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);
31 # n.b.: we get EAGAIN with public-inbox-httpd, and EINTR on other PSGI servers
32 use Errno qw(EAGAIN EINTR);
36 # declares a command to spawn (but does not spawn it).
37 # $cmd is the command to spawn
38 # $cmd_env is the environ for the child process (not PSGI env)
39 # $opt can include redirects and perhaps other process spawning options
41 my ($class, $cmd, $cmd_env, $opt) = @_;
42 bless { args => [ $cmd, $cmd_env, $opt ] }, $class;
46 my ($self, $start_cb, $limiter) = @_;
48 my ($cmd, $cmd_env, $opts) = @{$self->{args}};
49 my %opts = %{$opts || {}};
50 $self->{limiter} = $limiter;
51 foreach my $k (PublicInbox::Spawn::RLIMITS()) {
52 if (defined(my $rlimit = $limiter->{$k})) {
57 ($self->{rpipe}, $self->{pid}) = popen_rd($cmd, $cmd_env, \%opts);
59 # drop any IO handles opt was holding open via $opt->{hold}
60 # No need to hold onto the descriptor once the child process has it.
61 $self->{args} = $cmd; # keep this around for logging
63 if (defined $self->{pid}) {
64 $limiter->{running}++;
72 my ($child_error) = @_; # typically $?
73 my $exitstatus = ($child_error >> 8) or return;
74 my $sig = $child_error & 127;
75 my $msg = "exit status=$exitstatus";
76 $msg .= " signal=$sig" if $sig;
82 $env->{'psgi.errors'}->print($msg, "\n");
85 # callback for dwaitpid
86 sub waitpid_err ($$) {
87 my ($self, $pid) = @_;
88 my $xpid = delete $self->{pid};
90 if ($pid > 0) { # success!
92 } elsif ($pid < 0) { # ??? does this happen in our case?
93 $err = "W: waitpid($xpid, 0) => $pid: $!";
94 } # else should not be called with pid == 0
96 my ($env, $qx_cb, $qx_arg, $qx_buf) =
97 delete @$self{qw(psgi_env qx_cb qx_arg qx_buf)};
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 eval { $qx_cb->($qx_buf, $qx_arg) } if $qx_cb;
120 my $pid = $self->{pid};
121 # PublicInbox::DS may not be loaded
122 eval { PublicInbox::DS::dwaitpid($pid, \&waitpid_err, $self) };
123 # done if we're running in PublicInbox::DS::EventLoop
125 # non public-inbox-{httpd,nntpd} callers may block:
126 my $ret = waitpid($pid, 0);
127 waitpid_err($self, $ret);
133 if (delete $self->{rpipe}) {
136 my ($env, $qx_cb, $qx_arg, $qx_buf) =
137 delete @$self{qw(psgi_env qx_cb qx_arg qx_buf)};
138 eval { $qx_cb->($qx_buf, $qx_arg) } if $qx_cb;
143 my ($self, $limiter, $start_cb) = @_;
144 if ($limiter->{running} < $limiter->{max}) {
145 _do_spawn($self, $start_cb, $limiter);
147 push @{$limiter->{run_queue}}, [ $self, $start_cb ];
151 sub psgi_qx_init_cb {
153 my $async = delete $self->{async};
155 my $qx_fh = $self->{qx_fh};
157 $r = sysread($self->{rpipe}, $buf, 65536);
159 $async->async_pass($self->{psgi_env}->{'psgix.io'},
161 } elsif (defined $r) {
162 $r ? $qx_fh->write($buf) : event_step($self, undef);
164 return if $! == EAGAIN; # try again when notified
165 goto reread if $! == EINTR;
166 event_step($self, $!);
172 if (my $async = $self->{psgi_env}->{'pi-httpd.async'}) {
173 # PublicInbox::HTTPD::Async->new(rpipe, $cb, cb_arg, $end_obj)
174 $self->{async} = $async->($self->{rpipe},
175 \&psgi_qx_init_cb, $self, $self);
176 # init_cb will call ->async_pass or ->close
177 } else { # generic PSGI
178 psgi_qx_init_cb($self) while $self->{qx_fh};
182 # Similar to `backtick` or "qx" ("perldoc -f qx"), it calls $qx_cb with
183 # the stdout of the given command when done; but respects the given limiter
184 # $env is the PSGI env. As with ``/qx; only use this when output is small
187 my ($self, $env, $limiter, $qx_cb, $qx_arg) = @_;
188 $self->{psgi_env} = $env;
190 open(my $qx_fh, '+>', \$qx_buf) or die; # PerlIO::scalar
191 $self->{qx_cb} = $qx_cb;
192 $self->{qx_arg} = $qx_arg;
193 $self->{qx_fh} = $qx_fh;
194 $self->{qx_buf} = \$qx_buf;
195 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
196 start($self, $limiter, \&psgi_qx_start);
199 # this is called on pipe EOF to reap the process, may be called
200 # via PublicInbox::DS event loop OR via GetlineBody for generic
203 my ($self, $err) = @_; # $err: $!
204 log_err($self->{psgi_env}, "psgi_{return,qx} $err") if defined($err);
206 my ($fh, $qx_fh) = delete(@$self{qw(fh qx_fh)});
207 $fh->close if $fh; # async-only (psgi_return)
212 # typically used for reading CGI headers
213 # we must loop until EAGAIN for EPOLLET in HTTPD/Async.pm
214 # We also need to check EINTR for generic PSGI servers.
217 my $hdr_buf = $self->{hdr_buf};
219 my $r = sysread($self->{rpipe}, $$hdr_buf, 4096,
223 $ret = $self->{parse_hdr}->($total_rd, $hdr_buf);
225 # caller should notify us when it's ready:
226 return if $! == EAGAIN;
227 next if $! == EINTR; # immediate retry
228 log_err($self->{psgi_env}, "error reading header: $!");
229 $ret = [ 500, [], [ "Internal error\n" ] ];
231 } until (defined $ret);
232 delete $self->{parse_hdr}; # done parsing headers
236 sub psgi_return_init_cb {
238 my $r = rd_hdr($self) or return;
239 my $env = $self->{psgi_env};
240 my $wcb = delete $env->{'qspawn.wcb'};
241 my $async = delete $self->{async};
242 if (scalar(@$r) == 3) { # error
244 # calls rpipe->close && ->event_step
247 $self->{rpipe}->close;
252 # done reading headers, handoff to read body
253 my $fh = $wcb->($r); # scalar @$r == 2
255 $async->async_pass($env->{'psgix.io'}, $fh,
256 delete($self->{hdr_buf}));
257 } else { # for synchronous PSGI servers
258 require PublicInbox::GetlineBody;
259 $r->[2] = PublicInbox::GetlineBody->new($self->{rpipe},
261 ${$self->{hdr_buf}});
265 # Workaround a leak under Perl 5.16.3 when combined with
266 # Plack::Middleware::Deflater:
270 sub psgi_return_start { # may run later, much later...
272 if (my $async = $self->{psgi_env}->{'pi-httpd.async'}) {
273 # PublicInbox::HTTPD::Async->new(rpipe, $cb, $cb_arg, $end_obj)
274 $self->{async} = $async->($self->{rpipe},
275 \&psgi_return_init_cb, $self, $self);
276 } else { # generic PSGI
277 psgi_return_init_cb($self) while $self->{parse_hdr};
281 # Used for streaming the stdout of one process as a PSGI response.
283 # $env is the PSGI env.
284 # optional keys in $env:
285 # $env->{'qspawn.wcb'} - the write callback from the PSGI server
286 # optional, use this if you've already
287 # captured it elsewhere. If not given,
288 # psgi_return will return an anonymous
289 # sub for the PSGI server to call
291 # $limiter - the Limiter object to use (uses the def_limiter if not given)
293 # $parse_hdr - Initial read function; often for parsing CGI header output.
294 # It will be given the return value of sysread from the pipe
295 # and a string ref of the current buffer. Returns an arrayref
296 # for PSGI responses. 2-element arrays in PSGI mean the
297 # body will be streamed, later, via writes (push-based) to
298 # psgix.io. 3-element arrays means the body is available
299 # immediately (or streamed via ->getline (pull-based)).
301 my ($self, $env, $limiter, $parse_hdr) = @_;
302 $self->{psgi_env} = $env;
303 $self->{hdr_buf} = \(my $hdr_buf = '');
304 $self->{parse_hdr} = $parse_hdr;
305 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
307 # the caller already captured the PSGI write callback from
308 # the PSGI server, so we can call ->start, here:
309 $env->{'qspawn.wcb'} and
310 return start($self, $limiter, \&psgi_return_start);
312 # the caller will return this sub to the PSGI server, so
313 # it can set the response callback (that is, for
314 # PublicInbox::HTTP, the chunked_wcb or identity_wcb callback),
315 # but other HTTP servers are supported:
317 $env->{'qspawn.wcb'} = $_[0];
318 start($self, $limiter, \&psgi_return_start);
322 package PublicInbox::Qspawn::Limiter;
327 my ($class, $max) = @_;
329 # 32 is same as the git-daemon connection limit
333 # RLIMIT_CPU => undef,
334 # RLIMIT_DATA => undef,
335 # RLIMIT_CORE => undef,
340 my ($self, $name, $config) = @_;
341 foreach my $rlim (PublicInbox::Spawn::RLIMITS()) {
344 $k = "publicinboxlimiter.$name.$k";
345 defined(my $v = $config->{$k}) or next;
346 my @rlimit = split(/\s*,\s*/, $v);
347 if (scalar(@rlimit) == 1) {
348 push @rlimit, $rlimit[0];
349 } elsif (scalar(@rlimit) != 2) {
350 warn "could not parse $k: $v\n";
352 eval { require BSD::Resource };
354 warn "BSD::Resource missing for $rlim";
357 foreach my $i (0..$#rlimit) {
358 next if $rlimit[$i] ne 'INFINITY';
359 $rlimit[$i] = BSD::Resource::RLIM_INFINITY();
361 $self->{$rlim} = \@rlimit;