2 # -*- mode: perl; perl-indent-level: 2; indent-tabs-mode: nil; -*-
6 crunch-job: Execute job steps, save snapshots as requested, collate output.
10 Obtain job details from Arvados, run tasks on compute nodes (typically
11 invoked by scheduler on controller):
13 crunch-job --job x-y-z --git-dir /path/to/repo/.git
15 Obtain job details from command line, run tasks on local machine
16 (typically invoked by application or developer on VM):
18 crunch-job --job '{"script_version":"/path/to/working/tree","script":"scriptname",...}'
20 crunch-job --job '{"repository":"https://github.com/curoverse/arvados.git","script_version":"master","script":"scriptname",...}'
28 If the job is already locked, steal the lock and run it anyway.
32 Path to a .git directory (or a git URL) where the commit given in the
33 job's C<script_version> attribute is to be found. If this is I<not>
34 given, the job's C<repository> attribute will be used.
38 Arvados API authorization token to use during the course of the job.
42 Do not clear per-job/task temporary directories during initial job
43 setup. This can speed up development and debugging when running jobs
48 UUID of the job to run, or a JSON-encoded job resource without a
49 UUID. If the latter is given, a new job object will be created.
53 =head1 RUNNING JOBS LOCALLY
55 crunch-job's log messages appear on stderr along with the job tasks'
56 stderr streams. The log is saved in Keep at each checkpoint and when
59 If the job succeeds, the job's output locator is printed on stdout.
61 While the job is running, the following signals are accepted:
65 =item control-C, SIGINT, SIGQUIT
67 Save a checkpoint, terminate any job tasks that are running, and stop.
71 Save a checkpoint and continue.
75 Refresh node allocation (i.e., check whether any nodes have been added
76 or unallocated) and attributes of the Job record that should affect
77 behavior (e.g., cancel job if cancelled_at becomes non-nil).
85 use POSIX ':sys_wait_h';
86 use POSIX qw(strftime);
87 use Fcntl qw(F_GETFL F_SETFL O_NONBLOCK);
91 use Digest::MD5 qw(md5_hex);
97 use File::Path qw( make_path remove_tree );
99 use constant TASK_TEMPFAIL => 111;
100 use constant EX_TEMPFAIL => 75;
101 use constant EX_RETRY_UNLOCKED => 93;
103 $ENV{"TMPDIR"} ||= "/tmp";
104 unless (defined $ENV{"CRUNCH_TMP"}) {
105 $ENV{"CRUNCH_TMP"} = $ENV{"TMPDIR"} . "/crunch-job";
106 if ($ENV{"USER"} ne "crunch" && $< != 0) {
107 # use a tmp dir unique for my uid
108 $ENV{"CRUNCH_TMP"} .= "-$<";
112 # Create the tmp directory if it does not exist
113 if ( ! -d $ENV{"CRUNCH_TMP"} ) {
114 make_path $ENV{"CRUNCH_TMP"} or die "Failed to create temporary working directory: " . $ENV{"CRUNCH_TMP"};
117 $ENV{"JOB_WORK"} = $ENV{"CRUNCH_TMP"} . "/work";
118 $ENV{"CRUNCH_INSTALL"} = "$ENV{CRUNCH_TMP}/opt";
119 $ENV{"CRUNCH_WORK"} = $ENV{"JOB_WORK"}; # deprecated
120 mkdir ($ENV{"JOB_WORK"});
129 my $docker_bin = "docker.io";
130 GetOptions('force-unlock' => \$force_unlock,
131 'git-dir=s' => \$git_dir,
132 'job=s' => \$jobspec,
133 'job-api-token=s' => \$job_api_token,
134 'no-clear-tmp' => \$no_clear_tmp,
135 'resume-stash=s' => \$resume_stash,
136 'docker-bin=s' => \$docker_bin,
139 if (defined $job_api_token) {
140 $ENV{ARVADOS_API_TOKEN} = $job_api_token;
143 my $have_slurm = exists $ENV{SLURM_JOB_ID} && exists $ENV{SLURM_NODELIST};
148 $main::ENV{CRUNCH_DEBUG} = 1;
152 $main::ENV{CRUNCH_DEBUG} = 0;
155 my $arv = Arvados->new('apiVersion' => 'v1');
164 if ($jobspec =~ /^[-a-z\d]+$/)
166 # $jobspec is an Arvados UUID, not a JSON job specification
167 $Job = api_call("jobs/get", uuid => $jobspec);
172 $local_job = JSON::decode_json($jobspec);
176 # Make sure our workers (our slurm nodes, localhost, or whatever) are
177 # at least able to run basic commands: they aren't down or severely
180 if (($Job || $local_job)->{docker_image_locator}) {
181 $cmd = [$docker_bin, 'ps', '-q'];
183 Log(undef, "Sanity check is `@$cmd`");
184 srun(["srun", "--nodes=\Q$ENV{SLURM_NNODES}\E", "--ntasks-per-node=1"],
188 Log(undef, "Sanity check failed: ".exit_status_s($?));
191 Log(undef, "Sanity check OK");
194 my $User = api_call("users/current");
197 if (!$force_unlock) {
198 # Claim this job, and make sure nobody else does
199 eval { api_call("jobs/lock", uuid => $Job->{uuid}); };
201 Log(undef, "Error while locking job, exiting ".EX_TEMPFAIL);
210 map { croak ("No $_ specified") unless $local_job->{$_} }
211 qw(script script_version script_parameters);
214 $local_job->{'is_locked_by_uuid'} = $User->{'uuid'};
215 $local_job->{'started_at'} = gmtime;
216 $local_job->{'state'} = 'Running';
218 $Job = api_call("jobs/create", job => $local_job);
220 $job_id = $Job->{'uuid'};
222 my $keep_logfile = $job_id . '.log.txt';
223 log_writer_start($keep_logfile);
225 $Job->{'runtime_constraints'} ||= {};
226 $Job->{'runtime_constraints'}->{'max_tasks_per_node'} ||= 0;
227 my $max_ncpus = $Job->{'runtime_constraints'}->{'max_tasks_per_node'};
229 my $gem_versions = `gem list --quiet arvados-cli 2>/dev/null`;
231 $gem_versions =~ s/^arvados-cli \(/ with arvados-cli Gem version(s) /;
232 chomp($gem_versions);
233 chop($gem_versions); # Closing parentheses
238 "running from " . ((-e $0) ? realpath($0) : "stdin") . $gem_versions);
240 Log (undef, "check slurm allocation");
243 # Should use $ENV{SLURM_TASKS_PER_NODE} instead of sinfo? (eg. "4(x3),2,4(x2)")
247 my $localcpus = 0 + `grep -cw ^processor /proc/cpuinfo` || 1;
248 push @sinfo, "$localcpus localhost";
250 if (exists $ENV{SLURM_NODELIST})
252 push @sinfo, `sinfo -h --format='%c %N' --nodes=\Q$ENV{SLURM_NODELIST}\E`;
256 my ($ncpus, $slurm_nodelist) = split;
257 $ncpus = $max_ncpus if $max_ncpus && $ncpus > $max_ncpus;
260 while ($slurm_nodelist =~ s/^([^\[,]+?(\[.*?\])?)(,|$)//)
263 if ($nodelist =~ /\[((\d+)(-(\d+))?(,(\d+)(-(\d+))?)*)\]/)
266 foreach (split (",", $ranges))
279 push @nodelist, map {
281 $n =~ s/\[[-,\d]+\]/$_/;
288 push @nodelist, $nodelist;
291 foreach my $nodename (@nodelist)
293 Log (undef, "node $nodename - $ncpus slots");
294 my $node = { name => $nodename,
296 # The number of consecutive times a task has been dispatched
297 # to this node and failed.
299 # The number of consecutive times that SLURM has reported
300 # a node failure since the last successful task.
302 # Don't dispatch work to this node until this time
303 # (in seconds since the epoch) has passed.
305 foreach my $cpu (1..$ncpus)
307 push @slot, { node => $node,
311 push @node, @nodelist;
316 # Ensure that we get one jobstep running on each allocated node before
317 # we start overloading nodes with concurrent steps
319 @slot = sort { $a->{cpu} <=> $b->{cpu} } @slot;
322 $Job->update_attributes(
323 'tasks_summary' => { 'failed' => 0,
328 Log (undef, "start");
329 $SIG{'INT'} = sub { $main::please_freeze = 1; };
330 $SIG{'QUIT'} = sub { $main::please_freeze = 1; };
331 $SIG{'TERM'} = \&croak;
332 $SIG{'TSTP'} = sub { $main::please_freeze = 1; };
333 $SIG{'ALRM'} = sub { $main::please_info = 1; };
334 $SIG{'CONT'} = sub { $main::please_continue = 1; };
335 $SIG{'HUP'} = sub { $main::please_refresh = 1; };
337 $main::please_freeze = 0;
338 $main::please_info = 0;
339 $main::please_continue = 0;
340 $main::please_refresh = 0;
341 my $jobsteps_must_output_keys = 0; # becomes 1 when any task outputs a key
343 grep { $ENV{$1} = $2 if /^(NOCACHE.*?)=(.*)/ } split ("\n", $$Job{knobs});
344 $ENV{"CRUNCH_JOB_UUID"} = $job_id;
345 $ENV{"JOB_UUID"} = $job_id;
348 my @jobstep_todo = ();
349 my @jobstep_done = ();
350 my @jobstep_tomerge = ();
351 my $jobstep_tomerge_level = 0;
352 my $squeue_checked = 0;
353 my $latest_refresh = scalar time;
357 if (defined $Job->{thawedfromkey})
359 thaw ($Job->{thawedfromkey});
363 my $first_task = api_call("job_tasks/create", job_task => {
364 'job_uuid' => $Job->{'uuid'},
369 push @jobstep, { 'level' => 0,
371 'arvados_task' => $first_task,
373 push @jobstep_todo, 0;
379 must_lock_now("$ENV{CRUNCH_TMP}/.lock", "a job is already running here.");
382 my $build_script = handle_readall(\*DATA);
383 my $nodelist = join(",", @node);
384 my $git_tar_count = 0;
386 if (!defined $no_clear_tmp) {
387 # Clean out crunch_tmp/work, crunch_tmp/opt, crunch_tmp/src*
388 Log (undef, "Clean work dirs");
390 my $cleanpid = fork();
393 # Find FUSE mounts under $CRUNCH_TMP and unmount them.
394 # Then clean up work directories.
395 # TODO: When #5036 is done and widely deployed, we can limit mount's
396 # -t option to simply fuse.keep.
397 srun (["srun", "--nodelist=$nodelist", "-D", $ENV{'TMPDIR'}],
398 ['bash', '-ec', '-o', 'pipefail', 'mount -t fuse,fuse.keep | awk "(index(\$3, \"$CRUNCH_TMP\") == 1){print \$3}" | xargs -r -n 1 fusermount -u -z; sleep 1; rm -rf $JOB_WORK $CRUNCH_INSTALL $CRUNCH_TMP/task $CRUNCH_TMP/src* $CRUNCH_TMP/*.cid']);
403 last if $cleanpid == waitpid (-1, WNOHANG);
404 freeze_if_want_freeze ($cleanpid);
405 select (undef, undef, undef, 0.1);
408 Log(undef, "Clean work dirs: exit ".exit_status_s($?));
409 exit(EX_RETRY_UNLOCKED);
413 # If this job requires a Docker image, install that.
414 my ($docker_locator, $docker_stream, $docker_hash, $docker_limitmem, $dockeruserarg);
415 if ($docker_locator = $Job->{docker_image_locator}) {
416 ($docker_stream, $docker_hash) = find_docker_image($docker_locator);
419 croak("No Docker image hash found from locator $docker_locator");
421 $docker_stream =~ s/^\.//;
422 my $docker_install_script = qq{
423 if ! $docker_bin images -q --no-trunc --all | grep -qxF \Q$docker_hash\E; then
424 arv-get \Q$docker_locator$docker_stream/$docker_hash.tar\E | $docker_bin load
427 my $docker_pid = fork();
428 if ($docker_pid == 0)
430 srun (["srun", "--nodelist=" . join(',', @node)],
431 ["/bin/sh", "-ec", $docker_install_script]);
436 last if $docker_pid == waitpid (-1, WNOHANG);
437 freeze_if_want_freeze ($docker_pid);
438 select (undef, undef, undef, 0.1);
442 croak("Installing Docker image from $docker_locator exited "
446 # Determine whether this version of Docker supports memory+swap limits.
447 srun(["srun", "--nodelist=" . $node[0]],
448 ["/bin/sh", "-ec", "$docker_bin run --help | grep -qe --memory-swap="],
450 $docker_limitmem = ($? == 0);
452 # Find a non-root Docker user to use.
453 # Tries the default user for the container, then 'crunch', then 'nobody',
454 # testing for whether the actual user id is non-zero. This defends against
455 # mistakes but not malice, but we intend to harden the security in the future
456 # so we don't want anyone getting used to their jobs running as root in their
458 my @tryusers = ("", "crunch", "nobody");
459 foreach my $try_user (@tryusers) {
461 if ($try_user eq "") {
462 Log(undef, "Checking if container default user is not UID 0");
465 Log(undef, "Checking if user '$try_user' is not UID 0");
466 $try_user_arg = "--user=$try_user";
468 srun(["srun", "--nodelist=" . $node[0]],
470 "a=`$docker_bin run $try_user_arg $docker_hash id --user` && " .
474 $dockeruserarg = $try_user_arg;
475 if ($try_user eq "") {
476 Log(undef, "Container will run with default user");
478 Log(undef, "Container will run with $dockeruserarg");
484 if (!defined $dockeruserarg) {
485 croak("Could not find a user in container that is not UID 0 (tried default user, @tryusers) or there was a problem running 'id' in the container.");
488 if ($Job->{arvados_sdk_version}) {
489 # The job also specifies an Arvados SDK version. Add the SDKs to the
490 # tar file for the build script to install.
491 Log(undef, sprintf("Packing Arvados SDK version %s for installation",
492 $Job->{arvados_sdk_version}));
493 add_git_archive("git", "--git-dir=$git_dir", "archive",
494 "--prefix=.arvados.sdk/",
495 $Job->{arvados_sdk_version}, "sdk");
499 if (!defined $git_dir && $Job->{'script_version'} =~ m{^/}) {
500 # If script_version looks like an absolute path, *and* the --git-dir
501 # argument was not given -- which implies we were not invoked by
502 # crunch-dispatch -- we will use the given path as a working
503 # directory instead of resolving script_version to a git commit (or
504 # doing anything else with git).
505 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{'script_version'};
506 $ENV{"CRUNCH_SRC"} = $Job->{'script_version'};
509 # Resolve the given script_version to a git commit sha1. Also, if
510 # the repository is remote, clone it into our local filesystem: this
511 # ensures "git archive" will work, and is necessary to reliably
512 # resolve a symbolic script_version like "master^".
513 $ENV{"CRUNCH_SRC"} = "$ENV{CRUNCH_TMP}/src";
515 Log (undef, "Looking for version ".$Job->{script_version}." from repository ".$Job->{repository});
517 $ENV{"CRUNCH_SRC_COMMIT"} = $Job->{script_version};
519 # If we're running under crunch-dispatch, it will have already
520 # pulled the appropriate source tree into its own repository, and
521 # given us that repo's path as $git_dir.
523 # If we're running a "local" job, we might have to fetch content
524 # from a remote repository.
526 # (Currently crunch-dispatch gives a local path with --git-dir, but
527 # we might as well accept URLs there too in case it changes its
529 my $repo = $git_dir || $Job->{'repository'};
531 # Repository can be remote or local. If remote, we'll need to fetch it
532 # to a local dir before doing `git log` et al.
535 if ($repo =~ m{://|^[^/]*:}) {
536 # $repo is a git url we can clone, like git:// or https:// or
537 # file:/// or [user@]host:repo.git. Note "user/name@host:foo" is
538 # not recognized here because distinguishing that from a local
539 # path is too fragile. If you really need something strange here,
540 # use the ssh:// form.
541 $repo_location = 'remote';
542 } elsif ($repo =~ m{^\.*/}) {
543 # $repo is a local path to a git index. We'll also resolve ../foo
544 # to ../foo/.git if the latter is a directory. To help
545 # disambiguate local paths from named hosted repositories, this
546 # form must be given as ./ or ../ if it's a relative path.
547 if (-d "$repo/.git") {
548 $repo = "$repo/.git";
550 $repo_location = 'local';
552 # $repo is none of the above. It must be the name of a hosted
554 my $arv_repo_list = api_call("repositories/list",
555 'filters' => [['name','=',$repo]]);
556 my @repos_found = @{$arv_repo_list->{'items'}};
557 my $n_found = $arv_repo_list->{'serverResponse'}->{'items_available'};
559 Log(undef, "Repository '$repo' -> "
560 . join(", ", map { $_->{'uuid'} } @repos_found));
563 croak("Error: Found $n_found repositories with name '$repo'.");
565 $repo = $repos_found[0]->{'fetch_url'};
566 $repo_location = 'remote';
568 Log(undef, "Using $repo_location repository '$repo'");
569 $ENV{"CRUNCH_SRC_URL"} = $repo;
571 # Resolve given script_version (we'll call that $treeish here) to a
572 # commit sha1 ($commit).
573 my $treeish = $Job->{'script_version'};
575 if ($repo_location eq 'remote') {
576 # We minimize excess object-fetching by re-using the same bare
577 # repository in CRUNCH_TMP/.git for multiple crunch-jobs -- we
578 # just keep adding remotes to it as needed.
579 my $local_repo = $ENV{'CRUNCH_TMP'}."/.git";
580 my $gitcmd = "git --git-dir=\Q$local_repo\E";
582 # Set up our local repo for caching remote objects, making
584 if (!-d $local_repo) {
585 make_path($local_repo) or croak("Error: could not create $local_repo");
587 # This works (exits 0 and doesn't delete fetched objects) even
588 # if $local_repo is already initialized:
589 `$gitcmd init --bare`;
591 croak("Error: $gitcmd init --bare exited ".exit_status_s($?));
594 # If $treeish looks like a hash (or abbrev hash) we look it up in
595 # our local cache first, since that's cheaper. (We don't want to
596 # do that with tags/branches though -- those change over time, so
597 # they should always be resolved by the remote repo.)
598 if ($treeish =~ /^[0-9a-f]{7,40}$/s) {
599 # Hide stderr because it's normal for this to fail:
600 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E 2>/dev/null`;
602 # Careful not to resolve a branch named abcdeff to commit 1234567:
603 $sha1 =~ /^$treeish/ &&
604 $sha1 =~ /^([0-9a-f]{40})$/s) {
606 Log(undef, "Commit $commit already present in $local_repo");
610 if (!defined $commit) {
611 # If $treeish isn't just a hash or abbrev hash, or isn't here
612 # yet, we need to fetch the remote to resolve it correctly.
614 # First, remove all local heads. This prevents a name that does
615 # not exist on the remote from resolving to (or colliding with)
616 # a previously fetched branch or tag (possibly from a different
618 remove_tree("$local_repo/refs/heads", {keep_root => 1});
620 Log(undef, "Fetching objects from $repo to $local_repo");
621 `$gitcmd fetch --no-progress --tags ''\Q$repo\E \Q+refs/heads/*:refs/heads/*\E`;
623 croak("Error: `$gitcmd fetch` exited ".exit_status_s($?));
627 # Now that the data is all here, we will use our local repo for
628 # the rest of our git activities.
632 my $gitcmd = "git --git-dir=\Q$repo\E";
633 my $sha1 = `$gitcmd rev-list -n1 ''\Q$treeish\E`;
634 unless ($? == 0 && $sha1 =~ /^([0-9a-f]{40})$/) {
635 croak("`$gitcmd rev-list` exited "
637 .", '$treeish' not found, giving up");
640 Log(undef, "Version $treeish is commit $commit");
642 if ($commit ne $Job->{'script_version'}) {
643 # Record the real commit id in the database, frozentokey, logs,
644 # etc. -- instead of an abbreviation or a branch name which can
645 # become ambiguous or point to a different commit in the future.
646 if (!$Job->update_attributes('script_version' => $commit)) {
647 croak("Error: failed to update job's script_version attribute");
651 $ENV{"CRUNCH_SRC_COMMIT"} = $commit;
652 add_git_archive("$gitcmd archive ''\Q$commit\E");
655 my $git_archive = combined_git_archive();
656 if (!defined $git_archive) {
657 Log(undef, "Skip install phase (no git archive)");
659 Log(undef, "Warning: This probably means workers have no source tree!");
664 my $install_script_tries_left = 3;
665 for (my $attempts = 0; $attempts < 3; $attempts++) {
666 Log(undef, "Run install script on all workers");
668 my @srunargs = ("srun",
669 "--nodelist=$nodelist",
670 "-D", $ENV{'TMPDIR'}, "--job-name=$job_id");
671 my @execargs = ("sh", "-c",
672 "mkdir -p $ENV{CRUNCH_INSTALL} && cd $ENV{CRUNCH_TMP} && perl -");
674 $ENV{"CRUNCH_GIT_ARCHIVE_HASH"} = md5_hex($git_archive);
675 my ($install_stderr_r, $install_stderr_w);
676 pipe $install_stderr_r, $install_stderr_w or croak("pipe() failed: $!");
677 set_nonblocking($install_stderr_r);
678 my $installpid = fork();
679 if ($installpid == 0)
681 close($install_stderr_r);
682 fcntl($install_stderr_w, F_SETFL, 0) or croak($!); # no close-on-exec
683 open(STDOUT, ">&", $install_stderr_w);
684 open(STDERR, ">&", $install_stderr_w);
685 srun (\@srunargs, \@execargs, {}, $build_script . $git_archive);
688 close($install_stderr_w);
689 # Tell freeze_if_want_freeze how to kill the child, otherwise the
690 # "waitpid(installpid)" loop won't get interrupted by a freeze:
691 $proc{$installpid} = {};
693 # Track whether anything appears on stderr other than slurm errors
694 # ("srun: ...") and the "starting: ..." message printed by the
695 # srun subroutine itself:
696 my $stderr_anything_from_script = 0;
697 my $match_our_own_errors = '^(srun: error: |starting: \[)';
698 while ($installpid != waitpid(-1, WNOHANG)) {
699 freeze_if_want_freeze ($installpid);
700 # Wait up to 0.1 seconds for something to appear on stderr, then
701 # do a non-blocking read.
702 my $bits = fhbits($install_stderr_r);
703 select ($bits, undef, $bits, 0.1);
704 if (0 < sysread ($install_stderr_r, $stderr_buf, 8192, length($stderr_buf)))
706 while ($stderr_buf =~ /^(.*?)\n/) {
708 substr $stderr_buf, 0, 1+length($line), "";
709 Log(undef, "stderr $line");
710 if ($line !~ /$match_our_own_errors/) {
711 $stderr_anything_from_script = 1;
716 delete $proc{$installpid};
717 $install_exited = $?;
718 close($install_stderr_r);
719 if (length($stderr_buf) > 0) {
720 if ($stderr_buf !~ /$match_our_own_errors/) {
721 $stderr_anything_from_script = 1;
723 Log(undef, "stderr $stderr_buf")
726 Log (undef, "Install script exited ".exit_status_s($install_exited));
727 last if $install_exited == 0 || $main::please_freeze;
728 # If the install script fails but doesn't print an error message,
729 # the next thing anyone is likely to do is just run it again in
730 # case it was a transient problem like "slurm communication fails
731 # because the network isn't reliable enough". So we'll just do
732 # that ourselves (up to 3 attempts in total). OTOH, if there is an
733 # error message, the problem is more likely to have a real fix and
734 # we should fail the job so the fixing process can start, instead
735 # of doing 2 more attempts.
736 last if $stderr_anything_from_script;
739 foreach my $tar_filename (map { tar_filename_n($_); } (1..$git_tar_count)) {
740 unlink($tar_filename);
743 if ($install_exited != 0) {
748 foreach (qw (script script_version script_parameters runtime_constraints))
752 (ref($Job->{$_}) ? JSON::encode_json($Job->{$_}) : $Job->{$_}));
754 foreach (split (/\n/, $Job->{knobs}))
756 Log (undef, "knob " . $_);
761 $main::success = undef;
767 my $thisround_succeeded = 0;
768 my $thisround_failed = 0;
769 my $thisround_failed_multiple = 0;
770 my $working_slot_count = scalar(@slot);
772 @jobstep_todo = sort { $jobstep[$a]->{level} <=> $jobstep[$b]->{level}
773 or $a <=> $b } @jobstep_todo;
774 my $level = $jobstep[$jobstep_todo[0]]->{level};
776 my $initial_tasks_this_level = 0;
777 foreach my $id (@jobstep_todo) {
778 $initial_tasks_this_level++ if ($jobstep[$id]->{level} == $level);
781 # If the number of tasks scheduled at this level #T is smaller than the number
782 # of slots available #S, only use the first #T slots, or the first slot on
783 # each node, whichever number is greater.
785 # When we dispatch tasks later, we'll allocate whole-node resources like RAM
786 # based on these numbers. Using fewer slots makes more resources available
787 # to each individual task, which should normally be a better strategy when
788 # there are fewer of them running with less parallelism.
790 # Note that this calculation is not redone if the initial tasks at
791 # this level queue more tasks at the same level. This may harm
792 # overall task throughput for that level.
794 if ($initial_tasks_this_level < @node) {
795 @freeslot = (0..$#node);
796 } elsif ($initial_tasks_this_level < @slot) {
797 @freeslot = (0..$initial_tasks_this_level - 1);
799 @freeslot = (0..$#slot);
801 my $round_num_freeslots = scalar(@freeslot);
803 my %round_max_slots = ();
804 for (my $ii = $#freeslot; $ii >= 0; $ii--) {
805 my $this_slot = $slot[$freeslot[$ii]];
806 my $node_name = $this_slot->{node}->{name};
807 $round_max_slots{$node_name} ||= $this_slot->{cpu};
808 last if (scalar(keys(%round_max_slots)) >= @node);
811 Log(undef, "start level $level with $round_num_freeslots slots");
814 my $progress_is_dirty = 1;
815 my $progress_stats_updated = 0;
817 update_progress_stats();
821 for (my $todo_ptr = 0; $todo_ptr <= $#jobstep_todo; $todo_ptr ++)
823 # Don't create new tasks if we already know the job's final result.
824 last if defined($main::success);
826 my $id = $jobstep_todo[$todo_ptr];
827 my $Jobstep = $jobstep[$id];
828 if ($Jobstep->{level} != $level)
833 pipe $reader{$id}, "writer" or croak("pipe() failed: $!");
834 set_nonblocking($reader{$id});
836 my $childslot = $freeslot[0];
837 my $childnode = $slot[$childslot]->{node};
838 my $childslotname = join (".",
839 $slot[$childslot]->{node}->{name},
840 $slot[$childslot]->{cpu});
842 my $childpid = fork();
845 $SIG{'INT'} = 'DEFAULT';
846 $SIG{'QUIT'} = 'DEFAULT';
847 $SIG{'TERM'} = 'DEFAULT';
849 foreach (values (%reader))
853 fcntl ("writer", F_SETFL, 0) or croak ($!); # no close-on-exec
854 open(STDOUT,">&writer");
855 open(STDERR,">&writer");
860 delete $ENV{"GNUPGHOME"};
861 $ENV{"TASK_UUID"} = $Jobstep->{'arvados_task'}->{'uuid'};
862 $ENV{"TASK_QSEQUENCE"} = $id;
863 $ENV{"TASK_SEQUENCE"} = $level;
864 $ENV{"JOB_SCRIPT"} = $Job->{script};
865 while (my ($param, $value) = each %{$Job->{script_parameters}}) {
866 $param =~ tr/a-z/A-Z/;
867 $ENV{"JOB_PARAMETER_$param"} = $value;
869 $ENV{"TASK_SLOT_NODE"} = $slot[$childslot]->{node}->{name};
870 $ENV{"TASK_SLOT_NUMBER"} = $slot[$childslot]->{cpu};
871 $ENV{"TASK_WORK"} = $ENV{"CRUNCH_TMP"}."/task/$childslotname";
872 $ENV{"HOME"} = $ENV{"TASK_WORK"};
873 $ENV{"TASK_KEEPMOUNT"} = $ENV{"TASK_WORK"}.".keep";
874 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
875 $ENV{"CRUNCH_NODE_SLOTS"} = $round_max_slots{$ENV{TASK_SLOT_NODE}};
876 $ENV{"PATH"} = $ENV{"CRUNCH_INSTALL"} . "/bin:" . $ENV{"PATH"};
882 "--nodelist=".$childnode->{name},
883 qw(-n1 -c1 -N1 -D), $ENV{'TMPDIR'},
884 "--job-name=$job_id.$id.$$",
887 my $stdbuf = " stdbuf --output=0 --error=0 ";
889 my $arv_file_cache = "";
890 if (defined($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'})) {
891 $arv_file_cache = "--file-cache=" . ($Job->{'runtime_constraints'}->{'keep_cache_mb_per_task'} * 1024 * 1024);
895 "if [ -e $ENV{TASK_WORK} ]; then rm -rf $ENV{TASK_WORK}; fi; "
896 ."mkdir -p $ENV{CRUNCH_TMP} $ENV{JOB_WORK} $ENV{TASK_WORK} $ENV{TASK_KEEPMOUNT} "
897 ."&& cd $ENV{CRUNCH_TMP} "
898 # These environment variables get used explicitly later in
899 # $command. No tool is expected to read these values directly.
900 .q{&& MEM=$(awk '($1 == "MemTotal:"){print $2}' </proc/meminfo) }
901 .q{&& SWAP=$(awk '($1 == "SwapTotal:"){print $2}' </proc/meminfo) }
902 ."&& MEMLIMIT=\$(( (\$MEM * 95) / ($ENV{CRUNCH_NODE_SLOTS} * 100) )) "
903 ."&& let SWAPLIMIT=\$MEMLIMIT+\$SWAP ";
904 $command .= "&& exec arv-mount --by-pdh --crunchstat-interval=10 --allow-other $arv_file_cache $ENV{TASK_KEEPMOUNT} --exec ";
907 my $containername = "$Jobstep->{arvados_task}->{uuid}-$Jobstep->{failures}";
908 my $cidfile = "$ENV{CRUNCH_TMP}/$containername.cid";
909 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -cgroup-parent=docker -cgroup-cid=$cidfile -poll=10000 ";
910 $command .= "$docker_bin run --name=$containername --attach=stdout --attach=stderr --attach=stdin -i \Q$dockeruserarg\E --cidfile=$cidfile --sig-proxy ";
911 # We only set memory limits if Docker lets us limit both memory and swap.
912 # Memory limits alone have been supported longer, but subprocesses tend
913 # to get SIGKILL if they exceed that without any swap limit set.
914 # See #5642 for additional background.
915 if ($docker_limitmem) {
916 $command .= "--memory=\${MEMLIMIT}k --memory-swap=\${SWAPLIMIT}k ";
919 # The source tree and $destdir directory (which we have
920 # installed on the worker host) are available in the container,
921 # under the same path.
922 $command .= "--volume=\Q$ENV{CRUNCH_SRC}:$ENV{CRUNCH_SRC}:ro\E ";
923 $command .= "--volume=\Q$ENV{CRUNCH_INSTALL}:$ENV{CRUNCH_INSTALL}:ro\E ";
925 # Currently, we make arv-mount's mount point appear at /keep
926 # inside the container (instead of using the same path as the
927 # host like we do with CRUNCH_SRC and CRUNCH_INSTALL). However,
928 # crunch scripts and utilities must not rely on this. They must
929 # use $TASK_KEEPMOUNT.
930 $command .= "--volume=\Q$ENV{TASK_KEEPMOUNT}:/keep:ro\E ";
931 $ENV{TASK_KEEPMOUNT} = "/keep";
933 # TASK_WORK is almost exactly like a docker data volume: it
934 # starts out empty, is writable, and persists until no
935 # containers use it any more. We don't use --volumes-from to
936 # share it with other containers: it is only accessible to this
937 # task, and it goes away when this task stops.
939 # However, a docker data volume is writable only by root unless
940 # the mount point already happens to exist in the container with
941 # different permissions. Therefore, we [1] assume /tmp already
942 # exists in the image and is writable by the crunch user; [2]
943 # avoid putting TASK_WORK inside CRUNCH_TMP (which won't be
944 # writable if they are created by docker while setting up the
945 # other --volumes); and [3] create $TASK_WORK inside the
946 # container using $build_script.
947 $command .= "--volume=/tmp ";
948 $ENV{"TASK_WORK"} = "/tmp/crunch-job-task-work/$childslotname";
949 $ENV{"HOME"} = $ENV{"TASK_WORK"};
950 $ENV{"TASK_TMPDIR"} = $ENV{"TASK_WORK"}; # deprecated
952 # TODO: Share a single JOB_WORK volume across all task
953 # containers on a given worker node, and delete it when the job
954 # ends (and, in case that doesn't work, when the next job
957 # For now, use the same approach as TASK_WORK above.
958 $ENV{"JOB_WORK"} = "/tmp/crunch-job-work";
960 while (my ($env_key, $env_val) = each %ENV)
962 if ($env_key =~ /^(ARVADOS|CRUNCH|JOB|TASK)_/) {
963 $command .= "--env=\Q$env_key=$env_val\E ";
966 $command .= "--env=\QHOME=$ENV{HOME}\E ";
967 $command .= "\Q$docker_hash\E ";
969 if ($Job->{arvados_sdk_version}) {
971 $command .= "perl - \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E";
973 $command .= "/bin/sh -c \'python -c " .
974 '"from pkg_resources import get_distribution as get; print \"Using Arvados SDK version\", get(\"arvados-python-client\").version"' .
975 ">&2 2>/dev/null; " .
976 "mkdir -p \"$ENV{JOB_WORK}\" \"$ENV{TASK_WORK}\" && " .
977 "if which stdbuf >/dev/null ; then " .
978 " exec $stdbuf \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
980 " exec \Q$ENV{CRUNCH_SRC}/crunch_scripts/$Job->{script}\E ;" .
985 $command .= "crunchstat -cgroup-root=/sys/fs/cgroup -poll=10000 ";
987 $command .= "perl - $ENV{CRUNCH_SRC}/crunch_scripts/" . $Job->{"script"};
990 my @execargs = ('bash', '-c', $command);
991 srun (\@srunargs, \@execargs, undef, $build_script);
992 # exec() failed, we assume nothing happened.
993 die "srun() failed on build script\n";
996 if (!defined $childpid)
1003 $proc{$childpid} = { jobstep => $id,
1006 jobstepname => "$job_id.$id.$childpid",
1008 croak ("assert failed: \$slot[$childslot]->{'pid'} exists") if exists $slot[$childslot]->{pid};
1009 $slot[$childslot]->{pid} = $childpid;
1011 Log ($id, "job_task ".$Jobstep->{'arvados_task'}->{'uuid'});
1012 Log ($id, "child $childpid started on $childslotname");
1013 $Jobstep->{starttime} = time;
1014 $Jobstep->{node} = $childnode->{name};
1015 $Jobstep->{slotindex} = $childslot;
1016 delete $Jobstep->{stderr};
1017 delete $Jobstep->{finishtime};
1018 delete $Jobstep->{tempfail};
1020 $Jobstep->{'arvados_task'}->{started_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{starttime});
1021 $Jobstep->{'arvados_task'}->save;
1023 splice @jobstep_todo, $todo_ptr, 1;
1026 $progress_is_dirty = 1;
1030 ($round_num_freeslots > @freeslot && $todo_ptr+1 > $#jobstep_todo))
1032 last THISROUND if $main::please_freeze;
1033 if ($main::please_info)
1035 $main::please_info = 0;
1037 create_output_collection();
1039 update_progress_stats();
1044 if (!$gotsome || ($latest_refresh + 2 < scalar time))
1046 check_refresh_wanted();
1048 update_progress_stats();
1049 select (undef, undef, undef, 0.1);
1051 elsif (time - $progress_stats_updated >= 30 || $progress_is_dirty)
1053 update_progress_stats();
1055 $working_slot_count = scalar(grep { $_->{node}->{fail_count} == 0 &&
1056 $_->{node}->{hold_count} < 4 } @slot);
1057 if (($thisround_failed_multiple >= 8 && $thisround_succeeded == 0) ||
1058 ($thisround_failed_multiple >= 16 && $thisround_failed_multiple > $thisround_succeeded))
1060 my $message = "Repeated failure rate too high ($thisround_failed_multiple/"
1061 .($thisround_failed+$thisround_succeeded)
1062 .") -- giving up on this round";
1063 Log (undef, $message);
1067 # move slots from freeslot to holdslot (or back to freeslot) if necessary
1068 for (my $i=$#freeslot; $i>=0; $i--) {
1069 if ($slot[$freeslot[$i]]->{node}->{hold_until} > scalar time) {
1070 push @holdslot, (splice @freeslot, $i, 1);
1073 for (my $i=$#holdslot; $i>=0; $i--) {
1074 if ($slot[$holdslot[$i]]->{node}->{hold_until} <= scalar time) {
1075 push @freeslot, (splice @holdslot, $i, 1);
1079 # give up if no nodes are succeeding
1080 if ($working_slot_count < 1) {
1081 Log(undef, "Every node has failed -- giving up");
1088 push @freeslot, splice @holdslot;
1089 map { $slot[$freeslot[$_]]->{node}->{losing_streak} = 0 } (0..$#freeslot);
1092 Log (undef, "wait for last ".(scalar keys %proc)." children to finish");
1095 if ($main::please_continue) {
1096 $main::please_continue = 0;
1099 $main::please_info = 0, freeze(), create_output_collection(), save_meta(1) if $main::please_info;
1101 if (!reapchildren())
1103 check_refresh_wanted();
1105 update_progress_stats();
1106 select (undef, undef, undef, 0.1);
1107 killem (keys %proc) if $main::please_freeze;
1111 update_progress_stats();
1112 freeze_if_want_freeze();
1115 if (!defined $main::success)
1117 if (!@jobstep_todo) {
1119 } elsif ($working_slot_count < 1) {
1120 save_output_collection();
1122 exit(EX_RETRY_UNLOCKED);
1123 } elsif ($thisround_succeeded == 0 &&
1124 ($thisround_failed == 0 || $thisround_failed > 4)) {
1125 my $message = "stop because $thisround_failed tasks failed and none succeeded";
1126 Log (undef, $message);
1131 goto ONELEVEL if !defined $main::success;
1134 release_allocation();
1136 my $collated_output = save_output_collection();
1137 Log (undef, "finish");
1142 if ($collated_output && $main::success) {
1143 $final_state = 'Complete';
1145 $final_state = 'Failed';
1147 $Job->update_attributes('state' => $final_state);
1149 exit (($final_state eq 'Complete') ? 0 : 1);
1153 sub update_progress_stats
1155 $progress_stats_updated = time;
1156 return if !$progress_is_dirty;
1157 my ($todo, $done, $running) = (scalar @jobstep_todo,
1158 scalar @jobstep_done,
1159 scalar keys(%proc));
1160 $Job->{'tasks_summary'} ||= {};
1161 $Job->{'tasks_summary'}->{'todo'} = $todo;
1162 $Job->{'tasks_summary'}->{'done'} = $done;
1163 $Job->{'tasks_summary'}->{'running'} = $running;
1164 $Job->update_attributes('tasks_summary' => $Job->{'tasks_summary'});
1165 Log (undef, "status: $done done, $running running, $todo todo");
1166 $progress_is_dirty = 0;
1173 my $pid = waitpid (-1, WNOHANG);
1174 return 0 if $pid <= 0;
1176 my $whatslot = ($slot[$proc{$pid}->{slot}]->{node}->{name}
1178 . $slot[$proc{$pid}->{slot}]->{cpu});
1179 my $jobstepid = $proc{$pid}->{jobstep};
1180 my $elapsed = time - $proc{$pid}->{time};
1181 my $Jobstep = $jobstep[$jobstepid];
1183 my $childstatus = $?;
1184 my $exitvalue = $childstatus >> 8;
1185 my $exitinfo = "exit ".exit_status_s($childstatus);
1186 $Jobstep->{'arvados_task'}->reload;
1187 my $task_success = $Jobstep->{'arvados_task'}->{success};
1189 Log ($jobstepid, "child $pid on $whatslot $exitinfo success=$task_success");
1191 if (!defined $task_success) {
1192 # task did not indicate one way or the other --> fail
1193 Log($jobstepid, sprintf(
1194 "ERROR: Task process exited %s, but never updated its task record to indicate success and record its output.",
1195 exit_status_s($childstatus)));
1196 $Jobstep->{'arvados_task'}->{success} = 0;
1197 $Jobstep->{'arvados_task'}->save;
1204 $temporary_fail ||= $Jobstep->{tempfail};
1205 $temporary_fail ||= ($exitvalue == TASK_TEMPFAIL);
1207 ++$thisround_failed;
1208 ++$thisround_failed_multiple if $Jobstep->{'failures'} >= 1;
1210 # Check for signs of a failed or misconfigured node
1211 if (++$slot[$proc{$pid}->{slot}]->{node}->{losing_streak} >=
1212 2+$slot[$proc{$pid}->{slot}]->{node}->{ncpus}) {
1213 # Don't count this against jobstep failure thresholds if this
1214 # node is already suspected faulty and srun exited quickly
1215 if ($slot[$proc{$pid}->{slot}]->{node}->{hold_until} &&
1217 Log ($jobstepid, "blaming failure on suspect node " .
1218 $slot[$proc{$pid}->{slot}]->{node}->{name});
1219 $temporary_fail ||= 1;
1221 ban_node_by_slot($proc{$pid}->{slot});
1224 Log ($jobstepid, sprintf('failure (#%d, %s) after %d seconds',
1225 ++$Jobstep->{'failures'},
1226 $temporary_fail ? 'temporary' : 'permanent',
1229 if (!$temporary_fail || $Jobstep->{'failures'} >= 3) {
1230 # Give up on this task, and the whole job
1233 # Put this task back on the todo queue
1234 push @jobstep_todo, $jobstepid;
1235 $Job->{'tasks_summary'}->{'failed'}++;
1239 ++$thisround_succeeded;
1240 $slot[$proc{$pid}->{slot}]->{node}->{losing_streak} = 0;
1241 $slot[$proc{$pid}->{slot}]->{node}->{hold_until} = 0;
1242 $slot[$proc{$pid}->{slot}]->{node}->{fail_count} = 0;
1243 push @jobstep_done, $jobstepid;
1244 Log ($jobstepid, "success in $elapsed seconds");
1246 $Jobstep->{exitcode} = $childstatus;
1247 $Jobstep->{finishtime} = time;
1248 $Jobstep->{'arvados_task'}->{finished_at} = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($Jobstep->{finishtime});
1249 $Jobstep->{'arvados_task'}->save;
1250 process_stderr ($jobstepid, $task_success);
1251 Log ($jobstepid, sprintf("task output (%d bytes): %s",
1252 length($Jobstep->{'arvados_task'}->{output}),
1253 $Jobstep->{'arvados_task'}->{output}));
1255 close $reader{$jobstepid};
1256 delete $reader{$jobstepid};
1257 delete $slot[$proc{$pid}->{slot}]->{pid};
1258 push @freeslot, $proc{$pid}->{slot};
1261 if ($task_success) {
1263 my $newtask_list = [];
1264 my $newtask_results;
1266 $newtask_results = api_call(
1269 'created_by_job_task_uuid' => $Jobstep->{'arvados_task'}->{uuid}
1271 'order' => 'qsequence',
1272 'offset' => scalar(@$newtask_list),
1274 push(@$newtask_list, @{$newtask_results->{items}});
1275 } while (@{$newtask_results->{items}});
1276 foreach my $arvados_task (@$newtask_list) {
1278 'level' => $arvados_task->{'sequence'},
1280 'arvados_task' => $arvados_task
1282 push @jobstep, $jobstep;
1283 push @jobstep_todo, $#jobstep;
1287 $progress_is_dirty = 1;
1291 sub check_refresh_wanted
1293 my @stat = stat $ENV{"CRUNCH_REFRESH_TRIGGER"};
1294 if (@stat && $stat[9] > $latest_refresh) {
1295 $latest_refresh = scalar time;
1296 my $Job2 = api_call("jobs/get", uuid => $jobspec);
1297 for my $attr ('cancelled_at',
1298 'cancelled_by_user_uuid',
1299 'cancelled_by_client_uuid',
1301 $Job->{$attr} = $Job2->{$attr};
1303 if ($Job->{'state'} ne "Running") {
1304 if ($Job->{'state'} eq "Cancelled") {
1305 Log (undef, "Job cancelled at " . $Job->{'cancelled_at'} . " by user " . $Job->{'cancelled_by_user_uuid'});
1307 Log (undef, "Job state unexpectedly changed to " . $Job->{'state'});
1310 $main::please_freeze = 1;
1317 my $last_squeue_check = $squeue_checked;
1319 # Do not call `squeue` or check the kill list more than once every
1321 return if $last_squeue_check > time - 15;
1322 $squeue_checked = time;
1324 # Look for children from which we haven't received stderr data since
1325 # the last squeue check. If no such children exist, all procs are
1326 # alive and there's no need to even look at squeue.
1328 # As long as the crunchstat poll interval (10s) is shorter than the
1329 # squeue check interval (15s) this should make the squeue check an
1331 my $silent_procs = 0;
1332 for my $jobstep (values %proc)
1334 if ($jobstep->{stderr_at} < $last_squeue_check)
1339 return if $silent_procs == 0;
1341 # use killem() on procs whose killtime is reached
1342 while (my ($pid, $jobstep) = each %proc)
1344 if (exists $jobstep->{killtime}
1345 && $jobstep->{killtime} <= time
1346 && $jobstep->{stderr_at} < $last_squeue_check)
1349 if ($jobstep->{stderr_at}) {
1350 $sincewhen = " in last " . (time - $jobstep->{stderr_at}) . "s";
1352 Log($jobstep->{jobstep}, "killing orphaned srun process $pid (task not in slurm queue, no stderr received$sincewhen)");
1359 # here is an opportunity to check for mysterious problems with local procs
1363 # Get a list of steps still running. Note: squeue(1) says --steps
1364 # selects a format (which we override anyway) and allows us to
1365 # specify which steps we're interested in (which we don't).
1366 # Importantly, it also changes the meaning of %j from "job name" to
1367 # "step name" and (although this isn't mentioned explicitly in the
1368 # docs) switches from "one line per job" mode to "one line per step"
1369 # mode. Without it, we'd just get a list of one job, instead of a
1371 my @squeue = `squeue --jobs=\Q$ENV{SLURM_JOB_ID}\E --steps --format='%j' --noheader`;
1374 Log(undef, "warning: squeue exit status $? ($!)");
1379 # which of my jobsteps are running, according to squeue?
1381 for my $jobstepname (@squeue)
1383 $ok{$jobstepname} = 1;
1386 # Check for child procs >60s old and not mentioned by squeue.
1387 while (my ($pid, $jobstep) = each %proc)
1389 if ($jobstep->{time} < time - 60
1390 && $jobstep->{jobstepname}
1391 && !exists $ok{$jobstep->{jobstepname}}
1392 && !exists $jobstep->{killtime})
1394 # According to slurm, this task has ended (successfully or not)
1395 # -- but our srun child hasn't exited. First we must wait (30
1396 # seconds) in case this is just a race between communication
1397 # channels. Then, if our srun child process still hasn't
1398 # terminated, we'll conclude some slurm communication
1399 # error/delay has caused the task to die without notifying srun,
1400 # and we'll kill srun ourselves.
1401 $jobstep->{killtime} = time + 30;
1402 Log($jobstep->{jobstep}, "notice: task is not in slurm queue but srun process $pid has not exited");
1408 sub release_allocation
1412 Log (undef, "release job allocation");
1413 system "scancel $ENV{SLURM_JOB_ID}";
1421 foreach my $job (keys %reader)
1424 while (0 < sysread ($reader{$job}, $buf, 8192))
1426 print STDERR $buf if $ENV{CRUNCH_DEBUG};
1427 $jobstep[$job]->{stderr_at} = time;
1428 $jobstep[$job]->{stderr} .= $buf;
1429 preprocess_stderr ($job);
1430 if (length ($jobstep[$job]->{stderr}) > 16384)
1432 substr ($jobstep[$job]->{stderr}, 0, 8192) = "";
1441 sub preprocess_stderr
1445 while ($jobstep[$job]->{stderr} =~ /^(.*?)\n/) {
1447 substr $jobstep[$job]->{stderr}, 0, 1+length($line), "";
1448 Log ($job, "stderr $line");
1449 if ($line =~ /srun: error: (SLURM job $ENV{SLURM_JOB_ID} has expired|Unable to confirm allocation for job $ENV{SLURM_JOB_ID})/) {
1451 $main::please_freeze = 1;
1453 elsif ($line =~ /srun: error: Node failure on/) {
1454 my $job_slot_index = $jobstep[$job]->{slotindex};
1455 $slot[$job_slot_index]->{node}->{fail_count}++;
1456 $jobstep[$job]->{tempfail} = 1;
1457 ban_node_by_slot($job_slot_index);
1459 elsif ($line =~ /srun: error: (Unable to create job step|.*: Communication connection failure)/) {
1460 $jobstep[$job]->{tempfail} = 1;
1461 ban_node_by_slot($jobstep[$job]->{slotindex});
1463 elsif ($line =~ /arvados\.errors\.Keep/) {
1464 $jobstep[$job]->{tempfail} = 1;
1473 my $task_success = shift;
1474 preprocess_stderr ($job);
1477 Log ($job, "stderr $_");
1478 } split ("\n", $jobstep[$job]->{stderr});
1485 if (!open($keep, "-|", "arv-get", "--retries", retry_count(), $hash)) {
1486 Log(undef, "fetch_block run error from arv-get $hash: $!");
1489 my $output_block = "";
1492 my $bytes = sysread($keep, $buf, 1024 * 1024);
1493 if (!defined $bytes) {
1494 Log(undef, "fetch_block read error from arv-get: $!");
1495 $output_block = undef;
1497 } elsif ($bytes == 0) {
1498 # sysread returns 0 at the end of the pipe.
1501 # some bytes were read into buf.
1502 $output_block .= $buf;
1507 Log(undef, "fetch_block arv-get exited " . exit_status_s($?));
1508 $output_block = undef;
1510 return $output_block;
1513 # Create a collection by concatenating the output of all tasks (each
1514 # task's output is either a manifest fragment, a locator for a
1515 # manifest fragment stored in Keep, or nothing at all). Return the
1516 # portable_data_hash of the new collection.
1517 sub create_output_collection
1519 Log (undef, "collate");
1521 my ($child_out, $child_in);
1522 my $pid = open2($child_out, $child_in, 'python', '-c', q{
1525 print (arvados.api("v1").collections().
1526 create(body={"manifest_text": sys.stdin.read()}).
1527 execute(num_retries=int(sys.argv[1]))["portable_data_hash"])
1531 my $manifest_size = 0;
1535 my $output = $_->{'arvados_task'}->{output};
1536 next if (!defined($output));
1538 if ($output =~ /^[0-9a-f]{32}(\+\S+)*$/) {
1539 $next_write = fetch_block($output);
1541 $next_write = $output;
1543 if (defined($next_write)) {
1544 if (!defined(syswrite($child_in, $next_write))) {
1545 # There's been an error writing. Stop the loop.
1546 # We'll log details about the exit code later.
1549 $manifest_size += length($next_write);
1552 my $uuid = $_->{'arvados_task'}->{'uuid'};
1553 Log (undef, "Error retrieving '$output' output by task $task_idx ($uuid)");
1558 Log(undef, "collated output manifest text to send to API server is $manifest_size bytes with access tokens");
1561 my $s = IO::Select->new($child_out);
1562 if ($s->can_read(120)) {
1563 sysread($child_out, $joboutput, 1024 * 1024);
1566 Log(undef, "output collection creation exited " . exit_status_s($?));
1572 Log (undef, "timed out while creating output collection");
1573 foreach my $signal (2, 2, 2, 15, 15, 9) {
1574 kill($signal, $pid);
1575 last if waitpid($pid, WNOHANG) == -1;
1584 # Calls create_output_collection, logs the result, and returns it.
1585 # If that was successful, save that as the output in the job record.
1586 sub save_output_collection {
1587 my $collated_output = create_output_collection();
1589 if (!$collated_output) {
1590 Log(undef, "Failed to write output collection");
1593 Log(undef, "job output $collated_output");
1594 $Job->update_attributes('output' => $collated_output);
1596 return $collated_output;
1603 my $sig = 2; # SIGINT first
1604 if (exists $proc{$_}->{"sent_$sig"} &&
1605 time - $proc{$_}->{"sent_$sig"} > 4)
1607 $sig = 15; # SIGTERM if SIGINT doesn't work
1609 if (exists $proc{$_}->{"sent_$sig"} &&
1610 time - $proc{$_}->{"sent_$sig"} > 4)
1612 $sig = 9; # SIGKILL if SIGTERM doesn't work
1614 if (!exists $proc{$_}->{"sent_$sig"})
1616 Log ($proc{$_}->{jobstep}, "sending 2x signal $sig to pid $_");
1618 select (undef, undef, undef, 0.1);
1621 kill $sig, $_; # srun wants two SIGINT to really interrupt
1623 $proc{$_}->{"sent_$sig"} = time;
1624 $proc{$_}->{"killedafter"} = time - $proc{$_}->{"time"};
1634 vec($bits,fileno($_),1) = 1;
1640 # Send log output to Keep via arv-put.
1642 # $log_pipe_in and $log_pipe_out are the input and output filehandles to the arv-put pipe.
1643 # $log_pipe_out_buf is a string containing all output read from arv-put so far.
1644 # $log_pipe_out_select is an IO::Select object around $log_pipe_out.
1645 # $log_pipe_pid is the pid of the arv-put subprocess.
1647 # The only functions that should access these variables directly are:
1649 # log_writer_start($logfilename)
1650 # Starts an arv-put pipe, reading data on stdin and writing it to
1651 # a $logfilename file in an output collection.
1653 # log_writer_read_output([$timeout])
1654 # Read output from $log_pipe_out and append it to $log_pipe_out_buf.
1655 # Passes $timeout to the select() call, with a default of 0.01.
1656 # Returns the result of the last read() call on $log_pipe_out, or
1657 # -1 if read() wasn't called because select() timed out.
1658 # Only other log_writer_* functions should need to call this.
1660 # log_writer_send($txt)
1661 # Writes $txt to the output log collection.
1663 # log_writer_finish()
1664 # Closes the arv-put pipe and returns the output that it produces.
1666 # log_writer_is_active()
1667 # Returns a true value if there is currently a live arv-put
1668 # process, false otherwise.
1670 my ($log_pipe_in, $log_pipe_out, $log_pipe_out_buf, $log_pipe_out_select,
1673 sub log_writer_start($)
1675 my $logfilename = shift;
1676 $log_pipe_pid = open2($log_pipe_out, $log_pipe_in,
1680 '--filename', $logfilename,
1682 $log_pipe_out_buf = "";
1683 $log_pipe_out_select = IO::Select->new($log_pipe_out);
1686 sub log_writer_read_output {
1687 my $timeout = shift || 0.01;
1689 while ($read && $log_pipe_out_select->can_read($timeout)) {
1690 $read = read($log_pipe_out, $log_pipe_out_buf, 65536,
1691 length($log_pipe_out_buf));
1693 if (!defined($read)) {
1694 Log(undef, "error reading log manifest from arv-put: $!");
1699 sub log_writer_send($)
1702 print $log_pipe_in $txt;
1703 log_writer_read_output();
1706 sub log_writer_finish()
1708 return unless $log_pipe_pid;
1710 close($log_pipe_in);
1712 my $logger_failed = 0;
1713 my $read_result = log_writer_read_output(120);
1714 if ($read_result == -1) {
1715 $logger_failed = -1;
1716 Log (undef, "timed out reading from 'arv-put'");
1717 } elsif ($read_result != 0) {
1718 $logger_failed = -2;
1719 Log(undef, "failed to read arv-put log manifest to EOF");
1722 waitpid($log_pipe_pid, 0);
1724 $logger_failed ||= $?;
1725 Log(undef, "log_writer_finish: arv-put exited " . exit_status_s($?))
1728 close($log_pipe_out);
1729 my $arv_put_output = $logger_failed ? undef : $log_pipe_out_buf;
1730 $log_pipe_pid = $log_pipe_in = $log_pipe_out = $log_pipe_out_buf =
1731 $log_pipe_out_select = undef;
1733 return $arv_put_output;
1736 sub log_writer_is_active() {
1737 return $log_pipe_pid;
1740 sub Log # ($jobstep_id, $logmessage)
1742 if ($_[1] =~ /\n/) {
1743 for my $line (split (/\n/, $_[1])) {
1748 my $fh = select STDERR; $|=1; select $fh;
1749 my $message = sprintf ("%s %d %s %s", $job_id, $$, @_);
1750 $message =~ s{([^ -\176])}{"\\" . sprintf ("%03o", ord($1))}ge;
1753 if (log_writer_is_active() || -t STDERR) {
1754 my @gmtime = gmtime;
1755 $datetime = sprintf ("%04d-%02d-%02d_%02d:%02d:%02d",
1756 $gmtime[5]+1900, $gmtime[4]+1, @gmtime[3,2,1,0]);
1758 print STDERR ((-t STDERR) ? ($datetime." ".$message) : $message);
1760 if (log_writer_is_active()) {
1761 log_writer_send($datetime . " " . $message);
1768 my ($package, $file, $line) = caller;
1769 my $message = "@_ at $file line $line\n";
1770 Log (undef, $message);
1771 freeze() if @jobstep_todo;
1772 create_output_collection() if @jobstep_todo;
1782 if ($Job->{'state'} eq 'Cancelled') {
1783 $Job->update_attributes('finished_at' => scalar gmtime);
1785 $Job->update_attributes('state' => 'Failed');
1792 my $justcheckpoint = shift; # false if this will be the last meta saved
1793 return if $justcheckpoint; # checkpointing is not relevant post-Warehouse.pm
1794 return unless log_writer_is_active();
1795 my $log_manifest = log_writer_finish();
1796 return unless defined($log_manifest);
1799 my $prev_log_coll = api_call("collections/get", uuid => $Job->{log});
1800 $log_manifest = $prev_log_coll->{manifest_text} . $log_manifest;
1803 my $log_coll = api_call(
1804 "collections/create", ensure_unique_name => 1, collection => {
1805 manifest_text => $log_manifest,
1806 owner_uuid => $Job->{owner_uuid},
1807 name => sprintf("Log from %s job %s", $Job->{script}, $Job->{uuid}),
1809 Log(undef, "log collection is " . $log_coll->{portable_data_hash});
1810 $Job->update_attributes('log' => $log_coll->{portable_data_hash});
1814 sub freeze_if_want_freeze
1816 if ($main::please_freeze)
1818 release_allocation();
1821 # kill some srun procs before freeze+stop
1822 map { $proc{$_} = {} } @_;
1825 killem (keys %proc);
1826 select (undef, undef, undef, 0.1);
1828 while (($died = waitpid (-1, WNOHANG)) > 0)
1830 delete $proc{$died};
1835 create_output_collection();
1845 Log (undef, "Freeze not implemented");
1852 croak ("Thaw not implemented");
1868 $s =~ s{\\(.)}{$1 eq "n" ? "\n" : $1}ge;
1875 my $srunargs = shift;
1876 my $execargs = shift;
1877 my $opts = shift || {};
1879 my $args = $have_slurm ? [@$srunargs, @$execargs] : $execargs;
1881 $Data::Dumper::Terse = 1;
1882 $Data::Dumper::Indent = 0;
1883 my $show_cmd = Dumper($args);
1884 $show_cmd =~ s/(TOKEN\\*=)[^\s\']+/${1}[...]/g;
1885 $show_cmd =~ s/\n/ /g;
1886 if ($opts->{fork}) {
1887 Log(undef, "starting: $show_cmd");
1889 # This is a child process: parent is in charge of reading our
1890 # stderr and copying it to Log() if needed.
1891 warn "starting: $show_cmd\n";
1894 if (defined $stdin) {
1895 my $child = open STDIN, "-|";
1896 defined $child or die "no fork: $!";
1898 print $stdin or die $!;
1899 close STDOUT or die $!;
1904 return system (@$args) if $opts->{fork};
1907 warn "ENV size is ".length(join(" ",%ENV));
1908 die "exec failed: $!: @$args";
1912 sub ban_node_by_slot {
1913 # Don't start any new jobsteps on this node for 60 seconds
1915 $slot[$slotid]->{node}->{hold_until} = 60 + scalar time;
1916 $slot[$slotid]->{node}->{hold_count}++;
1917 Log (undef, "backing off node " . $slot[$slotid]->{node}->{name} . " for 60 seconds");
1922 my ($lockfile, $error_message) = @_;
1923 open L, ">", $lockfile or croak("$lockfile: $!");
1924 if (!flock L, LOCK_EX|LOCK_NB) {
1925 croak("Can't lock $lockfile: $error_message\n");
1929 sub find_docker_image {
1930 # Given a Keep locator, check to see if it contains a Docker image.
1931 # If so, return its stream name and Docker hash.
1932 # If not, return undef for both values.
1933 my $locator = shift;
1934 my ($streamname, $filename);
1935 my $image = api_call("collections/get", uuid => $locator);
1937 foreach my $line (split(/\n/, $image->{manifest_text})) {
1938 my @tokens = split(/\s+/, $line);
1940 $streamname = shift(@tokens);
1941 foreach my $filedata (grep(/^\d+:\d+:/, @tokens)) {
1942 if (defined($filename)) {
1943 return (undef, undef); # More than one file in the Collection.
1945 $filename = (split(/:/, $filedata, 3))[2];
1950 if (defined($filename) and ($filename =~ /^([0-9A-Fa-f]{64})\.tar$/)) {
1951 return ($streamname, $1);
1953 return (undef, undef);
1958 # Calculate the number of times an operation should be retried,
1959 # assuming exponential backoff, and that we're willing to retry as
1960 # long as tasks have been running. Enforce a minimum of 3 retries.
1961 my ($starttime, $endtime, $timediff, $retries);
1963 $starttime = $jobstep[0]->{starttime};
1964 $endtime = $jobstep[-1]->{finishtime};
1966 if (!defined($starttime)) {
1968 } elsif (!defined($endtime)) {
1969 $timediff = time - $starttime;
1971 $timediff = ($endtime - $starttime) - (time - $endtime);
1973 if ($timediff > 0) {
1974 $retries = int(log($timediff) / log(2));
1976 $retries = 1; # Use the minimum.
1978 return ($retries > 3) ? $retries : 3;
1982 # Pass in two function references.
1983 # This method will be called with the remaining arguments.
1984 # If it dies, retry it with exponential backoff until it succeeds,
1985 # or until the current retry_count is exhausted. After each failure
1986 # that can be retried, the second function will be called with
1987 # the current try count (0-based), next try time, and error message.
1988 my $operation = shift;
1989 my $retry_callback = shift;
1990 my $retries = retry_count();
1991 foreach my $try_count (0..$retries) {
1992 my $next_try = time + (2 ** $try_count);
1993 my $result = eval { $operation->(@_); };
1996 } elsif ($try_count < $retries) {
1997 $retry_callback->($try_count, $next_try, $@);
1998 my $sleep_time = $next_try - time;
1999 sleep($sleep_time) if ($sleep_time > 0);
2002 # Ensure the error message ends in a newline, so Perl doesn't add
2003 # retry_op's line number to it.
2009 # Pass in a /-separated API method name, and arguments for it.
2010 # This function will call that method, retrying as needed until
2011 # the current retry_count is exhausted, with a log on the first failure.
2012 my $method_name = shift;
2013 my $log_api_retry = sub {
2014 my ($try_count, $next_try_at, $errmsg) = @_;
2015 $errmsg =~ s/\s*\bat \Q$0\E line \d+\.?\s*//;
2016 $errmsg =~ s/\s/ /g;
2017 $errmsg =~ s/\s+$//;
2019 if ($next_try_at < time) {
2020 $retry_msg = "Retrying.";
2022 my $next_try_fmt = strftime "%Y-%m-%dT%H:%M:%SZ", gmtime($next_try_at);
2023 $retry_msg = "Retrying at $next_try_fmt.";
2025 Log(undef, "API method $method_name failed: $errmsg. $retry_msg");
2028 foreach my $key (split(/\//, $method_name)) {
2029 $method = $method->{$key};
2031 return retry_op(sub { $method->execute(@_); }, $log_api_retry, @_);
2035 # Given a $?, return a human-readable exit code string like "0" or
2036 # "1" or "0 with signal 1" or "1 with signal 11".
2037 my $exitcode = shift;
2038 my $s = $exitcode >> 8;
2039 if ($exitcode & 0x7f) {
2040 $s .= " with signal " . ($exitcode & 0x7f);
2042 if ($exitcode & 0x80) {
2043 $s .= " with core dump";
2048 sub handle_readall {
2049 # Pass in a glob reference to a file handle.
2050 # Read all its contents and return them as a string.
2051 my $fh_glob_ref = shift;
2053 return <$fh_glob_ref>;
2056 sub tar_filename_n {
2058 return sprintf("%s/git.%s.%d.tar", $ENV{CRUNCH_TMP}, $job_id, $n);
2061 sub add_git_archive {
2062 # Pass in a git archive command as a string or list, a la system().
2063 # This method will save its output to be included in the archive sent to the
2067 if (!open(GIT_ARCHIVE, ">", tar_filename_n($git_tar_count))) {
2068 croak("Failed to save git archive: $!");
2070 my $git_pid = open2(">&GIT_ARCHIVE", $git_input, @_);
2072 waitpid($git_pid, 0);
2075 croak("Failed to save git archive: git exited " . exit_status_s($?));
2079 sub combined_git_archive {
2080 # Combine all saved tar archives into a single archive, then return its
2081 # contents in a string. Return undef if no archives have been saved.
2082 if ($git_tar_count < 1) {
2085 my $base_tar_name = tar_filename_n(1);
2086 foreach my $tar_to_append (map { tar_filename_n($_); } (2..$git_tar_count)) {
2087 my $tar_exit = system("tar", "-Af", $base_tar_name, $tar_to_append);
2088 if ($tar_exit != 0) {
2089 croak("Error preparing build archive: tar -A exited " .
2090 exit_status_s($tar_exit));
2093 if (!open(GIT_TAR, "<", $base_tar_name)) {
2094 croak("Could not open build archive: $!");
2096 my $tar_contents = handle_readall(\*GIT_TAR);
2098 return $tar_contents;
2101 sub set_nonblocking {
2103 my $flags = fcntl ($fh, F_GETFL, 0) or croak ($!);
2104 fcntl ($fh, F_SETFL, $flags | O_NONBLOCK) or croak ($!);
2110 # This is crunch-job's internal dispatch script. crunch-job running on the API
2111 # server invokes this script on individual compute nodes, or localhost if we're
2112 # running a job locally. It gets called in two modes:
2114 # * No arguments: Installation mode. Read a tar archive from the DATA
2115 # file handle; it includes the Crunch script's source code, and
2116 # maybe SDKs as well. Those should be installed in the proper
2117 # locations. This runs outside of any Docker container, so don't try to
2118 # introspect Crunch's runtime environment.
2120 # * With arguments: Crunch script run mode. This script should set up the
2121 # environment, then run the command specified in the arguments. This runs
2122 # inside any Docker container.
2125 use File::Path qw( make_path remove_tree );
2126 use POSIX qw(getcwd);
2128 use constant TASK_TEMPFAIL => 111;
2130 # Map SDK subdirectories to the path environments they belong to.
2131 my %SDK_ENVVARS = ("perl/lib" => "PERLLIB", "ruby/lib" => "RUBYLIB");
2133 my $destdir = $ENV{"CRUNCH_SRC"};
2134 my $archive_hash = $ENV{"CRUNCH_GIT_ARCHIVE_HASH"};
2135 my $repo = $ENV{"CRUNCH_SRC_URL"};
2136 my $install_dir = $ENV{"CRUNCH_INSTALL"} || (getcwd() . "/opt");
2137 my $job_work = $ENV{"JOB_WORK"};
2138 my $task_work = $ENV{"TASK_WORK"};
2140 open(STDOUT_ORIG, ">&", STDOUT);
2141 open(STDERR_ORIG, ">&", STDERR);
2143 for my $dir ($destdir, $job_work, $task_work) {
2146 -e $dir or die "Failed to create temporary directory ($dir): $!";
2151 remove_tree($task_work, {keep_root => 1});
2154 ### Crunch script run mode
2156 # We want to do routine logging during task 0 only. This gives the user
2157 # the information they need, but avoids repeating the information for every
2160 if ($ENV{TASK_SEQUENCE} eq "0") {
2163 printf STDERR_ORIG "[Crunch] $msg\n", @_;
2169 my $python_src = "$install_dir/python";
2170 my $venv_dir = "$job_work/.arvados.venv";
2171 my $venv_built = -e "$venv_dir/bin/activate";
2172 if ((!$venv_built) and (-d $python_src) and can_run("virtualenv")) {
2173 shell_or_die(undef, "virtualenv", "--quiet", "--system-site-packages",
2174 "--python=python2.7", $venv_dir);
2175 shell_or_die(TASK_TEMPFAIL, "$venv_dir/bin/pip", "--quiet", "install", "-I", $python_src);
2177 $Log->("Built Python SDK virtualenv");
2180 my @pysdk_version_cmd = ("python", "-c",
2181 "from pkg_resources import get_distribution as get; print get('arvados-python-client').version");
2183 $Log->("Running in Python SDK virtualenv");
2184 @pysdk_version_cmd = ();
2185 my $orig_argv = join(" ", map { quotemeta($_); } @ARGV);
2186 @ARGV = ("/bin/sh", "-ec",
2187 ". \Q$venv_dir/bin/activate\E; exec $orig_argv");
2188 } elsif (-d $python_src) {
2189 $Log->("Warning: virtualenv not found inside Docker container default " .
2190 "\$PATH. Can't install Python SDK.");
2193 if (@pysdk_version_cmd) {
2194 open(my $pysdk_version_pipe, "-|", @pysdk_version_cmd);
2195 my $pysdk_version = <$pysdk_version_pipe>;
2196 close($pysdk_version_pipe);
2198 chomp($pysdk_version);
2199 $Log->("Using Arvados SDK version $pysdk_version");
2201 # A lot could've gone wrong here, but pretty much all of it means that
2202 # Python won't be able to load the Arvados SDK.
2203 $Log->("Warning: Arvados SDK not found");
2207 while (my ($sdk_dir, $sdk_envkey) = each(%SDK_ENVVARS)) {
2208 my $sdk_path = "$install_dir/$sdk_dir";
2210 if ($ENV{$sdk_envkey}) {
2211 $ENV{$sdk_envkey} = "$sdk_path:" . $ENV{$sdk_envkey};
2213 $ENV{$sdk_envkey} = $sdk_path;
2215 $Log->("Arvados SDK added to %s", $sdk_envkey);
2220 die "Cannot exec `@ARGV`: $!";
2223 ### Installation mode
2224 open L, ">", "$destdir.lock" or die "$destdir.lock: $!";
2226 if (readlink ("$destdir.archive_hash") eq $archive_hash && -d $destdir) {
2227 # This exact git archive (source + arvados sdk) is already installed
2228 # here, so there's no need to reinstall it.
2230 # We must consume our DATA section, though: otherwise the process
2231 # feeding it to us will get SIGPIPE.
2233 while (read(DATA, $buf, 65536)) { }
2238 unlink "$destdir.archive_hash";
2242 # Ignore SIGPIPE: we check retval of close() instead. See perlipc(1).
2243 local $SIG{PIPE} = "IGNORE";
2244 warn "Extracting archive: $archive_hash\n";
2245 # --ignore-zeros is necessary sometimes: depending on how much NUL
2246 # padding tar -A put on our combined archive (which in turn depends
2247 # on the length of the component archives) tar without
2248 # --ignore-zeros will exit before consuming stdin and cause close()
2249 # to fail on the resulting SIGPIPE.
2250 if (!open(TARX, "|-", "tar", "--ignore-zeros", "-xC", $destdir)) {
2251 die "Error launching 'tar -xC $destdir': $!";
2253 # If we send too much data to tar in one write (> 4-5 MiB), it stops, and we
2254 # get SIGPIPE. We must feed it data incrementally.
2256 while (read(DATA, $tar_input, 65536)) {
2257 print TARX $tar_input;
2260 die "'tar -xC $destdir' exited $?: $!";
2266 my $sdk_root = "$destdir/.arvados.sdk/sdk";
2268 foreach my $sdk_lang (("python",
2269 map { (split /\//, $_, 2)[0]; } keys(%SDK_ENVVARS))) {
2270 if (-d "$sdk_root/$sdk_lang") {
2271 if (!rename("$sdk_root/$sdk_lang", "$install_dir/$sdk_lang")) {
2272 die "Failed to install $sdk_lang SDK: $!";
2278 my $python_dir = "$install_dir/python";
2279 if ((-d $python_dir) and can_run("python2.7")) {
2280 open(my $egg_info_pipe, "-|",
2281 "python2.7 \Q$python_dir/setup.py\E egg_info 2>&1 >/dev/null");
2282 my @egg_info_errors = <$egg_info_pipe>;
2283 close($egg_info_pipe);
2286 if (@egg_info_errors and (($egg_info_errors[-1] =~ /\bgit\b/) or ($egg_info_errors[-1] =~ /\[Errno 2\]/))) {
2287 # egg_info apparently failed because it couldn't ask git for a build tag.
2288 # Specify no build tag.
2289 open(my $pysdk_cfg, ">>", "$python_dir/setup.cfg");
2290 print $pysdk_cfg "\n[egg_info]\ntag_build =\n";
2293 my $egg_info_exit = $? >> 8;
2294 foreach my $errline (@egg_info_errors) {
2297 warn "python setup.py egg_info failed: exit $egg_info_exit";
2298 exit ($egg_info_exit || 1);
2303 # Hide messages from the install script (unless it fails: shell_or_die
2304 # will show $destdir.log in that case).
2305 open(STDOUT, ">>", "$destdir.log");
2306 open(STDERR, ">&", STDOUT);
2308 if (-e "$destdir/crunch_scripts/install") {
2309 shell_or_die (undef, "$destdir/crunch_scripts/install", $install_dir);
2310 } elsif (!-e "./install.sh" && -e "./tests/autotests.sh") {
2312 shell_or_die (undef, "./tests/autotests.sh", $install_dir);
2313 } elsif (-e "./install.sh") {
2314 shell_or_die (undef, "./install.sh", $install_dir);
2317 if ($archive_hash) {
2318 unlink "$destdir.archive_hash.new";
2319 symlink ($archive_hash, "$destdir.archive_hash.new") or die "$destdir.archive_hash.new: $!";
2320 rename ("$destdir.archive_hash.new", "$destdir.archive_hash") or die "$destdir.archive_hash: $!";
2326 my $command_name = shift;
2327 open(my $which, "-|", "which", $command_name);
2328 while (<$which>) { }
2335 my $exitcode = shift;
2337 if ($ENV{"DEBUG"}) {
2338 print STDERR "@_\n";
2340 if (system (@_) != 0) {
2343 my $exitstatus = sprintf("exit %d signal %d", $code >> 8, $code & 0x7f);
2344 open STDERR, ">&STDERR_ORIG";
2345 system ("cat $destdir.log >&2");
2346 warn "@_ failed ($err): $exitstatus";
2347 if (defined($exitcode)) {
2351 exit (($code >> 8) || 1);