[-1]->{finishtime};
1708 if (!defined($starttime)) {
1710 } elsif (!defined($endtime)) {
1711 $timediff = time - $starttime;
1713 $timediff = ($endtime - $starttime) - (time - $endtime);
1715 if ($timediff > 0) {
1716 $retries = int(log($timediff) / log(2));
1718 $retries = 1; # Use the minimum.
1720 return ($retries > 3) ? $retries : 3;
1724 # Pass in two function references.
1725 # This method will be called with the remaining arguments.
1726 # If it dies, retry it with exponential backoff until it succeeds,
1727 # or until the current retry_count is exhausted. After each failure
1728 # that can be retried, the second function will be called with
1729 # the current try count (0-based), next try time, and error message.
1730 my $operation = shift;
1731 my $retry_callback = shift;
1732 my $retries = retry_count();
1733 foreach my $try_count (0..$retries) {
1734 my $next_try = time + (2 ** $try_count);
1735 my $result = eval { $operation->(@_); };
1738 } elsif ($try_count < $retries) {
1739 $retry_callback->($try_count, $next_try, $@);
1740 my $sleep_time = $next_try - time;
1741 sleep($sleep_time) if ($sleep_time > 0);
1744 # Ensure the error message ends in a newline, so Perl doesn't add
1745 # retry_op's line number to it.
1751 # Pass in a /-separated API method name, and arguments for it.
1752 # This function will call that method, retrying as needed until
1753 # the current retry_count is exhausted, with a log on the first failure.
1754 my $method_name = shift;
1755 my $log_api_retry = sub {
1756 my ($try_count, $next_try_at, $errmsg) = @_;
1757 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
1758 $errmsg =~ s/\s/ /g;
1759 $errmsg =~ s/\s+$//;
1761 if ($next_try_at < time) {
1762 $retry_msg = "Retrying.";
1764 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
1765 $retry_msg = "Retrying at $next_try_fmt.";
1767 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
1770 foreach my $key (split(/\//, $method_name)) {
1771 $method = $method->{$key};
1773 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
1777 # Given a $?, return a human-readable exit code string like "0" or
1778 # "1" or "0 with signal 1" or "1 with signal 11".
1779 my $exitcode = shift;
1780 my $s = $exitcode >> 8;
1781 if ($exitcode & 0x7f) {
1782 $s .= " with signal " . ($exitcode & 0x7f);
1784 if ($exitcode & 0x80) {
1785 $s .= " with core dump";
1790 sub handle_readall {
1791 # Pass in a glob reference to a file handle.
1792 # Read all its contents and return them as a string.
1793 my $fh_glob_ref = shift;
1795 return <$fh_glob_ref>;
1798 sub tar_filename_n {
1800 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
1803 sub add_git_archive {
1804 # Pass in a git archive command as a string or list, a la system().
1805 # This method will save its output to be included in the archive sent to the
1809 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
1810 croak("Failed to save git archive: $!");
1812 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
1814 waitpid($git_pid, 0);
1817 croak("Failed to save git archive: git exited " . exit_status_s($?));
1821 sub combined_git_archive {
1822 # Combine all saved tar archives into a single archive, then return its
1823 # contents in a string. Return undef if no archives have been saved.
1824 if ($git_tar_count < 1) {
1827 my $base_tar_name = tar_filename_n(1);
1828 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
1829 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
1830 if ($tar_exit != 0) {
1831 croak("Error preparing build archive: tar -A exited " .
1832 exit_status_s($tar_exit));
1835 if (!open(GIT_TAR, "<", $base_tar_name)) {
1836 croak("Could not open build archive: $!");
1838 my $tar_contents = handle_readall(\*GIT_TAR);
1840 return $tar_contents;
1846 # This is crunch-job's internal dispatch script. crunch-job running on the API
1847 # server invokes this script on individual compute nodes, or localhost if we're
1848 # running a job locally. It gets called in two modes:
1850 # * No arguments: Installation mode. Read a tar archive from the DATA
1851 # file handle; it includes the Crunch script's source code, and
1852 # maybe SDKs as well. Those should be installed in the proper
1853 # locations. This runs outside of any Docker container, so don't try to
1854 # introspect Crunch's runtime environment.
1856 # * With arguments: Crunch script run mode. This script should set up the
1857 # environment, then run the command specified in the arguments. This runs
1858 # inside any Docker container.
1861 use File::Path qw( make_path remove_tree );
1862 use POSIX qw(getcwd);
1864 use constant TASK_TEMPFAIL => 111;
1866 # Map SDK subdirectories to the path environments they belong to.
1867 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
1869 my $destdir = $ENV{"CRUNCH_SRC"};
1870 my $commit = $ENV{"CRUNCH_SRC_COMMIT"};
1871 my $repo = $ENV{"CRUNCH_SRC_URL"};
1872 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
1873 my $job_work = $ENV{"JOB_WORK"};
1874 my $task_work = $ENV{"TASK_WORK"};
1876 for my $dir ($destdir, $job_work, $task_work) {
1879 -e $dir or die "Failed to create temporary directory ($dir): $!";
1884 remove_tree($task_work, {keep_root => 1});
1887 open(STDOUT_ORIG, ">&", STDOUT);
1888 open(STDERR_ORIG, ">&", STDERR);
1889 open(STDOUT, ">>", "$destdir.log");
1890 open(STDERR, ">&", STDOUT);
1892 ### Crunch script run mode
1894 # We want to do routine logging during task 0 only. This gives the user
1895 # the information they need, but avoids repeating the information for every
1898 if ($ENV{TASK_SEQUENCE} eq "0") {
1901 printf STDERR_ORIG "[Crunch] $msg\n", @_;
1907 my $python_src = "$install_dir/python";
1908 my $venv_dir = "$job_work/.arvados.venv";
1909 my $venv_built = -e "$venv_dir/bin/activate";
1910 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
1911 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
1912 "--python=python2.7", $venv_dir);
1913 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
1915 $Log->("Built Python SDK virtualenv");
1918 my $pip_bin = "pip";
1920 $Log->("Running in Python SDK virtualenv");
1921 $pip_bin = "$venv_dir/bin/pip";
1922 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
1923 @ARGV = ("/bin/sh", "-ec",
1924 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
1925 } elsif (-d $python_src) {
1926 $Log->("Warning: virtualenv not found inside Docker container default " .
1927 "\$PATH. Can't install Python SDK.");
1930 my $pkgs = `(\Q$pip_bin\E freeze 2>/dev/null | grep arvados) || dpkg-query --show '*arvados*'`;
1932 $Log->("Using Arvados SDK:");
1933 foreach my $line (split /\n/, $pkgs) {
1937 $Log->("Arvados SDK packages not found");
1940 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
1941 my $sdk_path = "$install_dir/$sdk_dir";
1943 if ($ENV{$sdk_envkey}) {
1944 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
1946 $ENV{$sdk_envkey} = $sdk_path;
1948 $Log->("Arvados SDK added to %s", $sdk_envkey);
1954 open(STDOUT, ">&", STDOUT_ORIG);
1955 open(STDERR, ">&", STDERR_ORIG);
1957 die "Cannot exec `@ARGV`: $!";
1960 ### Installation mode
1961 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
1963 if (readlink ("$destdir.commit") eq $commit && -d $destdir) {
1964 # This version already installed -> nothing to do.
1968 unlink "$destdir.commit";
1971 if (!open(TARX, "|-", "tar", "-xC", $destdir)) {
1972 die "Error launching 'tar -xC $destdir': $!";
1974 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
1975 # get SIGPIPE. We must feed it data incrementally.
1977 while (read(DATA, $tar_input, 65536)) {
1978 print TARX $tar_input;
1981 die "'tar -xC $destdir' exited $?: $!";
1986 my $sdk_root = "$destdir/.arvados.sdk/sdk";
1988 foreach my $sdk_lang (("python",
1989 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
1990 if (-d "$sdk_root/$sdk_lang") {
1991 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
1992 die "Failed to install $sdk_lang SDK: $!";
1998 my $python_dir = "$install_dir/python";
1999 if ((-d $python_dir) and can_run("python2.7") and
2000 (system("python2.7", "$python_dir/setup.py", "--quiet", "egg_info") != 0)) {
2001 # egg_info failed, probably when it asked git for a build tag.
2002 # Specify no build tag.
2003 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2004 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2008 if (-e "$destdir/crunch_scripts/install") {
2009 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2010 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2012 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2013 } elsif (-e "./install.sh") {
2014 shell_or_die (undef, "./install.sh", $install_dir);
2018 unlink "$destdir.commit.new";
2019 symlink ($commit, "$destdir.commit.new") or die "$destdir.commit.new: $!";
2020 rename ("$destdir.commit.new", "$destdir.commit") or die "$destdir.commit: $!";
2026 my $command_name = shift;
2027 open(my $which, "-|", "which", $command_name);
2028 while (<$which>) { }
2035 my $exitcode = shift;
2037 if ($ENV{"DEBUG"}) {
2038 print STDERR "@_\n";
2040 if (system (@_) != 0) {
2043 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2044 open STDERR, ">&STDERR_ORIG";
2045 system ("cat $destdir.log >&2");
2046 warn "@_ failed ($err): $exitstatus";
2047 if (defined($exitcode)) {
2051 exit (($code >> 8) || 1);