1 # Copyright (C) 2016-2021 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 the PublicInbox::DS::event_loop or any
16 # other external scheduling mechanism, you just need to call
17 # start() and finish() appropriately. However, public-inbox-httpd
18 # (which uses PublicInbox::DS) will be able to schedule this
19 # based on readability of stdout from the spawned process.
20 # See GitHTTPBackend.pm and SolverGit.pm for usage examples.
21 # It does not depend on any form of threading.
23 # This is useful for scheduling CGI execution of both long-lived
24 # git-http-backend(1) process (for "git clone") as well as short-lived
25 # processes such as git-apply(1).
27 package PublicInbox::Qspawn;
29 use PublicInbox::Spawn qw(popen_rd);
30 use PublicInbox::GzipFilter;
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 # $cmd_env is the environ for the child process (not PSGI env)
40 # $opt can include redirects and perhaps other process spawning options
42 my ($class, $cmd, $cmd_env, $opt) = @_;
43 bless { args => [ $cmd, $cmd_env, $opt ] }, $class;
47 my ($self, $start_cb, $limiter) = @_;
49 my ($cmd, $cmd_env, $opt) = @{delete $self->{args}};
50 my %o = %{$opt || {}};
51 $self->{limiter} = $limiter;
52 foreach my $k (@PublicInbox::Spawn::RLIMITS) {
53 if (defined(my $rlimit = $limiter->{$k})) {
57 $self->{cmd} = $o{quiet} ? undef : $cmd;
59 # popen_rd may die on EMFILE, ENFILE
60 $self->{rpipe} = popen_rd($cmd, $cmd_env, \%o);
62 die "E: $!" unless defined($self->{rpipe});
64 $limiter->{running}++;
65 $start_cb->($self); # EPOLL_CTL_ADD may ENOSPC/ENOMEM
67 finish($self, $@) if $@;
71 my ($child_error) = @_; # typically $?
72 my $exitstatus = ($child_error >> 8) or return;
73 my $sig = $child_error & 127;
74 my $msg = "exit status=$exitstatus";
75 $msg .= " signal=$sig" if $sig;
80 my ($self, $err) = @_;
82 my ($env, $qx_cb, $qx_arg, $qx_buf) =
83 delete @$self{qw(psgi_env qx_cb qx_arg qx_buf)};
85 # done, spawn whatever's in the queue
86 my $limiter = $self->{limiter};
87 my $running = --$limiter->{running};
89 if ($running < $limiter->{max}) {
90 if (my $next = shift(@{$limiter->{run_queue}})) {
91 _do_spawn(@$next, $limiter);
96 if (defined $self->{err}) {
97 $self->{err} .= "; $err";
101 if ($env && $self->{cmd}) {
102 warn join(' ', @{$self->{cmd}}) . ": $err";
106 eval { $qx_cb->($qx_buf, $qx_arg) };
107 } elsif (my $wcb = delete $env->{'qspawn.wcb'}) {
108 # have we started writing, yet?
109 require PublicInbox::WwwStatic;
110 $wcb->(PublicInbox::WwwStatic::r(500));
114 # callback for dwaitpid or ProcessPipe
115 sub waitpid_err { finalize($_[0], child_err($?)) }
118 my ($self, $err) = @_;
119 my $tied_pp = delete($self->{rpipe}) or return finalize($self, $err);
120 my PublicInbox::ProcessPipe $pp = tied *$tied_pp;
121 @$pp{qw(cb arg)} = (\&waitpid_err, $self); # for ->DESTROY
125 my ($self, $limiter, $start_cb) = @_;
126 if ($limiter->{running} < $limiter->{max}) {
127 _do_spawn($self, $start_cb, $limiter);
129 push @{$limiter->{run_queue}}, [ $self, $start_cb ];
133 sub psgi_qx_init_cb {
135 my $async = delete $self->{async};
137 my $qx_fh = $self->{qx_fh};
139 $r = sysread($self->{rpipe}, $buf, 65536);
141 $async->async_pass($self->{psgi_env}->{'psgix.io'},
143 } elsif (defined $r) {
144 $r ? (print $qx_fh $buf) : event_step($self, undef);
146 return if $! == EAGAIN; # try again when notified
147 goto reread if $! == EINTR;
148 event_step($self, $!);
154 if (my $async = $self->{psgi_env}->{'pi-httpd.async'}) {
155 # PublicInbox::HTTPD::Async->new(rpipe, $cb, cb_arg, $end_obj)
156 $self->{async} = $async->($self->{rpipe},
157 \&psgi_qx_init_cb, $self, $self);
158 # init_cb will call ->async_pass or ->close
159 } else { # generic PSGI
160 psgi_qx_init_cb($self) while $self->{qx_fh};
164 # Similar to `backtick` or "qx" ("perldoc -f qx"), it calls $qx_cb with
165 # the stdout of the given command when done; but respects the given limiter
166 # $env is the PSGI env. As with ``/qx; only use this when output is small
169 my ($self, $env, $limiter, $qx_cb, $qx_arg) = @_;
170 $self->{psgi_env} = $env;
172 open(my $qx_fh, '+>', \$qx_buf) or die; # PerlIO::scalar
173 $self->{qx_cb} = $qx_cb;
174 $self->{qx_arg} = $qx_arg;
175 $self->{qx_fh} = $qx_fh;
176 $self->{qx_buf} = \$qx_buf;
177 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
178 start($self, $limiter, \&psgi_qx_start);
181 # this is called on pipe EOF to reap the process, may be called
182 # via PublicInbox::DS event loop OR via GetlineBody for generic
185 my ($self, $err) = @_; # $err: $!
186 warn "psgi_{return,qx} $err" if defined($err);
188 my ($fh, $qx_fh) = delete(@$self{qw(fh qx_fh)});
189 $fh->close if $fh; # async-only (psgi_return)
194 # typically used for reading CGI headers
195 # We also need to check EINTR for generic PSGI servers.
198 my $hdr_buf = $self->{hdr_buf};
199 my ($ph_cb, $ph_arg) = @{$self->{parse_hdr}};
201 my $r = sysread($self->{rpipe}, $$hdr_buf, 4096,
205 eval { $ret = $ph_cb->($total_rd, $hdr_buf, $ph_arg) };
207 warn "parse_hdr: $@";
208 $ret = [ 500, [], [ "Internal error\n" ] ];
211 # caller should notify us when it's ready:
212 return if $! == EAGAIN;
213 next if $! == EINTR; # immediate retry
214 warn "error reading header: $!";
215 $ret = [ 500, [], [ "Internal error\n" ] ];
217 } until (defined $ret);
218 delete $self->{parse_hdr}; # done parsing headers
222 sub psgi_return_init_cb {
224 my $r = rd_hdr($self) or return;
225 my $env = $self->{psgi_env};
226 my $filter = delete $env->{'qspawn.filter'} //
227 PublicInbox::GzipFilter::qsp_maybe($r->[1], $env);
229 my $wcb = delete $env->{'qspawn.wcb'};
230 my $async = delete $self->{async};
231 if (scalar(@$r) == 3) { # error
233 # calls rpipe->close && ->event_step
236 $self->{rpipe}->close;
241 # done reading headers, handoff to read body
242 my $fh = $wcb->($r); # scalar @$r == 2
243 $fh = $filter->attach($fh) if $filter;
245 $async->async_pass($env->{'psgix.io'}, $fh,
246 delete($self->{hdr_buf}));
247 } else { # for synchronous PSGI servers
248 require PublicInbox::GetlineBody;
249 $r->[2] = PublicInbox::GetlineBody->new($self->{rpipe},
251 ${$self->{hdr_buf}}, $filter);
256 sub psgi_return_start { # may run later, much later...
258 if (my $async = $self->{psgi_env}->{'pi-httpd.async'}) {
259 # PublicInbox::HTTPD::Async->new(rpipe, $cb, $cb_arg, $end_obj)
260 $self->{async} = $async->($self->{rpipe},
261 \&psgi_return_init_cb, $self, $self);
262 } else { # generic PSGI
263 psgi_return_init_cb($self) while $self->{parse_hdr};
267 # Used for streaming the stdout of one process as a PSGI response.
269 # $env is the PSGI env.
270 # optional keys in $env:
271 # $env->{'qspawn.wcb'} - the write callback from the PSGI server
272 # optional, use this if you've already
273 # captured it elsewhere. If not given,
274 # psgi_return will return an anonymous
275 # sub for the PSGI server to call
277 # $env->{'qspawn.filter'} - filter object, responds to ->attach for
278 # pi-httpd.async and ->translate for generic
281 # $limiter - the Limiter object to use (uses the def_limiter if not given)
283 # $parse_hdr - Initial read function; often for parsing CGI header output.
284 # It will be given the return value of sysread from the pipe
285 # and a string ref of the current buffer. Returns an arrayref
286 # for PSGI responses. 2-element arrays in PSGI mean the
287 # body will be streamed, later, via writes (push-based) to
288 # psgix.io. 3-element arrays means the body is available
289 # immediately (or streamed via ->getline (pull-based)).
291 my ($self, $env, $limiter, $parse_hdr, $hdr_arg) = @_;
292 $self->{psgi_env} = $env;
293 $self->{hdr_buf} = \(my $hdr_buf = '');
294 $self->{parse_hdr} = [ $parse_hdr, $hdr_arg ];
295 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
297 # the caller already captured the PSGI write callback from
298 # the PSGI server, so we can call ->start, here:
299 $env->{'qspawn.wcb'} and
300 return start($self, $limiter, \&psgi_return_start);
302 # the caller will return this sub to the PSGI server, so
303 # it can set the response callback (that is, for
304 # PublicInbox::HTTP, the chunked_wcb or identity_wcb callback),
305 # but other HTTP servers are supported:
307 $env->{'qspawn.wcb'} = $_[0];
308 start($self, $limiter, \&psgi_return_start);
312 package PublicInbox::Qspawn::Limiter;
317 my ($class, $max) = @_;
319 # 32 is same as the git-daemon connection limit
323 # RLIMIT_CPU => undef,
324 # RLIMIT_DATA => undef,
325 # RLIMIT_CORE => undef,
330 my ($self, $name, $cfg) = @_;
331 foreach my $rlim (@PublicInbox::Spawn::RLIMITS) {
334 $k = "publicinboxlimiter.$name.$k";
335 defined(my $v = $cfg->{$k}) or next;
336 my @rlimit = split(/\s*,\s*/, $v);
337 if (scalar(@rlimit) == 1) {
338 push @rlimit, $rlimit[0];
339 } elsif (scalar(@rlimit) != 2) {
340 warn "could not parse $k: $v\n";
342 eval { require BSD::Resource };
344 warn "BSD::Resource missing for $rlim";
347 foreach my $i (0..$#rlimit) {
348 next if $rlimit[$i] ne 'INFINITY';
349 $rlimit[$i] = BSD::Resource::RLIM_INFINITY();
351 $self->{$rlim} = \@rlimit;