1 # Copyright (C) 2016-2019 all contributors <meta@public-inbox.org>
2 # License: AGPL-3.0+ <https://www.gnu.org/licenses/agpl-3.0.txt>
4 # Like most Perl modules in public-inbox, this is internal and
5 # NOT subject to any stability guarantees! It is only documented
8 # This is used to limit the number of processes spawned by the
9 # PSGI server, so it acts like a semaphore and queues up extra
10 # commands to be run if currently at the limit. Multiple "limiters"
11 # may be configured which give inboxes different channels to
12 # operate in. This can be useful to ensure smaller inboxes can
13 # be cloned while cloning of large inboxes is maxed out.
15 # This does not depend on PublicInbox::DS or any other external
16 # scheduling mechanism, you just need to call start() and finish()
17 # appropriately. However, public-inbox-httpd (which uses PublicInbox::DS)
18 # will be able to schedule this based on readability of stdout from
19 # the spawned process. See GitHTTPBackend.pm and SolverGit.pm for
20 # usage examples. It does not depend on any form of threading.
22 # This is useful for scheduling CGI execution of both long-lived
23 # git-http-backend(1) process (for "git clone") as well as short-lived
24 # processes such as git-apply(1).
26 package PublicInbox::Qspawn;
29 use PublicInbox::Spawn qw(popen_rd);
32 # n.b.: we get EAGAIN with public-inbox-httpd, and EINTR on other PSGI servers
33 use Errno qw(EAGAIN EINTR);
37 # declares a command to spawn (but does not spawn it).
38 # $cmd is the command to spawn
39 # $env is the environ for the child process
40 # $opt can include redirects and perhaps other process spawning options
42 my ($class, $cmd, $env, $opt) = @_;
43 bless { args => [ $cmd, $env, $opt ] }, $class;
47 my ($self, $start_cb, $limiter) = @_;
49 my ($cmd, $env, $opts) = @{$self->{args}};
50 my %opts = %{$opts || {}};
51 $self->{limiter} = $limiter;
52 foreach my $k (PublicInbox::Spawn::RLIMITS()) {
53 if (defined(my $rlimit = $limiter->{$k})) {
58 ($self->{rpipe}, $self->{pid}) = popen_rd($cmd, $env, \%opts);
60 # drop any IO handles opt was holding open via $opt->{hold}
61 # No need to hold onto the descriptor once the child process has it.
62 $self->{args} = $cmd; # keep this around for logging
64 if (defined $self->{pid}) {
65 $limiter->{running}++;
69 $start_cb->($self->{rpipe});
73 my ($child_error) = @_; # typically $?
74 my $exitstatus = ($child_error >> 8) or return;
75 my $sig = $child_error & 127;
76 my $msg = "exit status=$exitstatus";
77 $msg .= " signal=$sig" if $sig;
83 $env->{'psgi.errors'}->print($msg, "\n");
86 # callback for dwaitpid
87 sub waitpid_err ($$) {
88 my ($self, $pid) = @_;
89 my $xpid = delete $self->{pid};
91 if ($pid > 0) { # success!
93 } elsif ($pid < 0) { # ??? does this happen in our case?
94 $err = "W: waitpid($xpid, 0) => $pid: $!";
95 } # else should not be called with pid == 0
97 my $env = delete $self->{env};
99 # done, spawn whatever's in the queue
100 my $limiter = $self->{limiter};
101 my $running = --$limiter->{running};
103 if ($running < $limiter->{max}) {
104 if (my $next = shift(@{$limiter->{run_queue}})) {
105 _do_spawn(@$next, $limiter);
111 if ($env && !$env->{'qspawn.quiet'}) {
112 log_err($env, join(' ', @{$self->{args}}) . ": $err");
115 if (my $fin_cb = delete $self->{fin_cb}) {
120 sub do_waitpid ($;$$) {
121 my ($self, $env, $fin_cb) = @_;
122 my $pid = $self->{pid};
124 $self->{fin_cb} = $fin_cb;
125 # PublicInbox::DS may not be loaded
126 eval { PublicInbox::DS::dwaitpid($pid, \&waitpid_err, $self) };
127 # done if we're running in PublicInbox::DS::EventLoop
129 # non public-inbox-{httpd,nntpd} callers may block:
130 my $ret = waitpid($pid, 0);
131 waitpid_err($self, $ret);
136 my ($self, $env, $fin_cb) = @_;
137 if (delete $self->{rpipe}) {
138 do_waitpid($self, $env, $fin_cb);
140 eval { $fin_cb->() };
145 my ($self, $limiter, $start_cb) = @_;
146 if ($limiter->{running} < $limiter->{max}) {
147 _do_spawn($self, $start_cb, $limiter);
149 push @{$limiter->{run_queue}}, [ $self, $start_cb ];
153 # Similar to `backtick` or "qx" ("perldoc -f qx"), it calls $qx_cb with
154 # the stdout of the given command when done; but respects the given limiter
155 # $env is the PSGI env. As with ``/qx; only use this when output is small
158 my ($self, $env, $limiter, $qx_cb) = @_;
160 open(my $qx, '+>', \$scalar) or die; # PerlIO::scalar
162 my $err = $_[0]; # $!
163 log_err($env, "psgi_qx: $err") if defined($err);
164 finish($self, $env, sub { $qx_cb->(\$scalar) });
167 my $rpipe; # comes from popen_rd
168 my $async = $env->{'pi-httpd.async'};
172 $r = sysread($rpipe, $buf, 65536);
174 $async->async_pass($env->{'psgix.io'}, $qx, \$buf);
175 } elsif (defined $r) {
176 $r ? $qx->write($buf) : $end->();
178 return if $! == EAGAIN; # try again when notified
179 goto reread if $! == EINTR;
183 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
184 $self->start($limiter, sub { # start_cb, may run later, much later...
185 ($rpipe) = @_; # popen_rd result
187 # PublicInbox::HTTPD::Async->new($rpipe, $cb, $end)
188 $async = $async->($rpipe, $cb, $end);
189 # $cb will call ->async_pass or ->close
190 } else { # generic PSGI
196 # create a filter for "push"-based streaming PSGI writes used by HTTPD::Async
198 my ($fh, $filter) = @_;
199 Plack::Util::inline_object(
201 $fh->write($filter->(undef));
205 $fh->write($filter->($_[0]));
209 # Used for streaming the stdout of one process as a PSGI response.
211 # $env is the PSGI env.
212 # optional keys in $env:
213 # $env->{'qspawn.wcb'} - the write callback from the PSGI server
214 # optional, use this if you've already
215 # captured it elsewhere. If not given,
216 # psgi_return will return an anonymous
217 # sub for the PSGI server to call
219 # $env->{'qspawn.filter'} - filter callback, receives a string as input,
222 # $limiter - the Limiter object to use (uses the def_limiter if not given)
224 # $parse_hdr - Initial read function; often for parsing CGI header output.
225 # It will be given the return value of sysread from the pipe
226 # and a string ref of the current buffer. Returns an arrayref
227 # for PSGI responses. 2-element arrays in PSGI mean the
228 # body will be streamed, later, via writes (push-based) to
229 # psgix.io. 3-element arrays means the body is available
230 # immediately (or streamed via ->getline (pull-based)).
232 my ($self, $env, $limiter, $parse_hdr) = @_;
235 my $err = $_[0]; # $!
236 log_err($env, "psgi_return: $err") if defined($err);
238 $fh->close if $fh; # async-only
243 # typically used for reading CGI headers
244 # we must loop until EAGAIN for EPOLLET in HTTPD/Async.pm
245 # We also need to check EINTR for generic PSGI servers.
249 my $r = sysread($rpipe, $buf, 4096, length($buf));
252 $ret = $parse_hdr->($r ? $total_rd : 0, \$buf);
254 # caller should notify us when it's ready:
255 return if $! == EAGAIN;
256 next if $! == EINTR; # immediate retry
257 log_err($env, "error reading header: $!");
258 $ret = [ 500, [], [ "Internal error\n" ] ];
260 } until (defined $ret);
264 my $wcb = delete $env->{'qspawn.wcb'}; # or PSGI server supplies it
265 my $async = $env->{'pi-httpd.async'};
268 my $r = $rd_hdr->() or return;
269 $rd_hdr = undef; # done reading headers
270 my $filter = delete $env->{'qspawn.filter'};
271 if (scalar(@$r) == 3) { # error
273 $async->close; # calls rpipe->close and $end
280 # done reading headers, handoff to read body
281 $fh = $wcb->($r); # scalar @$r == 2
282 $fh = filter_fh($fh, $filter) if $filter;
283 $async->async_pass($env->{'psgix.io'}, $fh, \$buf);
284 } else { # for synchronous PSGI servers
285 require PublicInbox::GetlineBody;
286 $r->[2] = PublicInbox::GetlineBody->new($rpipe, $end,
291 # Workaround a leak under Perl 5.16.3 when combined with
292 # Plack::Middleware::Deflater:
295 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
296 my $start_cb = sub { # may run later, much later...
299 # PublicInbox::HTTPD::Async->new($rpipe, $cb, $end)
300 $async = $async->($rpipe, $cb, $end);
301 # $cb will call ->async_pass or ->close
302 } else { # generic PSGI
303 $cb->() while $rd_hdr;
307 # the caller already captured the PSGI write callback from
308 # the PSGI server, so we can call ->start, here:
309 return $self->start($limiter, $start_cb) if $wcb;
311 # the caller will return this sub to the PSGI server, so
312 # it can set the response callback (that is, for PublicInbox::HTTP,
313 # the chunked_wcb or identity_wcb callback), but other HTTP servers
317 $self->start($limiter, $start_cb);
321 package PublicInbox::Qspawn::Limiter;
326 my ($class, $max) = @_;
328 # 32 is same as the git-daemon connection limit
332 # RLIMIT_CPU => undef,
333 # RLIMIT_DATA => undef,
334 # RLIMIT_CORE => undef,
339 my ($self, $name, $config) = @_;
340 foreach my $rlim (PublicInbox::Spawn::RLIMITS()) {
343 $k = "publicinboxlimiter.$name.$k";
344 defined(my $v = $config->{$k}) or next;
345 my @rlimit = split(/\s*,\s*/, $v);
346 if (scalar(@rlimit) == 1) {
347 push @rlimit, $rlimit[0];
348 } elsif (scalar(@rlimit) != 2) {
349 warn "could not parse $k: $v\n";
351 eval { require BSD::Resource };
353 warn "BSD::Resource missing for $rlim";
356 foreach my $i (0..$#rlimit) {
357 next if $rlimit[$i] ne 'INFINITY';
358 $rlimit[$i] = BSD::Resource::RLIM_INFINITY();
360 $self->{$rlim} = \@rlimit;