1 # Copyright (C) 2016-2020 all contributors <meta@public-inbox.org>
2 # License: AGPL-3.0+ <https://www.gnu.org/licenses/agpl-3.0.txt>
4 # This allows vfork to be used for spawning subprocesses if
5 # ~/.cache/public-inbox/inline-c is writable or if PERL_INLINE_DIRECTORY
6 # is explicitly defined in the environment (and writable).
7 # Under Linux, vfork can make a big difference in spawning performance
8 # as process size increases (fork still needs to mark pages for CoW use).
9 # Currently, we only use this for code intended for long running
10 # daemons (inside the PSGI code (-httpd) and -nntpd). The short-lived
11 # scripts (-mda, -index, -learn, -init) either use IPC::run or standard
14 package PublicInbox::Spawn;
16 use parent qw(Exporter);
17 use Symbol qw(gensym);
18 use PublicInbox::ProcessPipe;
19 our @EXPORT_OK = qw/which spawn popen_rd/;
20 our @RLIMITS = qw(RLIMIT_CPU RLIMIT_CORE RLIMIT_DATA);
22 my $vfork_spawn = <<'VFORK_SPAWN';
23 #include <sys/types.h>
25 #include <sys/resource.h>
29 /* some platforms need alloca.h, but some don't */
30 #if defined(__GNUC__) && !defined(alloca)
31 # define alloca(sz) __builtin_alloca(sz)
38 * From the av_len apidoc:
39 * Note that, unlike what the name implies, it returns
40 * the highest index in the array, so to get the size of
41 * the array you need to use "av_len(av) + 1".
42 * This is unlike "sv_len", which returns what you would expect.
44 #define AV2C_COPY(dst, src) do { \
46 I32 top_index = av_len(src); \
47 I32 real_len = top_index + 1; \
48 I32 capa = real_len + 1; \
49 dst = alloca(capa * sizeof(char *)); \
50 for (i = 0; i < real_len; i++) { \
51 SV **sv = av_fetch(src, i, 0); \
52 dst[i] = SvPV_nolen(*sv); \
57 /* needs to be safe inside a vfork'ed process */
58 static void exit_err(int *cerrnum)
65 * unstable internal API. It'll be updated depending on
66 * whatever we'll need in the future.
67 * Be sure to update PublicInbox::SpawnPP if this changes
69 int pi_fork_exec(SV *redirref, SV *file, SV *cmdref, SV *envref, SV *rlimref,
72 AV *redir = (AV *)SvRV(redirref);
73 AV *cmd = (AV *)SvRV(cmdref);
74 AV *env = (AV *)SvRV(envref);
75 AV *rlim = (AV *)SvRV(rlimref);
76 const char *filename = SvPV_nolen(file);
79 sigset_t set, old, cset;
80 int ret, perrnum, cerrnum = 0;
85 ret = sigfillset(&set);
86 assert(ret == 0 && "BUG calling sigfillset");
87 ret = sigprocmask(SIG_SETMASK, &set, &old);
88 assert(ret == 0 && "BUG calling sigprocmask to block");
89 ret = sigemptyset(&cset);
90 assert(ret == 0 && "BUG calling sigemptyset");
91 ret = sigaddset(&cset, SIGCHLD);
92 assert(ret == 0 && "BUG calling sigaddset for SIGCHLD");
96 I32 i, child_fd, max = av_len(redir);
98 for (child_fd = 0; child_fd <= max; child_fd++) {
99 SV **parent = av_fetch(redir, child_fd, 0);
100 int parent_fd = SvIV(*parent);
101 if (parent_fd == child_fd)
103 if (dup2(parent_fd, child_fd) < 0)
106 for (sig = 1; sig < NSIG; sig++)
107 signal(sig, SIG_DFL); /* ignore errors on signals */
108 if (*cd && chdir(cd) < 0)
112 for (i = 0; i < max; i += 3) {
114 SV **res = av_fetch(rlim, i, 0);
115 SV **soft = av_fetch(rlim, i + 1, 0);
116 SV **hard = av_fetch(rlim, i + 2, 0);
118 rl.rlim_cur = SvIV(*soft);
119 rl.rlim_max = SvIV(*hard);
120 if (setrlimit(SvIV(*res), &rl) < 0)
125 * don't bother unblocking other signals for now, just SIGCHLD.
126 * we don't want signals to the group taking out a subprocess
128 (void)sigprocmask(SIG_UNBLOCK, &cset, NULL);
129 execve(filename, argv, envp);
133 ret = sigprocmask(SIG_SETMASK, &old, NULL);
134 assert(ret == 0 && "BUG calling sigprocmask to restore");
137 waitpid(pid, NULL, 0);
140 } else if (perrnum) {
147 my $inline_dir = $ENV{PERL_INLINE_DIRECTORY} //= (
148 $ENV{XDG_CACHE_HOME} //
149 ( ($ENV{HOME} // '/nonexistent').'/.cache' )
150 ).'/public-inbox/inline-c';
152 $vfork_spawn = undef unless -d $inline_dir && -w _;
153 if (defined $vfork_spawn) {
154 # Inline 0.64 or later has locking in multi-process env,
155 # but we support 0.5 on Debian wheezy
156 use Fcntl qw(:flock);
158 my $f = "$inline_dir/.public-inbox.lock";
159 open my $fh, '>', $f or die "failed to open $f: $!\n";
160 flock($fh, LOCK_EX) or die "LOCK_EX failed on $f: $!\n";
161 eval 'use Inline C => $vfork_spawn';
163 flock($fh, LOCK_UN) or die "LOCK_UN failed on $f: $!\n";
167 warn "Inline::C failed for vfork: $@\n";
168 $vfork_spawn = undef;
172 unless (defined $vfork_spawn) {
173 require PublicInbox::SpawnPP;
174 *pi_fork_exec = \&PublicInbox::SpawnPP::pi_fork_exec
179 return $file if index($file, '/') >= 0;
180 foreach my $p (split(':', $ENV{PATH})) {
188 my ($cmd, $env, $opts) = @_;
189 my $f = which($cmd->[0]);
190 defined $f or die "$cmd->[0]: command not found\n";
194 my %env = $env ? (%ENV, %$env) : %ENV;
195 while (my ($k, $v) = each %env) {
199 for my $child_fd (0..2) {
200 my $parent_fd = $opts->{$child_fd};
201 if (defined($parent_fd) && $parent_fd !~ /\A[0-9]+\z/) {
202 defined(my $fd = fileno($parent_fd)) or
203 die "$parent_fd not an IO GLOB? $!";
206 $redir->[$child_fd] = $parent_fd // $child_fd;
210 foreach my $l (@RLIMITS) {
211 defined(my $v = $opts->{$l}) or next;
212 my $r = eval "require BSD::Resource; BSD::Resource::$l();";
213 unless (defined $r) {
214 warn "$l undefined by BSD::Resource: $@\n";
217 push @$rlim, $r, @$v;
219 my $cd = $opts->{'-C'} // ''; # undef => NULL mapping doesn't work?
220 my $pid = pi_fork_exec($redir, $f, $cmd, \@env, $rlim, $cd);
221 die "fork_exec failed: $!\n" unless $pid > 0;
226 my ($cmd, $env, $opts) = @_;
227 pipe(my ($r, $w)) or die "pipe: $!\n";
229 $opts->{1} = fileno($w);
230 my $pid = spawn($cmd, $env, $opts);
231 return ($r, $pid) if wantarray;
233 tie *$ret, 'PublicInbox::ProcessPipe', $pid, $r;