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 # - ipc_do and ipc_worker_* is for a single worker/producer and uses pipes
6 # - wq_do and wq_worker* is for a single producer and multiple workers,
7 # using SOCK_SEQPACKET for work distribution
8 # use ipc_do when you need work done on a certain process
9 # use wq_do when your work can be done on any idle worker
10 package PublicInbox::IPC;
13 use parent qw(Exporter);
15 use PublicInbox::DS qw(dwaitpid);
16 use PublicInbox::Spawn;
17 use PublicInbox::OnDestroy;
18 use PublicInbox::WQWorker;
19 use Socket qw(AF_UNIX MSG_EOR SOCK_STREAM);
20 my $SEQPACKET = eval { Socket::SOCK_SEQPACKET() }; # portable enough?
21 use constant PIPE_BUF => $^O eq 'linux' ? 4096 : POSIX::_POSIX_PIPE_BUF();
22 our @EXPORT_OK = qw(ipc_freeze ipc_thaw);
23 my $WQ_MAX_WORKERS = 4096;
25 # ->imports at BEGIN turns sereal_*_with_object into custom ops on 5.14+
26 # and eliminate method call overhead
29 require Sereal::Encoder;
30 require Sereal::Decoder;
31 Sereal::Encoder->import('sereal_encode_with_object');
32 Sereal::Decoder->import('sereal_decode_with_object');
33 ($enc, $dec) = (Sereal::Encoder->new, Sereal::Decoder->new);
37 if ($enc && $dec) { # should be custom ops
38 *ipc_freeze = sub ($) { sereal_encode_with_object $enc, $_[0] };
39 *ipc_thaw = sub ($) { sereal_decode_with_object $dec, $_[0], my $ret };
41 eval { # some distros have Storable as a separate package from Perl
43 *ipc_freeze = \&Storable::freeze;
44 *ipc_thaw = \&Storable::thaw;
46 } // warn("Storable (part of Perl) missing: $@\n");
49 my $recv_cmd = PublicInbox::Spawn->can('recv_cmd4');
50 my $send_cmd = PublicInbox::Spawn->can('send_cmd4') // do {
51 require PublicInbox::CmdIPC4;
52 $recv_cmd //= PublicInbox::CmdIPC4->can('recv_cmd4');
53 PublicInbox::CmdIPC4->can('send_cmd4');
58 defined(my $len = <$r>) or return;
59 chop($len) eq "\n" or croak "no LF byte in $len";
60 defined(my $n = read($r, my $buf, $len)) or croak "read error: $!";
61 $n == $len or croak "short read: $n != $len";
67 my $buf = ipc_freeze($ref);
68 length($buf) . "\n" . $buf;
73 print $w _pack_rec($ref) or croak "print: $!";
76 sub ipc_return ($$$) {
77 my ($w, $ret, $exc) = @_;
78 _send_rec($w, $exc ? bless(\$exc, 'PublicInbox::IPC::Die') : $ret);
81 sub ipc_worker_loop ($$$) {
82 my ($self, $r_req, $w_res) = @_;
83 my ($rec, $wantarray, $sub, @args);
85 while ($rec = _get_rec($r_req)) {
86 ($wantarray, $sub, @args) = @$rec;
87 # no waiting if client doesn't care,
88 # this is the overwhelmingly likely case
89 if (!defined($wantarray)) {
90 eval { $self->$sub(@args) };
91 warn "$$ die: $@ (from nowait $sub)\n" if $@;
92 } elsif ($wantarray) {
93 my @ret = eval { $self->$sub(@args) };
94 ipc_return($w_res, \@ret, $@);
95 } else { # '' => wantscalar
96 my $ret = eval { $self->$sub(@args) };
97 ipc_return($w_res, \$ret, $@);
102 # starts a worker if Sereal or Storable is installed
103 sub ipc_worker_spawn {
104 my ($self, $ident, $oldset, $fields) = @_;
105 return unless $enc; # no Sereal or Storable
106 return if ($self->{-ipc_ppid} // -1) == $$; # idempotent
107 delete(@$self{qw(-ipc_req -ipc_res -ipc_ppid -ipc_pid)});
108 pipe(my ($r_req, $w_req)) or die "pipe: $!";
109 pipe(my ($r_res, $w_res)) or die "pipe: $!";
110 my $sigset = $oldset // PublicInbox::DS::block_signals();
111 $self->ipc_atfork_prepare;
112 my $seed = rand(0xffffffff);
113 my $pid = fork // die "fork: $!";
116 eval { PublicInbox::DS->Reset };
117 delete @$self{qw(-wq_s1 -wq_s2 -wq_workers -wq_ppid)};
118 $w_req = $r_res = undef;
119 $w_res->autoflush(1);
120 $SIG{$_} = 'IGNORE' for (qw(TERM INT QUIT));
122 PublicInbox::DS::sig_setmask($sigset);
123 # ensure we properly exit even if warn() dies:
124 my $end = PublicInbox::OnDestroy->new($$, sub { exit(!!$@) });
127 local @$self{keys %$fields} = values(%$fields);
128 my $on_destroy = $self->ipc_atfork_child;
130 ipc_worker_loop($self, $r_req, $w_res);
132 warn "worker $ident PID:$$ died: $@\n" if $@;
133 undef $end; # trigger exit
135 PublicInbox::DS::sig_setmask($sigset) unless $oldset;
136 $r_req = $w_res = undef;
137 $w_req->autoflush(1);
138 $self->{-ipc_req} = $w_req;
139 $self->{-ipc_res} = $r_res;
140 $self->{-ipc_ppid} = $$;
141 $self->{-ipc_pid} = $pid;
144 sub ipc_worker_reap { # dwaitpid callback
145 my ($args, $pid) = @_;
147 # TERM(15) is our default exit signal, PIPE(13) is likely w/ pager
149 warn "PID:$pid died with \$?=$?\n" if $s != 15 && $s != 13;
153 my ($self, @args) = @_;
154 my $cb = ref($args[0]) eq 'CODE' ? shift(@args) : \&ipc_worker_reap;
155 my $pids = delete $self->{"-wq_old_pids.$$"} or return;
156 dwaitpid($_, $cb, [$self, @args]) for @$pids;
159 # for base class, override in sub classes
160 sub ipc_atfork_prepare {}
162 sub wq_atexit_child {}
164 sub ipc_atfork_child {
166 my $io = delete($self->{-ipc_atfork_child_close}) or return;
171 # idempotent, can be called regardless of whether worker is active or not
172 sub ipc_worker_stop {
173 my ($self, $args) = @_;
174 my ($pid, $ppid) = delete(@$self{qw(-ipc_pid -ipc_ppid)});
175 my ($w_req, $r_res) = delete(@$self{qw(-ipc_req -ipc_res)});
176 if (!$w_req && !$r_res) {
177 die "unexpected PID:$pid without IPC pipes" if $pid;
180 die 'no PID with IPC pipes' unless $pid;
181 $w_req = $r_res = undef;
183 return if $$ != $ppid;
184 dwaitpid($pid, \&ipc_worker_reap, [$self, $args]);
187 # use this if we have multiple readers reading curl or "pigz -dc"
188 # and writing to the same store
191 require PublicInbox::Lock;
192 $self->{-ipc_lock} //= bless { lock_path => $f }, 'PublicInbox::Lock'
195 sub ipc_async_wait ($$) {
196 my ($self, $max) = @_; # max == -1 to wait for all
197 my $aif = $self->{-async_inflight} or return;
198 my $r_res = $self->{-ipc_res} or die 'BUG: no ipc_res';
199 while (my ($sub, $bytes, $cb, $cb_arg) = splice(@$aif, 0, 4)) {
200 my $ret = _get_rec($r_res) //
201 die "no response on $sub (req.size=$bytes)";
202 $self->{-async_inflight_bytes} -= $bytes;
204 eval { $cb->($cb_arg, $ret) };
205 warn "E: $sub callback error: $@\n" if $@;
206 return if --$max == 0;
210 # call $self->$sub(@args), on a worker if ipc_worker_spawn was used
212 my ($self, $sub, @args) = @_;
213 if (my $w_req = $self->{-ipc_req}) { # run in worker
214 my $ipc_lock = $self->{-ipc_lock};
215 my $lock = $ipc_lock ? $ipc_lock->lock_for_scope : undef;
216 if (defined(wantarray)) {
217 my $r_res = $self->{-ipc_res} or die 'BUG: no ipc_res';
218 ipc_async_wait($self, -1);
219 _send_rec($w_req, [ wantarray, $sub, @args ]);
220 my $ret = _get_rec($r_res) // die "no response on $sub";
221 die $$ret if ref($ret) eq 'PublicInbox::IPC::Die';
222 wantarray ? @$ret : $$ret;
223 } else { # likely, fire-and-forget into pipe
224 _send_rec($w_req, [ undef , $sub, @args ]);
226 } else { # run locally
232 my ($self, $sub, $sub_args, $cb, $cb_arg) = @_;
233 if (my $w_req = $self->{-ipc_req}) { # run in worker
234 my $rec = _pack_rec([ 1, $sub, @$sub_args ]);
235 my $cur_bytes = \($self->{-async_inflight_bytes} //= 0);
236 while (($$cur_bytes + length($rec)) > PIPE_BUF) {
237 ipc_async_wait($self, 1);
239 my $ipc_lock = $self->{-ipc_lock};
240 my $lock = $ipc_lock ? $ipc_lock->lock_for_scope : undef;
241 print $w_req $rec or croak "print: $!";
242 $$cur_bytes += length($rec);
243 push @{$self->{-async_inflight}},
244 $sub, length($rec), $cb, $cb_arg;
246 my $ret = [ eval { $self->$sub(@$sub_args) } ];
248 $ret = ( bless(\$exc, 'PublicInbox::IPC::Die') );
250 eval { $cb->($cb_arg, $ret) };
251 warn "E: $sub callback error: $@\n" if $@;
255 # needed when there's multiple IPC workers and the parent forking
256 # causes newer siblings to inherit older siblings sockets
257 sub ipc_sibling_atfork_child {
259 my ($pid, undef) = delete(@$self{qw(-ipc_pid -ipc_ppid)});
260 delete(@$self{qw(-ipc_req -ipc_res)});
261 $pid == $$ and die "BUG: $$ ipc_atfork_child called on itself";
265 my ($self, $s2, $len, $full_stream) = @_;
266 my @fds = $recv_cmd->($s2, my $buf, $len);
267 return if scalar(@fds) && !defined($fds[0]);
268 my $n = length($buf) or return 0;
271 if (open(my $cmdfh, '+<&=', $fd)) {
272 $self->{$nfd++} = $cmdfh;
273 $cmdfh->autoflush(1);
275 die "$$ open(+<&=$fd) (FD:$nfd): $!";
278 while ($full_stream && $n < $len) {
279 my $r = sysread($s2, $buf, $len - $n, $n) // croak "read: $!";
280 croak "read EOF after $n/$len bytes" if $r == 0;
283 # Sereal dies on truncated data, Storable returns undef
284 my $args = ipc_thaw($buf) // die "thaw error on buffer of size: $n";
286 my $sub = shift @$args;
287 eval { $self->$sub(@$args) };
288 warn "$$ wq_worker: $@" if $@;
289 delete @$self{0..($nfd-1)};
293 sub wq_worker_loop ($) {
295 my $wqw = PublicInbox::WQWorker->new($self);
296 PublicInbox::DS->SetPostLoopCallback(sub { $wqw->{sock} });
297 PublicInbox::DS->EventLoop;
298 PublicInbox::DS->Reset;
301 sub do_sock_stream { # via wq_do, for big requests
302 my ($self, $len) = @_;
303 recv_and_run($self, delete $self->{0}, $len, 1);
306 sub wq_do { # always async
307 my ($self, $sub, $ios, @args) = @_;
308 if (my $s1 = $self->{-wq_s1}) { # run in worker
309 my $fds = [ map { fileno($_) } @$ios ];
310 my $buf = ipc_freeze([$sub, @args]);
311 my $n = $send_cmd->($s1, $fds, $buf, MSG_EOR);
312 return if defined($n); # likely
313 croak "sendmsg: $! (check RLIMIT_NOFILE)" if $!{ETOOMANYREFS};
314 croak "sendmsg: $!" if !$!{EMSGSIZE};
315 socketpair(my $r, my $w, AF_UNIX, SOCK_STREAM, 0) or
316 croak "socketpair: $!";
317 $n = $send_cmd->($s1, [ fileno($r) ],
318 ipc_freeze(['do_sock_stream', length($buf)]),
319 MSG_EOR) // croak "sendmsg: $!";
321 $n = $send_cmd->($w, $fds, $buf, 0) // croak "sendmsg: $!";
322 while ($n < length($buf)) {
323 my $x = syswrite($w, $buf, length($buf) - $n, $n) //
324 croak "syswrite: $!";
325 croak "syswrite wrote 0 bytes" if $x == 0;
329 @$self{0..$#$ios} = @$ios;
330 eval { $self->$sub(@args) };
331 warn "wq_do: $@" if $@;
332 delete @$self{0..$#$ios}; # don't close
336 sub _wq_worker_start ($$$) {
337 my ($self, $oldset, $fields) = @_;
338 my $seed = rand(0xffffffff);
339 my $pid = fork // die "fork: $!";
342 eval { PublicInbox::DS->Reset };
343 delete @$self{qw(-wq_s1 -wq_workers -wq_ppid)};
344 $SIG{$_} = 'IGNORE' for (qw(PIPE));
345 $SIG{$_} = 'DEFAULT' for (qw(TTOU TTIN TERM QUIT INT CHLD));
346 local $0 = $self->{-wq_ident};
347 PublicInbox::DS::sig_setmask($oldset);
348 # ensure we properly exit even if warn() dies:
349 my $end = PublicInbox::OnDestroy->new($$, sub { exit(!!$@) });
352 local @$self{keys %$fields} = values(%$fields);
353 my $on_destroy = $self->ipc_atfork_child;
355 wq_worker_loop($self);
357 warn "worker $self->{-wq_ident} PID:$$ died: $@" if $@;
358 undef $end; # trigger exit
360 $self->{-wq_workers}->{$pid} = \undef;
364 # starts workqueue workers if Sereal or Storable is installed
365 sub wq_workers_start {
366 my ($self, $ident, $nr_workers, $oldset, $fields) = @_;
367 ($enc && $send_cmd && $recv_cmd && defined($SEQPACKET)) or return;
368 return if $self->{-wq_s1}; # idempotent
369 $self->{-wq_s1} = $self->{-wq_s2} = undef;
370 socketpair($self->{-wq_s1}, $self->{-wq_s2}, AF_UNIX, $SEQPACKET, 0) or
371 die "socketpair: $!";
372 $self->ipc_atfork_prepare;
374 $nr_workers = $WQ_MAX_WORKERS if $nr_workers > $WQ_MAX_WORKERS;
375 my $sigset = $oldset // PublicInbox::DS::block_signals();
376 $self->{-wq_workers} = {};
377 $self->{-wq_ident} = $ident;
378 _wq_worker_start($self, $sigset, $fields) for (1..$nr_workers);
379 PublicInbox::DS::sig_setmask($sigset) unless $oldset;
380 $self->{-wq_ppid} = $$;
383 sub wq_worker_incr { # SIGTTIN handler
384 my ($self, $oldset, $fields) = @_;
385 $self->{-wq_s2} or return;
386 return if wq_workers($self) >= $WQ_MAX_WORKERS;
387 $self->ipc_atfork_prepare;
388 my $sigset = $oldset // PublicInbox::DS::block_signals();
389 _wq_worker_start($self, $sigset, $fields);
390 PublicInbox::DS::sig_setmask($sigset) unless $oldset;
393 sub wq_exit { # wakes up wq_worker_decr_wait
394 send($_[0]->{-wq_s2}, $$, MSG_EOR) // die "$$ send: $!";
398 sub wq_worker_decr { # SIGTTOU handler, kills first idle worker
400 return unless wq_workers($self);
401 my $s2 = $self->{-wq_s2} // die 'BUG: no wq_s2';
402 $self->wq_do('wq_exit', [ $s2, $s2, $s2 ]);
403 # caller must call wq_worker_decr_wait in main loop
406 sub wq_worker_decr_wait {
407 my ($self, $timeout) = @_;
408 return if $self->{-wq_ppid} != $$; # can't reap siblings or parents
409 my $s1 = $self->{-wq_s1} // croak 'BUG: no wq_s1';
410 vec(my $rin = '', fileno($s1), 1) = 1;
411 select(my $rout = $rin, undef, undef, $timeout) or
412 croak 'timed out waiting for wq_exit';
413 recv($s1, my $pid, 64, 0) // croak "recv: $!";
414 my $workers = $self->{-wq_workers} // croak 'BUG: no wq_workers';
415 delete $workers->{$pid} // croak "BUG: PID:$pid invalid";
416 dwaitpid($pid, \&ipc_worker_reap, $self);
419 # set or retrieve number of workers
421 my ($self, $nr) = @_;
422 my $cur = $self->{-wq_workers} or return;
424 while (scalar(keys(%$cur)) > $nr) {
425 $self->wq_worker_decr;
426 $self->wq_worker_decr_wait;
428 $self->wq_worker_incr while scalar(keys(%$cur)) < $nr;
434 my ($self, $nohang) = @_;
435 delete @$self{qw(-wq_s1 -wq_s2)} or return;
436 my $ppid = delete $self->{-wq_ppid} or return;
437 my $workers = delete $self->{-wq_workers} // die 'BUG: no wq_workers';
438 return if $ppid != $$; # can't reap siblings or parents
439 my @pids = map { $_ + 0 } keys %$workers;
441 push @{$self->{"-wq_old_pids.$$"}}, @pids;
443 dwaitpid($_, \&ipc_worker_reap, $self) for @pids;
449 my $pids = $self->{"-wq_old_pids.$$"} or return;
454 my ($self, $sig) = @_;
455 my $workers = $self->{-wq_workers} or return;
456 kill($sig // 'TERM', keys %$workers);
459 sub WQ_MAX_WORKERS { $WQ_MAX_WORKERS }
463 my $ppid = $self->{-wq_ppid};
464 wq_kill($self) if $ppid && $ppid == $$;
467 ipc_worker_stop($self);
470 sub detect_nproc () {
471 # _SC_NPROCESSORS_ONLN = 84 on both Linux glibc and musl
472 return POSIX::sysconf(84) if $^O eq 'linux';
473 return POSIX::sysconf(58) if $^O eq 'freebsd';
476 # getconf(1) is POSIX, but *NPROCESSORS* vars are not
477 for (qw(_NPROCESSORS_ONLN NPROCESSORS_ONLN)) {
478 `getconf $_ 2>/dev/null` =~ /^(\d+)$/ and return $1;
480 for my $nproc (qw(nproc gnproc)) { # GNU coreutils nproc
481 `$nproc 2>/dev/null` =~ /^(\d+)$/ and return $1;
484 # should we bother with `sysctl hw.ncpu`? Those only give
485 # us total processor count, not online processor count.