gcf2: improve error handling and do not ->fail on wbuf

For historical reasons, both Danga::Socket::write and
PublicInbox::DS::write will return 0 when data is buffered;
so Gcf2Client must not call ->fail when DS::write returns 0.

We'll also improve robustness by recreating the entire
Gcf2Client object if it does die for other reasons, instead of
risking mismatched fields due to deferred close.

We also need to ensure we only get one EPOLLERR wakeup and
issue EPOLL_CTL_DEL if ->event_step is triggered by a dying
Gcf2 process, so always register the FD with EPOLLONESHOT.

diff --git a/lib/PublicInbox/Daemon.pm b/lib/PublicInbox/Daemon.pm
index 1520f8f20e563850fad496df741bdd64dc797cff..155707e1df49c4d0c17dd61551f19acc0538e2c5 100644 (file)
--- a/lib/PublicInbox/Daemon.pm
+++ b/lib/PublicInbox/Daemon.pm
@@ -654,14 +654,8 @@ sub run ($$$;$) {
        my $af_default = $default =~ /:8080\z/ ? 'httpready' : undef;
        my $for_destroy = daemonize();
 
-       # this wastes a bit of memory for non-PublicInbox::WWW -httpd users
-       # oh well...
-       eval {
-               require PublicInbox::Gcf2;
-               require PublicInbox::Gcf2Client;
-       };
-       local $PublicInbox::GitAsyncCat::GCF2C =
-                               PublicInbox::Gcf2Client::new() if !$@;
+       # localize GCF2C for tests:
+       local $PublicInbox::GitAsyncCat::GCF2C;
 
        daemon_loop($refresh, $post_accept, $tlsd, $af_default);
        PublicInbox::DS->Reset;

diff --git a/lib/PublicInbox/Gcf2Client.pm b/lib/PublicInbox/Gcf2Client.pm
index 42ff1bf33d5ef2898f6c4400f72db9bfa5d7b2cd..ab486de5895ad9e5190d3438b7adb029414dfa90 100644 (file)
--- a/lib/PublicInbox/Gcf2Client.pm
+++ b/lib/PublicInbox/Gcf2Client.pm
@@ -8,27 +8,28 @@ use parent qw(PublicInbox::DS);
 use PublicInbox::Git;
 use PublicInbox::Spawn qw(popen_rd);
 use IO::Handle ();
-use PublicInbox::Syscall qw(EPOLLONESHOT EPOLLOUT);
+use PublicInbox::Syscall qw(EPOLLONESHOT);
 # fields:
 #      async_cat => GitAsyncCat ref (read-only pipe)
 #      sock => writable pipe to Gcf2::loop
 
-sub new { bless($_[0] // {}, __PACKAGE__) }
 
-sub gcf2c_begin ($) {
-       my ($self) = @_;
+sub new  {
+       my ($rdr) = @_;
+       my $self = bless {}, __PACKAGE__;
        # ensure the child process has the same @INC we do:
        my $env = { PERL5LIB => join(':', @INC) };
        my ($out_r, $out_w);
        pipe($out_r, $out_w) or die "pipe failed: $!";
-       my $rdr = { 0 => $out_r, 2 => $self->{2} };
+       $rdr //= {};
+       $rdr->{0} = $out_r;
        my $cmd = [$^X, qw[-MPublicInbox::Gcf2 -e PublicInbox::Gcf2::loop()]];
        @$self{qw(in pid)} = popen_rd($cmd, $env, $rdr);
        fcntl($out_w, 1031, 4096) if $^O eq 'linux'; # 1031: F_SETPIPE_SZ
        $out_w->autoflush(1);
        $out_w->blocking(0);
-       $self->SUPER::new($out_w, 0); # EPOLL_CTL_ADD (a bit wasteful :x)
        $self->{inflight} = [];
+       $self->SUPER::new($out_w, EPOLLONESHOT); # detect errors once
 }
 
 sub fail {
@@ -39,22 +40,28 @@ sub fail {
 
 sub cat_async ($$$;$) {
        my ($self, $req, $cb, $arg) = @_;
-       my $inflight = $self->{inflight} // gcf2c_begin($self);
+       my $inflight = $self->{inflight};
 
-       # rare, I hope:
+       # {wbuf} is rare, I hope:
        cat_async_step($self, $inflight) if $self->{wbuf};
 
-       $self->write(\"$req\n") or $self->fail("gcf2c write: $!");
+       if (!$self->write(\"$req\n")) {
+               $self->fail("gcf2c write: $!") if !$self->{sock};
+       }
        push @$inflight, $req, $cb, $arg;
 }
 
 # ensure PublicInbox::Git::cat_async_step never calls cat_async_retry
 sub alternates_changed {}
 
-no warnings 'once';
-
 # this is the write-only end of a pipe, DS->EventLoop will call this
-*event_step = \&PublicInbox::DS::flush_write;
+sub event_step {
+       my ($self) = @_;
+       $self->flush_write;
+       $self->close if !$self->{in}; # process died
+}
+
+no warnings 'once';
 
 # used by GitAsyncCat
 *cat_async_step = \&PublicInbox::Git::cat_async_step;

diff --git a/lib/PublicInbox/GitAsyncCat.pm b/lib/PublicInbox/GitAsyncCat.pm
index b9dbe0ccf2251a0d434d6260880dc0d240bb431b..be51f673317eae824b85cc9f28f61aa9a221ad8a 100644 (file)
--- a/lib/PublicInbox/GitAsyncCat.pm
+++ b/lib/PublicInbox/GitAsyncCat.pm
@@ -53,14 +53,21 @@ sub event_step {
 
 sub git_async_cat ($$$$) {
        my ($git, $oid, $cb, $arg) = @_;
-       my $gitish = $GCF2C;
-       if ($gitish) {
+       my $gitish = $GCF2C //= eval {
+               require PublicInbox::Gcf2;
+               require PublicInbox::Gcf2Client;
+               PublicInbox::Gcf2Client::new();
+       } // 0; # 0: do not retry if libgit2 or Inline::C are missing
+       if ($gitish) { # Gcf2 active, {inflight} may be unset due to errors
+               $GCF2C->{inflight} or
+                       $gitish = $GCF2C = PublicInbox::Gcf2Client::new();
                $oid .= " $git->{git_dir}";
        } else {
                $gitish = $git;
        }
        $gitish->cat_async($oid, $cb, $arg);
        $gitish->{async_cat} //= do {
+               # read-only end of pipe (Gcf2Client is write-only end)
                my $self = bless { gitish => $gitish }, __PACKAGE__;
                $self->SUPER::new($gitish->{in}, EPOLLIN|EPOLLET);
                \undef; # this is a true ref()