1 # Copyright (C) 2016-2021 all contributors <meta@public-inbox.org>
2 # License: AGPL-3.0+ <https://www.gnu.org/licenses/agpl-3.0.txt>
4 # Generic PSGI server for convenience. It aims to provide
5 # a consistent experience for public-inbox admins so they don't have
6 # to learn different ways to admin both NNTP and HTTP components.
7 # There's nothing which depends on public-inbox, here.
8 # Each instance of this class represents a HTTP client socket
11 # httpd: PublicInbox::HTTPD ref
12 # env: PSGI env hashref
13 # input_left: bytes left to read in request body (e.g. POST/PUT)
14 # remote_addr: remote IP address as a string (e.g. "127.0.0.1")
15 # remote_port: peer port
16 # forward: response body object, response to ->getline + ->close
17 # alive: HTTP keepalive state:
18 # 0: drop connection when done
19 # 1: keep connection when done
20 # 2: keep connection, chunk responses
21 package PublicInbox::HTTP;
23 use parent qw(PublicInbox::DS);
25 use Plack::HTTPParser qw(parse_http_request); # XS or pure Perl
27 use HTTP::Status qw(status_message);
28 use HTTP::Date qw(time2str);
29 use PublicInbox::DS qw(msg_more);
30 use PublicInbox::Syscall qw(EPOLLIN EPOLLONESHOT);
31 use PublicInbox::Tmpfile;
33 CHUNK_START => -1, # [a-f0-9]+\r\n
34 CHUNK_END => -2, # \r\n
35 CHUNK_ZEND => -3, # \r\n
41 sub process_pipelineq () {
45 next unless $_->{sock};
50 # Use the same configuration parameter as git since this is primarily
51 # a slow-client sponge for git-http-backend
52 # TODO: support per-respository http.maxRequestBuffer somehow...
53 our $MAX_REQUEST_BUFFER = $ENV{GIT_HTTP_MAX_REQUEST_BUFFER} ||
56 open(my $null_io, '<', '/dev/null') or die "failed to open /dev/null: $!";
61 $now == $prev ? $http_date : ($http_date = time2str($prev = $now));
65 my ($class, $sock, $addr, $httpd) = @_;
66 my $self = bless { httpd => $httpd }, $class;
69 if ($sock->can('accept_SSL') && !$sock->accept_SSL) {
70 return CORE::close($sock) if $! != EAGAIN;
71 $ev = PublicInbox::TLS::epollbit() or return CORE::close($sock);
72 $wbuf = [ \&PublicInbox::DS::accept_tls_step ];
74 $self->{wbuf} = $wbuf if $wbuf;
75 ($self->{remote_addr}, $self->{remote_port}) =
76 PublicInbox::Daemon::host_with_port($addr);
77 $self->SUPER::new($sock, $ev | EPOLLONESHOT);
80 sub event_step { # called by PublicInbox::DS
83 return unless $self->flush_write && $self->{sock};
85 # only read more requests if we've drained the write buffer,
86 # otherwise we can be buffering infinitely w/o backpressure
88 return read_input($self) if ref($self->{env});
89 my $rbuf = $self->{rbuf} // (\(my $x = ''));
90 $self->do_read($rbuf, 8192, length($$rbuf)) or return;
91 rbuf_process($self, $rbuf);
95 my ($self, $rbuf) = @_;
96 $rbuf //= $self->{rbuf} // (\(my $x = ''));
98 my %env = %{$self->{httpd}->{env}}; # full hash copy
99 my $r = parse_http_request($$rbuf, \%env);
101 # We do not support Trailers in chunked requests, for now
102 # (they are rarely-used and git (as of 2.7.2) does not use them)
103 if ($r == -1 || $env{HTTP_TRAILER} ||
104 # this length-check is necessary for PURE_PERL=1:
105 ($r == -2 && length($$rbuf) > 0x4000)) {
106 return quit($self, 400);
108 if ($r < 0) { # incomplete
109 $self->rbuf_idle($rbuf);
110 return $self->requeue;
112 $$rbuf = substr($$rbuf, $r);
113 my $len = input_prepare($self, \%env);
114 defined $len or return write_err($self, undef); # EMFILE/ENFILE
116 $len ? read_input($self, $rbuf) : app_dispatch($self, undef, $rbuf);
119 sub read_input ($;$) {
120 my ($self, $rbuf) = @_;
121 $rbuf //= $self->{rbuf} // (\(my $x = ''));
122 my $env = $self->{env};
123 return read_input_chunked($self, $rbuf) if env_chunked($env);
125 # env->{CONTENT_LENGTH} (identity)
126 my $len = delete $self->{input_left};
127 my $input = $env->{'psgi.input'};
131 my $w = syswrite($input, $$rbuf, $len);
132 return write_err($self, $len) unless $w;
134 die "BUG: $len < 0 (w=$w)" if $len < 0;
135 if ($len == 0) { # next request may be pipelined
136 $$rbuf = substr($$rbuf, $w);
141 $self->do_read($rbuf, 8192) or return recv_err($self, $len);
142 # continue looping if $r > 0;
144 app_dispatch($self, $input, $rbuf);
148 my ($self, $input, $rbuf) = @_;
149 $self->rbuf_idle($rbuf);
150 my $env = $self->{env};
151 $self->{env} = undef; # for exists() check in ->busy
152 $env->{REMOTE_ADDR} = $self->{remote_addr};
153 $env->{REMOTE_PORT} = $self->{remote_port};
154 if (defined(my $host = $env->{HTTP_HOST})) {
155 $host =~ s/:([0-9]+)\z// and $env->{SERVER_PORT} = $1;
156 $env->{SERVER_NAME} = $host;
158 if (defined $input) {
159 sysseek($input, 0, SEEK_SET) or
160 die "BUG: psgi.input seek failed: $!";
162 # note: NOT $self->{sock}, we want our close (+ PublicInbox::DS::close),
163 # to do proper cleanup:
164 $env->{'psgix.io'} = $self; # for ->close or async_pass
165 my $res = Plack::Util::run_app($self->{httpd}->{app}, $env);
167 if (ref($res) eq 'CODE') {
168 $res->(sub { response_write($self, $env, $_[0]) });
170 response_write($self, $env, $res);
174 err($self, "response_write error: $@");
179 sub response_header_write {
180 my ($self, $env, $res) = @_;
181 my $proto = $env->{SERVER_PROTOCOL} or return; # HTTP/0.9 :P
182 my $status = $res->[0];
183 my $h = "$proto $status " . status_message($status) . "\r\n";
185 my $headers = $res->[1];
187 for (my $i = 0; $i < @$headers; $i += 2) {
188 my $k = $headers->[$i];
189 my $v = $headers->[$i + 1];
190 next if $k =~ /\A(?:Connection|Date)\z/i;
192 $len = $v if $k =~ /\AContent-Length\z/i;
193 if ($k =~ /\ATransfer-Encoding\z/i && $v =~ /\bchunked\b/i) {
199 my $conn = $env->{HTTP_CONNECTION} || '';
200 my $term = defined($len) || $chunked;
201 my $prot_persist = ($proto eq 'HTTP/1.1') && ($conn !~ /\bclose\b/i);
203 if (!$term && $prot_persist) { # auto-chunk
204 $chunked = $alive = 2;
205 $h .= "Transfer-Encoding: chunked\r\n";
206 # no need for "Connection: keep-alive" with HTTP/1.1
207 } elsif ($term && ($prot_persist || ($conn =~ /\bkeep-alive\b/i))) {
209 $h .= "Connection: keep-alive\r\n";
212 $h .= "Connection: close\r\n";
214 $h .= 'Date: ' . http_date() . "\r\n\r\n";
216 if (($len || $chunked) && $env->{REQUEST_METHOD} ne 'HEAD') {
224 # middlewares such as Deflater may write empty strings
225 sub chunked_write ($$) {
227 return if $_[1] eq '';
228 msg_more($self, sprintf("%x\r\n", length($_[1])));
229 msg_more($self, $_[1]);
231 # use $self->write(\"\n\n") if you care about real-time
232 # streaming responses, public-inbox WWW does not.
233 msg_more($self, "\r\n");
236 sub identity_write ($$) {
238 $self->write(\($_[1])) if $_[1] ne '';
241 sub next_request ($) {
244 # avoid recursion for pipelined requests
245 PublicInbox::DS::requeue(\&process_pipelineq) if !@$pipelineq;
246 push @$pipelineq, $self;
247 } else { # wait for next request
253 my ($self, $alive) = @_;
254 delete $self->{env}; # we're no longer busy
255 $self->write(\"0\r\n\r\n") if $alive == 2;
256 $self->write($alive ? \&next_request : \&close);
261 my $forward = $self->{forward};
263 # limit our own running time for fairness with other
264 # clients and to avoid buffering too much:
271 # may close in PublicInbox::DS::write
272 if ($self->{alive} == 2) {
273 chunked_write($self, $buf);
275 identity_write($self, $buf);
280 my $new_size = push(@{$self->{wbuf}}, \&getline_pull);
282 # wbuf may be populated by {chunked,identity}_write()
283 # above, no need to rearm if so:
284 $self->requeue if $new_size == 1;
288 err($self, "response ->getline error: $@");
292 if (delete $self->{forward}) {
293 eval { $forward->close };
295 err($self, "response ->close error: $@");
296 $self->close; # idempotent
299 response_done($self, delete $self->{alive});
303 my ($self, $env, $res) = @_;
304 my $alive = response_header_write($self, $env, $res);
305 if (defined(my $body = $res->[2])) {
306 if (ref $body eq 'ARRAY') {
308 chunked_write($self, $_) for @$body;
310 identity_write($self, $_) for @$body;
312 response_done($self, $alive);
314 $self->{forward} = $body;
315 $self->{alive} = $alive;
316 getline_pull($self); # kick-off!
318 # these are returned to the calling application:
319 } elsif ($alive == 2) {
320 bless [ $self, $alive ], 'PublicInbox::HTTP::Chunked';
322 bless [ $self, $alive ], 'PublicInbox::HTTP::Identity';
327 my ($self, $env) = @_;
330 # rfc 7230 3.3.2, 3.3.3,: favor Transfer-Encoding over Content-Length
331 my $hte = $env->{HTTP_TRANSFER_ENCODING};
333 # rfc7230 3.3.3, point 3 says only chunked is accepted
334 # as the final encoding. Since neither public-inbox-httpd,
335 # git-http-backend, or our WWW-related code uses "gzip",
336 # "deflate" or "compress" as the Transfer-Encoding, we'll
338 return quit($self, 400) if $hte !~ /\Achunked\z/i;
341 $input = tmpfile('http.input', $self->{sock});
343 $len = $env->{CONTENT_LENGTH};
346 return quit($self, 400) if $len !~ /\A[0-9]+\z/;
347 return quit($self, 413) if $len > $MAX_REQUEST_BUFFER;
348 $input = $len ? tmpfile('http.input', $self->{sock})
355 # TODO: expire idle clients on ENFILE / EMFILE
356 $env->{'psgi.input'} = $input // return;
358 $self->{input_left} = $len || 0;
361 sub env_chunked { ($_[0]->{HTTP_TRANSFER_ENCODING} // '') =~ /\Achunked\z/i }
364 eval { $_[0]->{httpd}->{env}->{'psgi.errors'}->print($_[1]."\n") };
368 my ($self, $len) = @_;
369 my $msg = $! || '(zero write)';
370 $msg .= " ($len bytes remaining)" if defined $len;
371 err($self, "error buffering to input: $msg");
376 my ($self, $len) = @_;
377 if ($! == EAGAIN) { # epoll/kevent watch already set by do_read
378 $self->{input_left} = $len;
380 err($self, "error reading input: $! ($len bytes remaining)");
384 sub read_input_chunked { # unlikely...
385 my ($self, $rbuf) = @_;
386 $rbuf //= $self->{rbuf} // (\(my $x = ''));
387 my $input = $self->{env}->{'psgi.input'};
388 my $len = delete $self->{input_left};
390 while (1) { # chunk start
391 if ($len == CHUNK_ZEND) {
392 $$rbuf =~ s/\A\r\n//s and
393 return app_dispatch($self, $input, $rbuf);
395 return quit($self, 400) if length($$rbuf) > 2;
397 if ($len == CHUNK_END) {
398 if ($$rbuf =~ s/\A\r\n//s) {
400 } elsif (length($$rbuf) > 2) {
401 return quit($self, 400);
404 if ($len == CHUNK_START) {
405 if ($$rbuf =~ s/\A([a-f0-9]+).*?\r\n//i) {
407 if (($len + -s $input) > $MAX_REQUEST_BUFFER) {
408 return quit($self, 413);
410 } elsif (length($$rbuf) > CHUNK_MAX_HDR) {
411 return quit($self, 400);
413 # will break from loop since $len >= 0
416 if ($len < 0) { # chunk header is trickled, read more
417 $self->do_read($rbuf, 8192, length($$rbuf)) or
418 return recv_err($self, $len);
419 # (implicit) goto chunk_start if $r > 0;
421 $len = CHUNK_ZEND if $len == 0;
423 # drain the current chunk
426 my $w = syswrite($input, $$rbuf, $len);
427 return write_err($self, "$len chunk") if !$w;
430 # we may have leftover data to parse
432 $$rbuf = substr($$rbuf, $w);
435 die "BUG: len < 0: $len";
441 # read more of current chunk
442 $self->do_read($rbuf, 8192) or
443 return recv_err($self, $len);
450 my ($self, $status) = @_;
451 my $h = "HTTP/1.1 $status " . status_message($status) . "\r\n\r\n";
454 undef; # input_prepare expects this
459 if (my $forward = delete $self->{forward}) {
460 eval { $forward->close };
461 err($self, "forward ->close error: $@") if $@;
463 $self->SUPER::close; # PublicInbox::DS::close
466 # for graceful shutdown in PublicInbox::Daemon:
469 ($self->{rbuf} || exists($self->{env}) || $self->{wbuf});
472 # runs $cb on the next iteration of the event loop at earliest
474 my ($self, $cb) = @_;
475 return unless exists $self->{sock};
476 $self->requeue if 1 == push(@{$self->{wbuf}}, $cb);
479 # Chunked and Identity packages are used for writing responses.
480 # They may be exposed to the PSGI application when the PSGI app
481 # returns a CODE ref for "push"-based responses
482 package PublicInbox::HTTP::Chunked;
486 # ([$http], $buf) = @_;
487 PublicInbox::HTTP::chunked_write($_[0]->[0], $_[1])
491 # $_[0] = [$http, $alive]
492 PublicInbox::HTTP::response_done(@{$_[0]});
495 package PublicInbox::HTTP::Identity;
497 our @ISA = qw(PublicInbox::HTTP::Chunked);
500 # ([$http], $buf) = @_;
501 PublicInbox::HTTP::identity_write($_[0]->[0], $_[1]);