tvl-depot/scripts/build-remote.pl.in
Eelco Dolstra 019176137f * When using a build hook, distinguish between transient failures
(e.g. an SSH connection problem) and permanent failures (i.e. the
  builder failed).  This matters to Hydra (it wants to know whether it
  makes sense to retry a build).
2009-01-13 11:39:09 +00:00

225 lines
6.8 KiB
Perl
Executable file

#! @perl@ -w
use strict;
use Fcntl ':flock';
use English '-no_match_vars';
# General operation:
#
# Try to find a free machine of type $neededSystem. We do this as
# follows:
# - We acquire an exclusive lock on $currentLoad/main-lock.
# - For each machine $machine of type $neededSystem and for each $slot
# less than the maximum load for that machine, we try to get an
# exclusive lock on $currentLoad/$machine-$slot (without blocking).
# If we get such a lock, we send "accept" to the caller. Otherwise,
# we send "postpone" and exit.
# - We release the exclusive lock on $currentLoad/main-lock.
# - We perform the build on $neededSystem.
# - We release the exclusive lock on $currentLoad/$machine-$slot.
#
# The nice thing about this scheme is that if we die prematurely, the
# locks are released automatically.
my $loadIncreased = 0;
my ($amWilling, $localSystem, $neededSystem, $drvPath, $mustRun, $maxSilentTime) = @ARGV;
$mustRun = 0 unless defined $mustRun;
$maxSilentTime = 0 unless defined $maxSilentTime;
sub sendReply {
my $reply = shift;
open OUT, ">&3" or die;
print OUT "$reply\n";
close OUT;
}
sub decline {
sendReply "decline";
exit 0;
}
my $currentLoad = $ENV{"NIX_CURRENT_LOAD"};
decline unless defined $currentLoad;
mkdir $currentLoad, 0777 or die unless -d $currentLoad;
my $conf = $ENV{"NIX_REMOTE_SYSTEMS"};
decline if !defined $conf || ! -e $conf;
my $canBuildLocally = $amWilling && ($localSystem eq $neededSystem);
# Otherwise find a willing remote machine.
my @machines;
my %curJobs;
# Read the list of machines.
open CONF, "< $conf" or die;
while (<CONF>) {
chomp;
s/\#.*$//g;
next if /^\s*$/;
/^\s*(\S+)\s+(\S+)\s+(\S+)\s+(\d+)\s*$/ or die;
push @machines,
{ hostName => $1
, systemType => $2
, sshKeys => $3
, maxJobs => $4
};
}
close CONF;
# Acquire the exclusive lock on $currentLoad/main-lock.
my $mainLock = "$currentLoad/main-lock";
open MAINLOCK, ">>$mainLock" or die;
flock(MAINLOCK, LOCK_EX) or die;
# Find a suitable system.
my $rightType = 0;
my $machine;
LOOP: foreach my $cur (@machines) {
if ($neededSystem eq $cur->{systemType}) {
$rightType = 1;
# We have a machine of the right type. Try to get a lock on
# one of the machine's lock files.
my $slot = 0;
while ($slot < $cur->{maxJobs} || ($mustRun && !$canBuildLocally)) {
my $slotLock = "$currentLoad/" . $cur->{systemType} . "-" . $cur->{hostName} . "-$slot";
open SLOTLOCK, ">>$slotLock" or die;
if (flock(SLOTLOCK, LOCK_EX | LOCK_NB)) {
print STDERR "warning: exceeding maximum load on " . $cur->{systemType} . "\n"
if $slot >= $cur->{maxJobs};
$machine = $cur;
last LOOP;
}
close SLOTLOCK;
$slot++;
}
}
}
close MAINLOCK;
# Didn't find one? Then decline or postpone.
if (!defined $machine) {
# Postpone if we have a machine of the right type, except if the
# local system can and wants to do the build.
if ($rightType && !$canBuildLocally) {
sendReply "postpone";
exit 0;
} else {
decline;
}
}
# Yes we did, accept.
sendReply "accept";
open IN, "<&4" or die;
my $x = <IN>;
chomp $x;
#print "got $x\n";
close IN;
if ($x ne "okay") {
exit 0;
}
# Do the actual job.
my $hostName = $machine->{hostName};
print "BUILDING REMOTE: $drvPath on $hostName\n";
# Make sure that we don't get any SSH passphrase or host key popups -
# if there is any problem it should fail, not do something
# interactive.
$ENV{"DISPLAY"} = "";
$ENV{"SSH_PASSWORD_FILE="} = "";
$ENV{"SSH_ASKPASS="} = "";
my $sshOpts = "-i " . $machine->{sshKeys} . " -x";
# Hack to support Cygwin: if we login without a password, we don't
# have exactly the same rights as when we do. This causes the
# Microsoft C compiler to fail with certain flags:
#
# http://connect.microsoft.com/VisualStudio/feedback/ViewFeedback.aspx?FeedbackID=99676
#
# So as a workaround, we pass a verbatim password. ssh tries to makes
# this very hard; the trick is to make it call SSH_ASKPASS to get the
# password. (It only calls this command when there is no controlling
# terminal, but Nix ensures that is is the case. When doing this
# manually, use setsid(1).)
if ($machine->{sshKeys} =~ /^password:/) {
my $passwordFile = $machine->{sshKeys};
$passwordFile =~ s/^password://;
$sshOpts = "ssh -x";
$ENV{"SSH_PASSWORD_FILE"} = $passwordFile;
$ENV{"SSH_ASKPASS"} = "/tmp/writepass";
open WRITEPASS, ">/tmp/writepass" or die;
print WRITEPASS "#! /bin/sh\ncat \"\$SSH_PASSWORD_FILE\"";
close WRITEPASS;
chmod 0755, "/tmp/writepass" or die;
}
my $inputs = `cat inputs`; die if ($? != 0);
$inputs =~ s/\n/ /g;
my $outputs = `cat outputs`; die if ($? != 0);
$outputs =~ s/\n/ /g;
print "COPYING INPUTS...\n";
my $maybeSign = "";
$maybeSign = "--sign" if -e "/nix/etc/nix/signing-key.sec";
system("NIX_SSHOPTS=\"$sshOpts\" nix-copy-closure $hostName $maybeSign $drvPath $inputs") == 0
or die "cannot copy inputs to $hostName: $?";
print "BUILDING...\n";
my $buildFlags = "--max-silent-time $maxSilentTime";
# `-tt' forces allocation of a pseudo-terminal. This is required to
# make the remote nix-store process receive a signal when the
# connection dies. Without it, the remote process might continue to
# run indefinitely (that is, until it next tries to write to
# stdout/stderr).
if (system("ssh -tt $sshOpts $hostName 'nix-store -rvvK $buildFlags $drvPath'") != 0) {
# If we couldn't run ssh or there was an ssh problem (indicated by
# exit code 255), then we return exit code 1; otherwise we assume
# that the builder failed, which we indicated to Nix using exit
# code 100. It's important to distinguish between the two because
# the first is a transient failure and the latter is permanent.
my $res = $? == -1 || ($? >> 8) == 255 ? 1 : 100;
print STDERR "remote build on $hostName failed: $?";
exit $res;
}
print "REMOTE BUILD DONE: $drvPath on $hostName\n";
foreach my $output (split '\n', $outputs) {
my $maybeSignRemote = "";
$maybeSignRemote = "--sign" if $UID != 0;
system("ssh $sshOpts $hostName 'nix-store --export $maybeSignRemote $output' > dump") == 0
or die "cannot copy $output from $hostName: $?";
# This doesn't work yet, since the caller has a lock on the output
# path. We should move towards lock-free invocation of build
# hooks and substitutes.
#system("nix-store --import < dump") == 0
# or die "cannot import $output: $?";
# Hack: skip the first 8 bytes (the nix-store --export next
# archive marker). The archive follows.
system("(dd bs=1 count=8 of=/dev/null && cat) < dump | nix-store --restore $output") == 0
or die "cannot restore $output: $?";
}