1 # Copyright (C) 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;
31 use Scalar::Util qw(blessed);
33 # n.b.: we get EAGAIN with public-inbox-httpd, and EINTR on other PSGI servers
34 use Errno qw(EAGAIN EINTR);
38 # declares a command to spawn (but does not spawn it).
39 # $cmd is the command to spawn
40 # $cmd_env is the environ for the child process (not PSGI env)
41 # $opt can include redirects and perhaps other process spawning options
42 # {qsp_err} is an optional error buffer callers may access themselves
44 my ($class, $cmd, $cmd_env, $opt) = @_;
45 bless { args => [ $cmd, $cmd_env, $opt ] }, $class;
49 my ($self, $start_cb, $limiter) = @_;
51 my ($cmd, $cmd_env, $opt) = @{delete $self->{args}};
52 my %o = %{$opt || {}};
53 $self->{limiter} = $limiter;
54 foreach my $k (@PublicInbox::Spawn::RLIMITS) {
55 if (defined(my $rlimit = $limiter->{$k})) {
60 $self->{-quiet} = 1 if $o{quiet};
61 $o{cb_arg} = [ \&waitpid_err, $self ];
63 # popen_rd may die on EMFILE, ENFILE
64 $self->{rpipe} = popen_rd($cmd, $cmd_env, \%o) // die "E: $!";
65 $limiter->{running}++;
66 $start_cb->($self); # EPOLL_CTL_ADD may ENOSPC/ENOMEM
68 finish($self, $@) if $@;
74 # process is done, spawn whatever's in the queue
75 my $limiter = delete $self->{limiter} or return;
76 my $running = --$limiter->{running};
78 if ($running < $limiter->{max}) {
79 if (my $next = shift(@{$limiter->{run_queue}})) {
80 _do_spawn(@$next, $limiter);
83 if (my $err = $self->{_err}) { # set by finish or waitpid_err
85 if (my $dst = $self->{qsp_err}) {
86 $$dst .= $$dst ? " $err" : "; $err";
88 warn "@{$self->{cmd}}: $err\n" if !$self->{-quiet};
91 my ($env, $qx_cb, $qx_arg, $qx_buf) =
92 delete @$self{qw(psgi_env qx_cb qx_arg qx_buf)};
94 eval { $qx_cb->($qx_buf, $qx_arg) };
96 warn "E: $@"; # hope qspawn.wcb can handle it
98 return if $self->{passed}; # another command chained it
99 if (my $wcb = delete $env->{'qspawn.wcb'}) {
100 # have we started writing, yet?
101 my $code = delete $env->{'qspawn.fallback'} // 500;
102 require PublicInbox::WwwStatic;
103 $wcb->(PublicInbox::WwwStatic::r($code));
107 sub DESTROY { finalize($_[0]) } # ->finalize is idempotent
109 sub waitpid_err { # callback for awaitpid
110 my (undef, $self) = @_; # $_[0]: pid
111 $self->{_err} = ''; # for defined check in ->finish
113 my $status = $? >> 8;
115 $self->{_err} .= "exit status=$status";
116 $self->{_err} .= " signal=$sig" if $sig;
118 finalize($self) if !$self->{rpipe};
122 my ($self, $err) = @_;
123 $self->{_err} //= $err; # only for $@
125 # we can safely finalize if pipe was closed before, or if
126 # {_err} is defined by waitpid_err. Deleting {rpipe} will
127 # trigger PublicInbox::ProcessPipe::DESTROY -> waitpid_err,
128 # but it may not fire right away if inside the event loop.
129 my $closed_before = !delete($self->{rpipe});
130 finalize($self) if $closed_before || defined($self->{_err});
134 my ($self, $limiter, $start_cb) = @_;
135 if ($limiter->{running} < $limiter->{max}) {
136 _do_spawn($self, $start_cb, $limiter);
138 push @{$limiter->{run_queue}}, [ $self, $start_cb ];
142 sub psgi_qx_init_cb { # this may be PublicInbox::HTTPD::Async {cb}
146 $r = sysread($self->{rpipe}, $buf, 65536);
148 return if $! == EAGAIN; # try again when notified
149 goto reread if $! == EINTR;
150 event_step($self, $!);
151 } elsif (my $as = delete $self->{async}) { # PublicInbox::HTTPD::Async
152 $as->async_pass($self->{psgi_env}->{'psgix.io'},
153 $self->{qx_fh}, \$buf);
154 } elsif ($r) { # generic PSGI:
155 print { $self->{qx_fh} } $buf;
157 event_step($self, undef);
163 if (my $async = $self->{psgi_env}->{'pi-httpd.async'}) {
164 # PublicInbox::HTTPD::Async->new(rpipe, $cb, cb_arg, $end_obj)
165 $self->{async} = $async->($self->{rpipe},
166 \&psgi_qx_init_cb, $self, $self);
167 # init_cb will call ->async_pass or ->close
168 } else { # generic PSGI
169 psgi_qx_init_cb($self) while $self->{qx_fh};
173 # Similar to `backtick` or "qx" ("perldoc -f qx"), it calls $qx_cb with
174 # the stdout of the given command when done; but respects the given limiter
175 # $env is the PSGI env. As with ``/qx; only use this when output is small
178 my ($self, $env, $limiter, $qx_cb, $qx_arg) = @_;
179 $self->{psgi_env} = $env;
181 open(my $qx_fh, '+>', \$qx_buf) or die; # PerlIO::scalar
182 $self->{qx_cb} = $qx_cb;
183 $self->{qx_arg} = $qx_arg;
184 $self->{qx_fh} = $qx_fh;
185 $self->{qx_buf} = \$qx_buf;
186 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
187 start($self, $limiter, \&psgi_qx_start);
190 # this is called on pipe EOF to reap the process, may be called
191 # via PublicInbox::DS event loop OR via GetlineBody for generic
194 my ($self, $err) = @_; # $err: $!
195 warn "psgi_{return,qx} $err" if defined($err);
197 my ($fh, $qx_fh) = delete(@$self{qw(qfh qx_fh)});
198 $fh->close if $fh; # async-only (psgi_return)
203 # typically used for reading CGI headers
204 # We also need to check EINTR for generic PSGI servers.
207 my $hdr_buf = $self->{hdr_buf};
208 my ($ph_cb, $ph_arg) = @{$self->{parse_hdr}};
209 until (defined($ret)) {
210 my $r = sysread($self->{rpipe}, $$hdr_buf, 4096,
214 eval { $ret = $ph_cb->($total_rd, $hdr_buf, $ph_arg) };
216 warn "parse_hdr: $@";
217 $ret = [ 500, [], [ "Internal error\n" ] ];
218 } elsif (!defined($ret) && !$r) {
220 EOF parsing headers from @{$self->{cmd}} ($self->{psgi_env}->{REQUEST_URI})
222 $ret = [ 500, [], [ "Internal error\n" ] ];
225 # caller should notify us when it's ready:
226 return if $! == EAGAIN;
227 next if $! == EINTR; # immediate retry
228 warn "error reading header: $!";
229 $ret = [ 500, [], [ "Internal error\n" ] ];
232 delete $self->{parse_hdr}; # done parsing headers
236 sub psgi_return_init_cb { # this may be PublicInbox::HTTPD::Async {cb}
238 my $r = rd_hdr($self) or return;
239 my $env = $self->{psgi_env};
242 # this is for RepoAtom since that can fire after parse_cgi_headers
243 if (ref($r) eq 'ARRAY' && blessed($r->[2]) && $r->[2]->can('attach')) {
246 $filter //= delete($env->{'qspawn.filter'}) // (ref($r) eq 'ARRAY' ?
247 PublicInbox::GzipFilter::qsp_maybe($r->[1], $env) : undef);
249 my $wcb = delete $env->{'qspawn.wcb'};
250 my $async = delete $self->{async}; # PublicInbox::HTTPD::Async
251 if (ref($r) ne 'ARRAY' || scalar(@$r) == 3) { # error
252 if ($async) { # calls rpipe->close && ->event_step
253 $async->close; # PublicInbox::HTTPD::Async::close
254 } else { # generic PSGI, use PublicInbox::ProcessPipe::CLOSE
255 delete($self->{rpipe})->close;
258 if (ref($r) eq 'ARRAY') { # error
260 } elsif (ref($r) eq 'CODE') { # chain another command
266 # done reading headers, handoff to read body
267 my $fh = $wcb->($r); # scalar @$r == 2
268 $fh = $filter->attach($fh) if $filter;
270 $async->async_pass($env->{'psgix.io'}, $fh,
271 delete($self->{hdr_buf}));
272 } else { # for synchronous PSGI servers
273 require PublicInbox::GetlineBody;
274 my $buf = delete $self->{hdr_buf};
275 $r->[2] = PublicInbox::GetlineBody->new($self->{rpipe},
276 \&event_step, $self, $$buf, $filter);
281 sub psgi_return_start { # may run later, much later...
283 if (my $cb = $self->{psgi_env}->{'pi-httpd.async'}) {
284 # PublicInbox::HTTPD::Async->new(rpipe, $cb, $cb_arg, $end_obj)
285 $self->{async} = $cb->($self->{rpipe},
286 \&psgi_return_init_cb, $self, $self);
287 } else { # generic PSGI
288 psgi_return_init_cb($self) while $self->{parse_hdr};
292 # Used for streaming the stdout of one process as a PSGI response.
294 # $env is the PSGI env.
295 # optional keys in $env:
296 # $env->{'qspawn.wcb'} - the write callback from the PSGI server
297 # optional, use this if you've already
298 # captured it elsewhere. If not given,
299 # psgi_return will return an anonymous
300 # sub for the PSGI server to call
302 # $env->{'qspawn.filter'} - filter object, responds to ->attach for
303 # pi-httpd.async and ->translate for generic
306 # $limiter - the Limiter object to use (uses the def_limiter if not given)
308 # $parse_hdr - Initial read function; often for parsing CGI header output.
309 # It will be given the return value of sysread from the pipe
310 # and a string ref of the current buffer. Returns an arrayref
311 # for PSGI responses. 2-element arrays in PSGI mean the
312 # body will be streamed, later, via writes (push-based) to
313 # psgix.io. 3-element arrays means the body is available
314 # immediately (or streamed via ->getline (pull-based)).
316 my ($self, $env, $limiter, $parse_hdr, $hdr_arg) = @_;
317 $self->{psgi_env} = $env;
318 $self->{hdr_buf} = \(my $hdr_buf = '');
319 $self->{parse_hdr} = [ $parse_hdr, $hdr_arg ];
320 $limiter ||= $def_limiter ||= PublicInbox::Qspawn::Limiter->new(32);
322 # the caller already captured the PSGI write callback from
323 # the PSGI server, so we can call ->start, here:
324 $env->{'qspawn.wcb'} and
325 return start($self, $limiter, \&psgi_return_start);
327 # the caller will return this sub to the PSGI server, so
328 # it can set the response callback (that is, for
329 # PublicInbox::HTTP, the chunked_wcb or identity_wcb callback),
330 # but other HTTP servers are supported:
332 $env->{'qspawn.wcb'} = $_[0];
333 start($self, $limiter, \&psgi_return_start);
337 package PublicInbox::Qspawn::Limiter;
341 my ($class, $max) = @_;
343 # 32 is same as the git-daemon connection limit
347 # RLIMIT_CPU => undef,
348 # RLIMIT_DATA => undef,
349 # RLIMIT_CORE => undef,
354 my ($self, $name, $cfg) = @_;
355 foreach my $rlim (@PublicInbox::Spawn::RLIMITS) {
358 $k = "publicinboxlimiter.$name.$k";
359 defined(my $v = $cfg->{$k}) or next;
360 my @rlimit = split(/\s*,\s*/, $v);
361 if (scalar(@rlimit) == 1) {
362 push @rlimit, $rlimit[0];
363 } elsif (scalar(@rlimit) != 2) {
364 warn "could not parse $k: $v\n";
366 eval { require BSD::Resource };
368 warn "BSD::Resource missing for $rlim";
371 foreach my $i (0..$#rlimit) {
372 next if $rlimit[$i] ne 'INFINITY';
373 $rlimit[$i] = BSD::Resource::RLIM_INFINITY();
375 $self->{$rlim} = \@rlimit;