-# Copyright (C) 2016-2019 all contributors <meta@public-inbox.org>
+# Copyright (C) 2016-2021 all contributors <meta@public-inbox.org>
# License: AGPL-3.0+ <https://www.gnu.org/licenses/agpl-3.0.txt>
#
# This allows vfork to be used for spawning subprocesses if
-# PERL_INLINE_DIRECTORY is explicitly defined in the environment.
+# ~/.cache/public-inbox/inline-c is writable or if PERL_INLINE_DIRECTORY
+# is explicitly defined in the environment (and writable).
# Under Linux, vfork can make a big difference in spawning performance
# as process size increases (fork still needs to mark pages for CoW use).
# Currently, we only use this for code intended for long running
# daemons (inside the PSGI code (-httpd) and -nntpd). The short-lived
# scripts (-mda, -index, -learn, -init) either use IPC::run or standard
# Perl routines.
+#
+# There'll probably be more OS-level C stuff here, down the line.
+# We don't want too many DSOs: https://udrepper.livejournal.com/8790.html
package PublicInbox::Spawn;
use strict;
-use warnings;
-use base qw(Exporter);
+use parent qw(Exporter);
use Symbol qw(gensym);
use PublicInbox::ProcessPipe;
-our @EXPORT_OK = qw/which spawn popen_rd/;
-sub RLIMITS () { qw(RLIMIT_CPU RLIMIT_CORE RLIMIT_DATA) }
+our @EXPORT_OK = qw(which spawn popen_rd run_die nodatacow_dir);
+our @RLIMITS = qw(RLIMIT_CPU RLIMIT_CORE RLIMIT_DATA);
my $vfork_spawn = <<'VFORK_SPAWN';
#include <sys/types.h>
-#include <sys/uio.h>
#include <sys/time.h>
#include <sys/resource.h>
#include <unistd.h>
#include <stdlib.h>
+#include <errno.h>
/* some platforms need alloca.h, but some don't */
#if defined(__GNUC__) && !defined(alloca)
const char *filename = SvPV_nolen(file);
pid_t pid;
char **argv, **envp;
- sigset_t set, old;
+ sigset_t set, old, cset;
int ret, perrnum, cerrnum = 0;
AV2C_COPY(argv, cmd);
assert(ret == 0 && "BUG calling sigfillset");
ret = sigprocmask(SIG_SETMASK, &set, &old);
assert(ret == 0 && "BUG calling sigprocmask to block");
+ ret = sigemptyset(&cset);
+ assert(ret == 0 && "BUG calling sigemptyset");
+ ret = sigaddset(&cset, SIGCHLD);
+ assert(ret == 0 && "BUG calling sigaddset for SIGCHLD");
pid = vfork();
if (pid == 0) {
int sig;
}
/*
- * don't bother unblocking, we don't want signals
- * to the group taking out a subprocess
+ * don't bother unblocking other signals for now, just SIGCHLD.
+ * we don't want signals to the group taking out a subprocess
*/
+ (void)sigprocmask(SIG_UNBLOCK, &cset, NULL);
execve(filename, argv, envp);
exit_err(&cerrnum);
}
}
VFORK_SPAWN
-my $inline_dir = $ENV{PERL_INLINE_DIRECTORY};
-$vfork_spawn = undef unless defined $inline_dir && -d $inline_dir && -w _;
+# btrfs on Linux is copy-on-write (COW) by default. As of Linux 5.7,
+# this still leads to fragmentation for SQLite and Xapian files where
+# random I/O happens, so we disable COW just for SQLite files and Xapian
+# directories. Disabling COW disables checksumming, so we only do this
+# for regeneratable files, and not canonical git storage (git doesn't
+# checksum refs, only data under $GIT_DIR/objects).
+my $set_nodatacow = $^O eq 'linux' ? <<'SET_NODATACOW' : '';
+#include <sys/ioctl.h>
+#include <sys/vfs.h>
+#include <linux/magic.h>
+#include <linux/fs.h>
+#include <dirent.h>
+#include <errno.h>
+#include <stdio.h>
+#include <string.h>
+
+void nodatacow_fd(int fd)
+{
+ struct statfs buf;
+ int val = 0;
+
+ if (fstatfs(fd, &buf) < 0) {
+ fprintf(stderr, "fstatfs: %s\\n", strerror(errno));
+ return;
+ }
+
+ /* only btrfs is known to have this problem, so skip for non-btrfs */
+ if (buf.f_type != BTRFS_SUPER_MAGIC)
+ return;
+
+ if (ioctl(fd, FS_IOC_GETFLAGS, &val) < 0) {
+ fprintf(stderr, "FS_IOC_GET_FLAGS: %s\\n", strerror(errno));
+ return;
+ }
+ val |= FS_NOCOW_FL;
+ if (ioctl(fd, FS_IOC_SETFLAGS, &val) < 0)
+ fprintf(stderr, "FS_IOC_SET_FLAGS: %s\\n", strerror(errno));
+}
+
+void nodatacow_dir(const char *dir)
+{
+ DIR *dh = opendir(dir);
+ int fd;
+
+ if (!dh) croak("opendir(%s): %s", dir, strerror(errno));
+ fd = dirfd(dh);
+ if (fd >= 0)
+ nodatacow_fd(fd);
+ /* ENOTSUP probably won't happen under Linux... */
+ closedir(dh);
+}
+SET_NODATACOW
+
+my $inline_dir = $ENV{PERL_INLINE_DIRECTORY} //= (
+ $ENV{XDG_CACHE_HOME} //
+ ( ($ENV{HOME} // '/nonexistent').'/.cache' )
+ ).'/public-inbox/inline-c';
+
+$set_nodatacow = $vfork_spawn = undef unless -d $inline_dir && -w _;
if (defined $vfork_spawn) {
# Inline 0.64 or later has locking in multi-process env,
# but we support 0.5 on Debian wheezy
my $f = "$inline_dir/.public-inbox.lock";
open my $fh, '>', $f or die "failed to open $f: $!\n";
flock($fh, LOCK_EX) or die "LOCK_EX failed on $f: $!\n";
- eval 'use Inline C => $vfork_spawn'; #, BUILD_NOISY => 1';
+ eval 'use Inline C => $vfork_spawn . $set_nodatacow';
my $err = $@;
+ my $ndc_err;
+ if ($err && $set_nodatacow) { # missing Linux kernel headers
+ $ndc_err = $err;
+ undef $set_nodatacow;
+ eval 'use Inline C => $vfork_spawn';
+ }
flock($fh, LOCK_UN) or die "LOCK_UN failed on $f: $!\n";
die $err if $err;
+ warn $ndc_err if $ndc_err;
};
if ($@) {
warn "Inline::C failed for vfork: $@\n";
- $vfork_spawn = undef;
+ $set_nodatacow = $vfork_spawn = undef;
}
}
require PublicInbox::SpawnPP;
*pi_fork_exec = \&PublicInbox::SpawnPP::pi_fork_exec
}
+unless ($set_nodatacow) {
+ require PublicInbox::NDC_PP;
+ no warnings 'once';
+ *nodatacow_fd = \&PublicInbox::NDC_PP::nodatacow_fd;
+ *nodatacow_dir = \&PublicInbox::NDC_PP::nodatacow_dir;
+}
+undef $set_nodatacow;
+undef $vfork_spawn;
sub which ($) {
my ($file) = @_;
}
my $rlim = [];
- foreach my $l (RLIMITS()) {
+ foreach my $l (@RLIMITS) {
defined(my $v = $opts->{$l}) or next;
my $r = eval "require BSD::Resource; BSD::Resource::$l();";
unless (defined $r) {
}
my $cd = $opts->{'-C'} // ''; # undef => NULL mapping doesn't work?
my $pid = pi_fork_exec($redir, $f, $cmd, \@env, $rlim, $cd);
- $pid < 0 ? undef : $pid;
+ die "fork_exec @$cmd failed: $!\n" unless $pid > 0;
+ $pid;
}
sub popen_rd {
$ret;
}
+sub run_die ($;$$) {
+ my ($cmd, $env, $rdr) = @_;
+ my $pid = spawn($cmd, $env, $rdr);
+ waitpid($pid, 0) == $pid or die "@$cmd did not finish";
+ $? == 0 or die "@$cmd failed: \$?=$?\n";
+}
+
1;