1 # Copyright (C) 2020-2021 all contributors <meta@public-inbox.org>
2 # License: AGPL-3.0+ <https://www.gnu.org/licenses/agpl-3.0.txt>
4 # base class for remote IPC calls and workqueues, requires Storable or Sereal
5 package PublicInbox::IPC;
8 use Carp qw(confess croak);
9 use PublicInbox::DS qw(dwaitpid);
10 use PublicInbox::Spawn;
11 use POSIX qw(WNOHANG);
12 use Socket qw(AF_UNIX MSG_EOR);
13 my $SEQPACKET = eval { Socket::SOCK_SEQPACKET() }; # portable enough?
14 use constant PIPE_BUF => $^O eq 'linux' ? 4096 : POSIX::_POSIX_PIPE_BUF();
15 my $WQ_MAX_WORKERS = 4096;
17 # ->imports at BEGIN turns sereal_*_with_object into custom ops on 5.14+
18 # and eliminate method call overhead
21 require Sereal::Encoder;
22 require Sereal::Decoder;
23 Sereal::Encoder->import('sereal_encode_with_object');
24 Sereal::Decoder->import('sereal_decode_with_object');
25 ($enc, $dec) = (Sereal::Encoder->new, Sereal::Decoder->new);
29 if ($enc && $dec) { # should be custom ops
30 *freeze = sub ($) { sereal_encode_with_object $enc, $_[0] };
31 *thaw = sub ($) { sereal_decode_with_object $dec, $_[0], my $ret };
33 eval { # some distros have Storable as a separate package from Perl
35 Storable->import(qw(freeze thaw));
37 } // warn("Storable (part of Perl) missing: $@\n");
40 my $recv_cmd = PublicInbox::Spawn->can('recv_cmd4');
41 my $send_cmd = PublicInbox::Spawn->can('send_cmd4') // do {
42 require PublicInbox::CmdIPC4;
43 $recv_cmd //= PublicInbox::CmdIPC4->can('recv_cmd4');
44 PublicInbox::CmdIPC4->can('send_cmd4');
47 sub wq_set_recv_modes {
48 my ($self, @modes) = @_;
49 $self->{-wq_recv_modes} = \@modes;
54 defined(my $len = <$r>) or return;
55 chop($len) eq "\n" or croak "no LF byte in $len";
56 defined(my $n = read($r, my $buf, $len)) or croak "read error: $!";
57 $n == $len or croak "short read: $n != $len";
63 my $buf = freeze($ref);
64 length($buf) . "\n" . $buf;
69 print $w _pack_rec($ref) or croak "print: $!";
72 sub ipc_return ($$$) {
73 my ($w, $ret, $exc) = @_;
74 _send_rec($w, $exc ? bless(\$exc, 'PublicInbox::IPC::Die') : $ret);
77 sub ipc_worker_loop ($$$) {
78 my ($self, $r_req, $w_res) = @_;
79 my ($rec, $wantarray, $sub, @args);
81 while ($rec = _get_rec($r_req)) {
82 ($wantarray, $sub, @args) = @$rec;
83 # no waiting if client doesn't care,
84 # this is the overwhelmingly likely case
85 if (!defined($wantarray)) {
86 eval { $self->$sub(@args) };
87 warn "$$ die: $@ (from nowait $sub)\n" if $@;
88 } elsif ($wantarray) {
89 my @ret = eval { $self->$sub(@args) };
90 ipc_return($w_res, \@ret, $@);
91 } else { # '' => wantscalar
92 my $ret = eval { $self->$sub(@args) };
93 ipc_return($w_res, \$ret, $@);
98 # starts a worker if Sereal or Storable is installed
99 sub ipc_worker_spawn {
100 my ($self, $ident, $oldset) = @_;
101 return unless $enc; # no Sereal or Storable
102 return if ($self->{-ipc_ppid} // -1) == $$; # idempotent
103 delete(@$self{qw(-ipc_req -ipc_res -ipc_ppid -ipc_pid)});
104 pipe(my ($r_req, $w_req)) or die "pipe: $!";
105 pipe(my ($r_res, $w_res)) or die "pipe: $!";
106 my $sigset = $oldset // PublicInbox::DS::block_signals();
107 $self->ipc_atfork_prepare;
108 my $seed = rand(0xffffffff);
109 my $pid = fork // die "fork: $!";
112 eval { PublicInbox::DS->Reset };
113 delete @$self{qw(-wq_s1 -wq_workers -wq_ppid)};
114 $w_req = $r_res = undef;
115 $w_res->autoflush(1);
116 $SIG{$_} = 'IGNORE' for (qw(TERM INT QUIT));
118 PublicInbox::DS::sig_setmask($sigset);
119 my $on_destroy = $self->ipc_atfork_child;
120 eval { ipc_worker_loop($self, $r_req, $w_res) };
121 die "worker $ident PID:$$ died: $@\n" if $@;
124 PublicInbox::DS::sig_setmask($sigset) unless $oldset;
125 $r_req = $w_res = undef;
126 $w_req->autoflush(1);
127 $self->{-ipc_req} = $w_req;
128 $self->{-ipc_res} = $r_res;
129 $self->{-ipc_ppid} = $$;
130 $self->{-ipc_pid} = $pid;
133 sub ipc_worker_reap { # dwaitpid callback
134 my ($self, $pid) = @_;
135 # SIGTERM (15) is our default exit signal
136 warn "PID:$pid died with \$?=$?\n" if $? && ($? & 127) != 15;
141 my $pids = delete $self->{"-wq_old_pids.$$"} or return;
142 dwaitpid($_, \&ipc_worker_reap, $self) for @$pids;
145 # for base class, override in sub classes
146 sub ipc_atfork_prepare {}
148 sub ipc_atfork_child {
150 my $io = delete($self->{-ipc_atfork_child_close}) or return;
155 # idempotent, can be called regardless of whether worker is active or not
156 sub ipc_worker_stop {
158 my ($pid, $ppid) = delete(@$self{qw(-ipc_pid -ipc_ppid)});
159 my ($w_req, $r_res) = delete(@$self{qw(-ipc_req -ipc_res)});
160 if (!$w_req && !$r_res) {
161 die "unexpected PID:$pid without IPC pipes" if $pid;
164 die 'no PID with IPC pipes' unless $pid;
165 $w_req = $r_res = undef;
167 return if $$ != $ppid;
168 dwaitpid($pid, \&ipc_worker_reap, $self);
171 # use this if we have multiple readers reading curl or "pigz -dc"
172 # and writing to the same store
175 require PublicInbox::Lock;
176 $self->{-ipc_lock} //= bless { lock_path => $f }, 'PublicInbox::Lock'
179 sub ipc_async_wait ($$) {
180 my ($self, $max) = @_; # max == -1 to wait for all
181 my $aif = $self->{-async_inflight} or return;
182 my $r_res = $self->{-ipc_res} or die 'BUG: no ipc_res';
183 while (my ($sub, $bytes, $cb, $cb_arg) = splice(@$aif, 0, 4)) {
184 my $ret = _get_rec($r_res) //
185 die "no response on $sub (req.size=$bytes)";
186 $self->{-async_inflight_bytes} -= $bytes;
188 eval { $cb->($cb_arg, $ret) };
189 warn "E: $sub callback error: $@\n" if $@;
190 return if --$max == 0;
194 # call $self->$sub(@args), on a worker if ipc_worker_spawn was used
196 my ($self, $sub, @args) = @_;
197 if (my $w_req = $self->{-ipc_req}) { # run in worker
198 my $ipc_lock = $self->{-ipc_lock};
199 my $lock = $ipc_lock ? $ipc_lock->lock_for_scope : undef;
200 if (defined(wantarray)) {
201 my $r_res = $self->{-ipc_res} or die 'BUG: no ipc_res';
202 ipc_async_wait($self, -1);
203 _send_rec($w_req, [ wantarray, $sub, @args ]);
204 my $ret = _get_rec($r_res) // die "no response on $sub";
205 die $$ret if ref($ret) eq 'PublicInbox::IPC::Die';
206 wantarray ? @$ret : $$ret;
207 } else { # likely, fire-and-forget into pipe
208 _send_rec($w_req, [ undef , $sub, @args ]);
210 } else { # run locally
216 my ($self, $sub, $sub_args, $cb, $cb_arg) = @_;
217 if (my $w_req = $self->{-ipc_req}) { # run in worker
218 my $rec = _pack_rec([ 1, $sub, @$sub_args ]);
219 my $cur_bytes = \($self->{-async_inflight_bytes} //= 0);
220 while (($$cur_bytes + length($rec)) > PIPE_BUF) {
221 ipc_async_wait($self, 1);
223 my $ipc_lock = $self->{-ipc_lock};
224 my $lock = $ipc_lock ? $ipc_lock->lock_for_scope : undef;
225 print $w_req $rec or croak "print: $!";
226 $$cur_bytes += length($rec);
227 push @{$self->{-async_inflight}},
228 $sub, length($rec), $cb, $cb_arg;
230 my $ret = [ eval { $self->$sub(@$sub_args) } ];
232 $ret = ( bless(\$exc, 'PublicInbox::IPC::Die') );
234 eval { $cb->($cb_arg, $ret) };
235 warn "E: $sub callback error: $@\n" if $@;
239 # needed when there's multiple IPC workers and the parent forking
240 # causes newer siblings to inherit older siblings sockets
241 sub ipc_sibling_atfork_child {
243 my ($pid, undef) = delete(@$self{qw(-ipc_pid -ipc_ppid)});
244 delete(@$self{qw(-ipc_req -ipc_res)});
245 $pid == $$ and die "BUG: $$ ipc_atfork_child called on itself";
248 sub wq_worker_loop ($) {
250 my $len = $self->{wq_req_len} // (4096 * 33);
251 my $s2 = $self->{-wq_s2} // die 'BUG: no -wq_s2';
253 my @fds = $recv_cmd->($s2, my $buf, $len) or return; # EOF
254 my @m = @{$self->{-wq_recv_modes} // [qw( +<&= >&= >&= )]};
257 my $mode = shift(@m);
258 if (open(my $cmdfh, $mode, $fd)) {
259 $self->{$nfd++} = $cmdfh;
260 $cmdfh->autoflush(1);
262 die "$$ open($mode$fd) (FD:$nfd): $!";
265 # Sereal dies on truncated data, Storable returns undef
266 my $args = thaw($buf) //
267 die "thaw error on buffer of size:".length($buf);
268 my $sub = shift @$args;
269 eval { $self->$sub(@$args) };
270 warn "$$ wq_worker: $@" if $@ &&
271 ref($@) ne 'PublicInbox::SIGPIPE';
272 delete @$self{0..($nfd-1)};
276 sub wq_do { # always async
277 my ($self, $sub, $ios, @args) = @_;
278 if (my $s1 = $self->{-wq_s1}) { # run in worker
279 my $fds = [ map { fileno($_) } @$ios ];
280 $send_cmd->($s1, $fds, freeze([$sub, @args]), MSG_EOR);
282 @$self{0..$#$ios} = @$ios;
283 eval { $self->$sub(@args) };
284 warn "wq_do: $@" if $@ && ref($@) ne 'PublicInbox::SIGPIPE';
285 delete @$self{0..$#$ios}; # don't close
289 sub _wq_worker_start ($$) {
290 my ($self, $oldset) = @_;
291 my $seed = rand(0xffffffff);
292 my $pid = fork // die "fork: $!";
295 eval { PublicInbox::DS->Reset };
296 delete @$self{qw(-wq_s1 -wq_workers -wq_ppid)};
297 $SIG{$_} = 'IGNORE' for (qw(PIPE TTOU TTIN));
298 $SIG{$_} = 'DEFAULT' for (qw(TERM QUIT INT CHLD));
299 local $0 = $self->{-wq_ident};
300 PublicInbox::DS::sig_setmask($oldset);
301 # ensure we properly exit even if warn() dies:
302 my $end = PublicInbox::OnDestroy->new($$, sub { exit(!!$@) });
303 my $on_destroy = $self->ipc_atfork_child;
304 eval { wq_worker_loop($self) };
305 warn "worker $self->{-wq_ident} PID:$$ died: $@" if $@;
307 undef $end; # trigger exit
309 $self->{-wq_workers}->{$pid} = \undef;
313 # starts workqueue workers if Sereal or Storable is installed
314 sub wq_workers_start {
315 my ($self, $ident, $nr_workers, $oldset) = @_;
316 ($enc && $send_cmd && $recv_cmd && defined($SEQPACKET)) or return;
317 return if $self->{-wq_s1}; # idempotent
318 $self->{-wq_s1} = $self->{-wq_s2} = undef;
319 socketpair($self->{-wq_s1}, $self->{-wq_s2}, AF_UNIX, $SEQPACKET, 0) or
320 die "socketpair: $!";
321 $self->ipc_atfork_prepare;
323 $nr_workers = $WQ_MAX_WORKERS if $nr_workers > $WQ_MAX_WORKERS;
324 my $sigset = $oldset // PublicInbox::DS::block_signals();
325 $self->{-wq_workers} = {};
326 $self->{-wq_ident} = $ident;
327 _wq_worker_start($self, $sigset) for (1..$nr_workers);
328 PublicInbox::DS::sig_setmask($sigset) unless $oldset;
329 $self->{-wq_ppid} = $$;
332 sub wq_worker_incr { # SIGTTIN handler
333 my ($self, $oldset) = @_;
334 $self->{-wq_s2} or return;
335 return if wq_workers($self) >= $WQ_MAX_WORKERS;
336 $self->ipc_atfork_prepare;
337 my $sigset = $oldset // PublicInbox::DS::block_signals();
338 _wq_worker_start($self, $sigset);
339 PublicInbox::DS::sig_setmask($sigset) unless $oldset;
342 sub wq_exit { # wakes up wq_worker_decr_wait
343 send($_[0]->{-wq_s2}, $$, MSG_EOR) // die "$$ send: $!";
347 sub wq_worker_decr { # SIGTTOU handler, kills first idle worker
349 return unless wq_workers($self);
350 my $s2 = $self->{-wq_s2} // die 'BUG: no wq_s2';
351 $self->wq_do('wq_exit', [ $s2, $s2, $s2 ]);
352 # caller must call wq_worker_decr_wait in main loop
355 sub wq_worker_decr_wait {
356 my ($self, $timeout) = @_;
357 return if $self->{-wq_ppid} != $$; # can't reap siblings or parents
358 my $s1 = $self->{-wq_s1} // croak 'BUG: no wq_s1';
359 vec(my $rin = '', fileno($s1), 1) = 1;
360 select(my $rout = $rin, undef, undef, $timeout) or
361 croak 'timed out waiting for wq_exit';
362 recv($s1, my $pid, 64, 0) // croak "recv: $!";
363 my $workers = $self->{-wq_workers} // croak 'BUG: no wq_workers';
364 delete $workers->{$pid} // croak "BUG: PID:$pid invalid";
365 dwaitpid($pid, \&ipc_worker_reap, $self);
368 # set or retrieve number of workers
370 my ($self, $nr) = @_;
371 my $cur = $self->{-wq_workers} or return;
373 while (scalar(keys(%$cur)) > $nr) {
374 $self->wq_worker_decr;
375 $self->wq_worker_decr_wait;
377 $self->wq_worker_incr while scalar(keys(%$cur)) < $nr;
383 my ($self, $nohang) = @_;
384 delete @$self{qw(-wq_s1 -wq_s2)} or return;
385 my $ppid = delete $self->{-wq_ppid} or return;
386 my $workers = delete $self->{-wq_workers} // die 'BUG: no wq_workers';
387 return if $ppid != $$; # can't reap siblings or parents
388 my @pids = map { $_ + 0 } keys %$workers;
390 push @{$self->{"-wq_old_pids.$$"}}, @pids;
392 dwaitpid($_, \&ipc_worker_reap, $self) for @pids;
398 my $pids = $self->{"-wq_old_pids.$$"} or return;
403 my ($self, $sig) = @_;
404 my $workers = $self->{-wq_workers} or return;
405 kill($sig // 'TERM', keys %$workers);
408 sub WQ_MAX_WORKERS { $WQ_MAX_WORKERS }
412 my $ppid = $self->{-wq_ppid};
413 wq_kill($self) if $ppid && $ppid == $$;
416 ipc_worker_stop($self);
419 # Sereal doesn't have dclone
420 sub deep_clone { thaw(freeze($_[-1])) }