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 ] }, $class;
49 my ($cmd, $env, $opts) = @{$self->{args}};
50 my %opts = %{$opts || {}};
51 my $limiter = $self->{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 $cb->($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 # done, spawn whatever's in the queue
98 my $limiter = $self->{limiter};
99 my $running = --$limiter->{running};
101 # limiter->{max} may change dynamically
102 if (($running || $limiter->{running}) < $limiter->{max}) {
103 if (my $next = shift @{$limiter->{run_queue}}) {
110 my $env = $self->{env} or return;
111 if (!$env->{'qspawn.quiet'}) {
112 log_err($env, join(' ', @{$self->{args}}) . ": $err");
116 sub do_waitpid ($;$) {
117 my ($self, $env) = @_;
118 my $pid = $self->{pid};
119 eval { # PublicInbox::DS may not be loaded
120 PublicInbox::DS::dwaitpid($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, $ret);
132 my ($self, $env) = @_;
133 if (delete $self->{rpipe}) {
134 do_waitpid($self, $env);
137 # limiter->{max} may change dynamically
138 my $limiter = $self->{limiter};
139 if ($limiter->{running} < $limiter->{max}) {
140 if (my $next = shift @{$limiter->{run_queue}}) {
144 $self->{err}; # may be meaningless if non-blocking
148 my ($self, $limiter, $cb) = @_;
149 $self->{limiter} = $limiter;
151 if ($limiter->{running} < $limiter->{max}) {
152 _do_spawn($self, $cb);
154 push @{$limiter->{run_queue}}, [ $self, $cb ];
158 # Similar to `backtick` or "qx" ("perldoc -f qx"), it calls $qx_cb with
159 # the stdout of the given command when done; but respects the given limiter
160 # $env is the PSGI env. As with ``/qx; only use this when output is small
163 my ($self, $env, $limiter, $qx_cb) = @_;
165 open(my $qx, '+>', \$scalar) or die; # PerlIO::scalar
167 my $err = $_[0]; # $!
168 log_err($env, "psgi_qx: $err") if defined($err);
170 eval { $qx_cb->(\$scalar) };
171 $qx = $scalar = undef;
173 my $rpipe; # comes from popen_rd
174 my $async = $env->{'pi-httpd.async'};
178 $r = sysread($rpipe, $buf, 65536);
180 $async->async_pass($env->{'psgix.io'}, $qx, \$buf);
181 } elsif (defined $r) {
182 $r ? $qx->write($buf) : $end->();
184 return if $! == EAGAIN; # try again when notified
185 goto reread if $! == EINTR;
189 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
190 $self->start($limiter, sub { # may run later, much later...
191 ($rpipe) = @_; # popen_rd result
193 # PublicInbox::HTTPD::Async->new($rpipe, $cb, $end)
194 $async = $async->($rpipe, $cb, $end);
195 } else { # generic PSGI
201 # create a filter for "push"-based streaming PSGI writes used by HTTPD::Async
203 my ($fh, $filter) = @_;
204 Plack::Util::inline_object(
206 $fh->write($filter->(undef));
210 $fh->write($filter->($_[0]));
214 # Used for streaming the stdout of one process as a PSGI response.
216 # $env is the PSGI env.
217 # optional keys in $env:
218 # $env->{'qspawn.wcb'} - the write callback from the PSGI server
219 # optional, use this if you've already
220 # captured it elsewhere. If not given,
221 # psgi_return will return an anonymous
222 # sub for the PSGI server to call
224 # $env->{'qspawn.filter'} - filter callback, receives a string as input,
227 # $limiter - the Limiter object to use (uses the def_limiter if not given)
229 # $parse_hdr - Initial read function; often for parsing CGI header output.
230 # It will be given the return value of sysread from the pipe
231 # and a string ref of the current buffer. Returns an arrayref
232 # for PSGI responses. 2-element arrays in PSGI mean the
233 # body will be streamed, later, via writes (push-based) to
234 # psgix.io. 3-element arrays means the body is available
235 # immediately (or streamed via ->getline (pull-based)).
237 my ($self, $env, $limiter, $parse_hdr) = @_;
240 my $err = $_[0]; # $!
241 log_err($env, "psgi_return: $err") if defined($err);
243 $fh->close if $fh; # async-only
248 # typically used for reading CGI headers
249 # we must loop until EAGAIN for EPOLLET in HTTPD/Async.pm
250 # We also need to check EINTR for generic PSGI servers.
254 my $r = sysread($rpipe, $buf, 4096, length($buf));
257 $ret = $parse_hdr->($r ? $total_rd : 0, \$buf);
259 # caller should notify us when it's ready:
260 return if $! == EAGAIN;
261 next if $! == EINTR; # immediate retry
262 log_err($env, "error reading header: $!");
263 $ret = [ 500, [], [ "Internal error\n" ] ];
265 } until (defined $ret);
269 my $wcb = delete $env->{'qspawn.wcb'};
270 my $async = $env->{'pi-httpd.async'};
273 my $r = $rd_hdr->() or return;
274 $rd_hdr = undef; # done reading headers
275 my $filter = delete $env->{'qspawn.filter'};
276 if (scalar(@$r) == 3) { # error
278 $async->close; # calls rpipe->close and $end
285 # done reading headers, handoff to read body
286 $fh = $wcb->($r); # scalar @$r == 2
287 $fh = filter_fh($fh, $filter) if $filter;
288 $async->async_pass($env->{'psgix.io'}, $fh, \$buf);
289 } else { # for synchronous PSGI servers
290 require PublicInbox::GetlineBody;
291 $r->[2] = PublicInbox::GetlineBody->new($rpipe, $end,
296 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
297 my $start_cb = sub { # may run later, much later...
300 # PublicInbox::HTTPD::Async->new($rpipe, $cb, $end)
301 $async = $async->($rpipe, $cb, $end);
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;