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
40 # Use the same configuration parameter as git since this is primarily
41 # a slow-client sponge for git-http-backend
42 # TODO: support per-respository http.maxRequestBuffer somehow...
43 our $MAX_REQUEST_BUFFER = $ENV{GIT_HTTP_MAX_REQUEST_BUFFER} ||
46 open(my $null_io, '<', '/dev/null') or die "failed to open /dev/null: $!";
51 $now == $prev ? $http_date : ($http_date = time2str($prev = $now));
55 my ($class, $sock, $addr, $httpd) = @_;
56 my $self = bless { httpd => $httpd }, $class;
59 if ($sock->can('accept_SSL') && !$sock->accept_SSL) {
60 return CORE::close($sock) if $! != EAGAIN;
61 $ev = PublicInbox::TLS::epollbit() or return CORE::close($sock);
62 $wbuf = [ \&PublicInbox::DS::accept_tls_step ];
64 $self->{wbuf} = $wbuf if $wbuf;
65 ($self->{remote_addr}, $self->{remote_port}) =
66 PublicInbox::Daemon::host_with_port($addr);
67 $self->SUPER::new($sock, $ev | EPOLLONESHOT);
70 sub event_step { # called by PublicInbox::DS
73 return unless $self->flush_write && $self->{sock};
75 # only read more requests if we've drained the write buffer,
76 # otherwise we can be buffering infinitely w/o backpressure
78 return read_input($self) if ref($self->{env});
80 my $rbuf = $self->{rbuf} // (\(my $x = ''));
81 my %env = %{$self->{httpd}->{env}}; # full hash copy
83 while (($r = parse_http_request($$rbuf, \%env)) < 0) {
84 # We do not support Trailers in chunked requests, for
85 # now (they are rarely-used and git (as of 2.7.2) does
87 # this length-check is necessary for PURE_PERL=1:
88 if ($r == -1 || $env{HTTP_TRAILER} ||
89 ($r == -2 && length($$rbuf) > 0x4000)) {
90 return quit($self, 400);
92 $self->do_read($rbuf, 8192, length($$rbuf)) or return;
94 return quit($self, 400) if grep(/\s/, keys %env); # stop smugglers
95 $$rbuf = substr($$rbuf, $r);
96 my $len = input_prepare($self, \%env) //
97 return write_err($self, undef); # EMFILE/ENFILE
99 $len ? read_input($self, $rbuf) : app_dispatch($self, undef, $rbuf);
102 sub read_input ($;$) {
103 my ($self, $rbuf) = @_;
104 $rbuf //= $self->{rbuf} // (\(my $x = ''));
105 my $env = $self->{env};
106 return read_input_chunked($self, $rbuf) if env_chunked($env);
108 # env->{CONTENT_LENGTH} (identity)
109 my $len = delete $self->{input_left};
110 my $input = $env->{'psgi.input'};
114 my $w = syswrite($input, $$rbuf, $len);
115 return write_err($self, $len) unless $w;
117 die "BUG: $len < 0 (w=$w)" if $len < 0;
118 if ($len == 0) { # next request may be pipelined
119 $$rbuf = substr($$rbuf, $w);
124 $self->do_read($rbuf, 8192) or return recv_err($self, $len);
125 # continue looping if $r > 0;
127 app_dispatch($self, $input, $rbuf);
131 my ($self, $input, $rbuf) = @_;
132 $self->rbuf_idle($rbuf);
133 my $env = $self->{env};
134 $self->{env} = undef; # for exists() check in ->busy
135 $env->{REMOTE_ADDR} = $self->{remote_addr};
136 $env->{REMOTE_PORT} = $self->{remote_port};
137 if (defined(my $host = $env->{HTTP_HOST})) {
138 $host =~ s/:([0-9]+)\z// and $env->{SERVER_PORT} = $1;
139 $env->{SERVER_NAME} = $host;
141 if (defined $input) {
142 sysseek($input, 0, SEEK_SET) or
143 die "BUG: psgi.input seek failed: $!";
145 # note: NOT $self->{sock}, we want our close (+ PublicInbox::DS::close),
146 # to do proper cleanup:
147 $env->{'psgix.io'} = $self; # for ->close or async_pass
148 my $res = Plack::Util::run_app($self->{httpd}->{app}, $env);
150 if (ref($res) eq 'CODE') {
151 $res->(sub { response_write($self, $env, $_[0]) });
153 response_write($self, $env, $res);
157 warn "response_write error: $@";
162 sub response_header_write {
163 my ($self, $env, $res) = @_;
164 my $proto = $env->{SERVER_PROTOCOL} or return; # HTTP/0.9 :P
165 my $status = $res->[0];
166 my $h = "$proto $status " . status_message($status) . "\r\n";
168 my $headers = $res->[1];
170 for (my $i = 0; $i < @$headers; $i += 2) {
171 my $k = $headers->[$i];
172 my $v = $headers->[$i + 1];
173 next if $k =~ /\A(?:Connection|Date)\z/i;
175 $len = $v if $k =~ /\AContent-Length\z/i;
176 if ($k =~ /\ATransfer-Encoding\z/i && $v =~ /\bchunked\b/i) {
182 my $conn = $env->{HTTP_CONNECTION} || '';
183 my $term = defined($len) || $chunked;
184 my $prot_persist = ($proto eq 'HTTP/1.1') && ($conn !~ /\bclose\b/i);
186 if (!$term && $prot_persist) { # auto-chunk
187 $chunked = $alive = 2;
188 $alive = 3 if $env->{REQUEST_METHOD} eq 'HEAD';
189 $h .= "Transfer-Encoding: chunked\r\n";
190 # no need for "Connection: keep-alive" with HTTP/1.1
191 } elsif ($term && ($prot_persist || ($conn =~ /\bkeep-alive\b/i))) {
193 $h .= "Connection: keep-alive\r\n";
196 $h .= "Connection: close\r\n";
198 $h .= 'Date: ' . http_date() . "\r\n\r\n";
200 if (($len || $chunked) && $env->{REQUEST_METHOD} ne 'HEAD') {
208 # middlewares such as Deflater may write empty strings
209 sub chunked_write ($$) {
211 return if $_[1] eq '';
212 msg_more($self, sprintf("%x\r\n", length($_[1])));
213 msg_more($self, $_[1]);
215 # use $self->write(\"\n\n") if you care about real-time
216 # streaming responses, public-inbox WWW does not.
217 msg_more($self, "\r\n");
220 sub identity_write ($$) {
222 $self->write(\($_[1])) if $_[1] ne '';
226 my ($self, $alive) = @_;
227 delete $self->{env}; # we're no longer busy
228 # HEAD requests set $alive = 3 so we don't send "0\r\n\r\n";
229 $self->write(\"0\r\n\r\n") if $alive == 2;
230 $self->write($alive ? $self->can('requeue') : \&close);
235 my $forward = $self->{forward};
237 # limit our own running time for fairness with other
238 # clients and to avoid buffering too much:
245 # may close in PublicInbox::DS::write
246 if ($self->{alive} == 2) {
247 chunked_write($self, $buf);
249 identity_write($self, $buf);
254 my $new_size = push(@{$self->{wbuf}}, \&getline_pull);
256 # wbuf may be populated by {chunked,identity}_write()
257 # above, no need to rearm if so:
258 $self->requeue if $new_size == 1;
262 warn "response ->getline error: $@";
266 if (delete $self->{forward}) {
267 eval { $forward->close };
269 warn "response ->close error: $@";
270 $self->close; # idempotent
273 response_done($self, delete $self->{alive});
277 my ($self, $env, $res) = @_;
278 my $alive = response_header_write($self, $env, $res);
279 if (defined(my $body = $res->[2])) {
280 if (ref $body eq 'ARRAY') {
282 chunked_write($self, $_) for @$body;
284 identity_write($self, $_) for @$body;
286 response_done($self, $alive);
288 $self->{forward} = $body;
289 $self->{alive} = $alive;
290 getline_pull($self); # kick-off!
292 # these are returned to the calling application:
293 } elsif ($alive >= 2) {
294 bless [ $self, $alive ], 'PublicInbox::HTTP::Chunked';
296 bless [ $self, $alive ], 'PublicInbox::HTTP::Identity';
301 my ($self, $env) = @_;
304 # rfc 7230 3.3.2, 3.3.3,: favor Transfer-Encoding over Content-Length
305 my $hte = $env->{HTTP_TRANSFER_ENCODING};
307 # rfc7230 3.3.3, point 3 says only chunked is accepted
308 # as the final encoding. Since neither public-inbox-httpd,
309 # git-http-backend, or our WWW-related code uses "gzip",
310 # "deflate" or "compress" as the Transfer-Encoding, we'll
312 return quit($self, 400) if $hte !~ /\Achunked\z/i;
315 $input = tmpfile('http.input', $self->{sock});
317 $len = $env->{CONTENT_LENGTH};
320 return quit($self, 400) if $len !~ /\A[0-9]+\z/;
321 return quit($self, 413) if $len > $MAX_REQUEST_BUFFER;
322 $input = $len ? tmpfile('http.input', $self->{sock})
329 # TODO: expire idle clients on ENFILE / EMFILE
330 $env->{'psgi.input'} = $input // return;
332 $self->{input_left} = $len || 0;
335 sub env_chunked { ($_[0]->{HTTP_TRANSFER_ENCODING} // '') =~ /\Achunked\z/i }
338 my ($self, $len) = @_;
339 my $msg = $! || '(zero write)';
340 $msg .= " ($len bytes remaining)" if defined $len;
341 warn "error buffering to input: $msg";
346 my ($self, $len) = @_;
347 if ($! == EAGAIN) { # epoll/kevent watch already set by do_read
348 $self->{input_left} = $len;
350 warn "error reading input: $! ($len bytes remaining)";
354 sub read_input_chunked { # unlikely...
355 my ($self, $rbuf) = @_;
356 $rbuf //= $self->{rbuf} // (\(my $x = ''));
357 my $input = $self->{env}->{'psgi.input'};
358 my $len = delete $self->{input_left};
360 while (1) { # chunk start
361 if ($len == CHUNK_ZEND) {
362 $$rbuf =~ s/\A\r\n//s and
363 return app_dispatch($self, $input, $rbuf);
365 return quit($self, 400) if length($$rbuf) > 2;
367 if ($len == CHUNK_END) {
368 if ($$rbuf =~ s/\A\r\n//s) {
370 } elsif (length($$rbuf) > 2) {
371 return quit($self, 400);
374 if ($len == CHUNK_START) {
375 if ($$rbuf =~ s/\A([a-f0-9]+).*?\r\n//i) {
377 if (($len + -s $input) > $MAX_REQUEST_BUFFER) {
378 return quit($self, 413);
380 } elsif (length($$rbuf) > CHUNK_MAX_HDR) {
381 return quit($self, 400);
383 # will break from loop since $len >= 0
386 if ($len < 0) { # chunk header is trickled, read more
387 $self->do_read($rbuf, 8192, length($$rbuf)) or
388 return recv_err($self, $len);
389 # (implicit) goto chunk_start if $r > 0;
391 $len = CHUNK_ZEND if $len == 0;
393 # drain the current chunk
396 my $w = syswrite($input, $$rbuf, $len);
397 return write_err($self, "$len chunk") if !$w;
400 # we may have leftover data to parse
402 $$rbuf = substr($$rbuf, $w);
405 die "BUG: len < 0: $len";
411 # read more of current chunk
412 $self->do_read($rbuf, 8192) or
413 return recv_err($self, $len);
420 my ($self, $status) = @_;
421 my $h = "HTTP/1.1 $status " . status_message($status) . "\r\n\r\n";
424 undef; # input_prepare expects this
429 if (my $forward = delete $self->{forward}) {
430 eval { $forward->close };
431 warn "forward ->close error: $@" if $@;
433 $self->SUPER::close; # PublicInbox::DS::close
436 sub busy { # for graceful shutdown in PublicInbox::Daemon:
438 defined($self->{rbuf}) || exists($self->{env}) || defined($self->{wbuf})
441 # runs $cb on the next iteration of the event loop at earliest
443 my ($self, $cb) = @_;
444 return unless exists $self->{sock};
445 $self->requeue if 1 == push(@{$self->{wbuf}}, $cb);
448 # Chunked and Identity packages are used for writing responses.
449 # They may be exposed to the PSGI application when the PSGI app
450 # returns a CODE ref for "push"-based responses
451 package PublicInbox::HTTP::Chunked;
455 # ([$http], $buf) = @_;
456 PublicInbox::HTTP::chunked_write($_[0]->[0], $_[1])
460 # $_[0] = [$http, $alive]
461 PublicInbox::HTTP::response_done(@{$_[0]});
464 package PublicInbox::HTTP::Identity;
466 our @ISA = qw(PublicInbox::HTTP::Chunked);
469 # ([$http], $buf) = @_;
470 PublicInbox::HTTP::identity_write($_[0]->[0], $_[1]);