WEBrick: prevent response splitting and header injection
This is a follow up to d9d4a28.
The commit prevented CRLR, but did not address an isolated CR or an
isolated LF.
Co-Authored-By: NARUSE, Yui <naruse@airemix.jp>
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@67819 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
Loop with String#scan without creating substrings
Create the substrings necessary parts only, instead of cutting the
rest of the buffer. Also removed a useless, probable typo, regexp.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@67818 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
webrick: support Proc objects as body responses
* lib/webrick/httpresponse.rb (send_body): call send_body_proc
(send_body_proc): new method
(class ChunkedWrapper): new class
* test/webrick/test_httpresponse.rb (test_send_body_proc): new test
(test_send_body_proc_chunked): ditto
[Feature #855]
webrick/httpresponse: IO.copy_stream for regular files
Remove the redundant _send_file method since its functionality
is unnecessary with IO.copy_stream. IO.copy_stream also allows
the use of sendfile under some OSes to speed up copies to
non-TLS sockets.
Testing with "curl >/dev/null" and "ruby -run -e httpd" to
read a 1G file over Linux loopback reveals a reduction from
around ~0.770 to ~0.490 seconds on the client side.
* lib/webrick/httpresponse.rb (send_body_io): use IO.copy_stream
(_send_file): remove
[Feature #14237]
webrick: use IO.copy_stream for single range response
This is also compatible with range responses generated
by Rack::File (tested with rack 2.0.3).
* lib/webrick/httpresponse.rb (send_body_io): use Content-Range
* lib/webrick/httpservlet/filehandler.rb (make_partial_content):
use File object for the single range case
* test/webrick/test_filehandler.rb (get_res_body): use send_body
to test result
test/webrick/test_filehandler.rb: stricter multipart range test
We need to ensure we generate compatibile output in
the face of future changes
* test/webrick/test_filehandler.rb (test_make_partial_content):
check response body
webrick: quiet warning for multi-part ranges
Content-Length is ignored by WEBrick::HTTPResponse even if we
calculate it, so instead we chunk responses to HTTP/1.1 clients
and terminate HTTP/1.0 connections.
* lib/webrick/httpservlet/filehandler.rb (make_partial_content):
quiet warning
webrick/httpresponse: make ChunkedWrapper copy_stream-compatible
The .write method needs to return the number of bytes written
to avoid confusing IO.copy_stream.
* lib/webrick/httpresponse.rb (ChunkedWrapper#write): return bytes written
(ChunkedWrapper#<<): return self
webrick: use IO.copy_stream for multipart response
Use the new Proc response body feature to generate a multipart
range response dynamically. We use a flat array to minimize
object overhead as much as possible; as many ranges may fit
into an HTTP request header.
* lib/webrick/httpservlet/filehandler.rb (multipart_body): new method
(make_partial_content): use multipart_body
get rid of test error/failure on Windows introduced at r62955
* lib/webrick/httpresponse.rb (send_body_io): use seek if NotImplementedError
is raised in IO.copy_stream with offset.
* lib/webrick/httpservlet/filehandler.rb (multipart_body): ditto.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@63012 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
webrick/httprequest: limit request headers size
We use the same 112 KB limit started (AFAIK) by Mongrel, Thin,
and Puma to prevent malicious users from using up all the memory
with a single request. This also limits the damage done by
excessive ranges in multipart Range: requests.
Due to the way we rely on IO#gets and the desire to keep
the code simple, the actual maximum header may be 4093 bytes
larger than 112 KB, but we're splitting hairs at that point.
* lib/webrick/httprequest.rb: define MAX_HEADER_LENGTH
(read_header): raise when headers exceed max length
webrick/httpservlet/cgihandler: reduce memory use
WEBrick::HTTPRequest#body can be passed a block to process the
body in chunks. Use this feature to avoid building a giant
string in memory.
* lib/webrick/httpservlet/cgihandler.rb (do_GET):
avoid reading entire request body into memory
(do_POST is aliased to do_GET, so it handles bodies)
webrick/httprequest: raise correct exception
"BadRequest" alone does not resolve correctly, it is in the
HTTPStatus namespace.
* lib/webrick/httprequest.rb (read_chunked): use correct exception
* test/webrick/test_httpserver.rb (test_eof_in_chunk): new test
webrick/httprequest: use InputBufferSize for chunked requests
While WEBrick::HTTPRequest#body provides a Proc interface
for streaming large request bodies, clients must not force
the server to use an excessively large chunk size.
* lib/webrick/httprequest.rb (read_chunk_size): limit each
read and block.call to :InputBufferSize in config.
* test/webrick/test_httpserver.rb (test_big_chunks): new test
webrick: add test for Digest auth-int
No changes to the actual code, this is a new test for
a feature for which no tests existed. I don't understand
the Digest authentication code well at all, but this is
necessary for the subsequent change.
* test/webrick/test_httpauth.rb (test_digest_auth_int): new test
(credentials_for_request): support bodies with POST
webrick/httpauth/digestauth: stream req.body
WARNING! WARNING! WARNING! LIKELY BROKEN CHANGE
Pass a proc to WEBrick::HTTPRequest#body to avoid reading a
potentially large request body into memory during
authentication.
WARNING! this will break apps completely which want to do
something with the body besides calculating the MD5 digest
of it.
Also, keep in mind that probably nobody uses "auth-int".
Servers such as Apache, lighttpd, nginx don't seem to
support it; nor does curl when using POST/PUT bodies;
and we didn't have tests for it until now...
* lib/webrick/httpauth/digestauth.rb (_authenticate): stream req.body
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@63004 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
webrick: prevent response splitting and header injection
Original patch by tenderlove (with minor style adjustments).
* lib/webrick/httpresponse.rb (send_header): call check_header
(check_header): raise on embedded CRLF in header value
* test/webrick/test_httpresponse.rb
(test_prevent_response_splitting_headers): new test
* (test_prevent_response_splitting_cookie_headers): ditto
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@63002 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
webrick: favor .write over << method
This will make the next change to use IO.copy_stream
easier-to-read. When we can drop Ruby 2.4 support in a few
years, this will allow us to use writev(2) with multiple
arguments for headers and chunked responses.
* lib/webrick/cgi.rb (write): new wrapper method
lib/webrick/httpresponse.rb: (send_header): use socket.write
(send_body_io): ditto
(send_body_string): ditto
(send_body_proc): ditto
(_write_data): ditto
(ChunkedWrapper#write): ditto
(_send_file): ditto
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@62988 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
webrick: avoid unnecessary IO#sync= call
Sockets and pipes are always created with FMODE_SYNC flag
already set (otherwise many things would be broken).
* lib/webrick/server.rb (accept_client): remove unnecessary
IO#sync= call
webrick: do not hang acceptor on slow TLS connections
OpenSSL::SSL::SSLSocket#accept may block indefinitely on clients
which negotiate the TCP connection, but fail (or are slow) to
negotiate the subsequent TLS handshake. This prevents the
multi-threaded WEBrick server from accepting other connections.
Since the TLS handshake (via OpenSSL::SSL::SSLSocket#accept)
consists of normal read/write traffic over TCP, handle it in the
per-client thread, instead.
Furthermore, using non-blocking accept() is useful for non-TLS
sockets anyways because spurious wakeups are possible from
select(2).
* lib/webrick/server.rb (accept_client): use TCPServer#accept_nonblock
and remove OpenSSL::SSL::SSLSocket#accept call
* lib/webrick/server.rb (start_thread): call OpenSSL::SSL::SSLSocket#accept
* test/webrick/test_ssl_server.rb (test_slow_connect): new test
[ruby-core:83221] [Bug #14005]
webrick: fix up r60172
By making the socket non-blocking in r60172, TLS/SSL negotiation
via the SSL_accept function must handle non-blocking sockets
properly and retry on SSL_ERROR_WANT_READ/SSL_ERROR_WANT_WRITE.
OpenSSL::SSL::SSLSocket#accept cannot do that properly with a
non-blocking socket, so it must use non-blocking logic of
OpenSSL::SSL::SSLSocket#accept_nonblock.
Thanks to MSP-Greg (Greg L) for finding this.
* lib/webrick/server.rb (start_thread): use SSL_accept properly
with non-blocking socket.
[Bug #14013] [Bug #14005]
webrick: fix up r60172 and revert r60189
Thanks to MSP-Greg (Greg L) for helping with this.
* lib/webrick/server.rb (start_thread): ignore ECONNRESET, ECONNABORTED,
EPROTO, and EINVAL on TLS negotiation errors the same way they
were ignored before r60172 in the accept_client method of the
main acceptor thread.
[Bug #14013] [Bug #14005]
webrick: fix up r60172 and r60208
Thanks to MSP-Greg (Greg L) for helping with this.
* lib/webrick/server.rb (start_thread): fix non-local return
introduced in r60208
webrick: fix up r60172 and r60210
Thanks to MSP-Greg (Greg L) for helping with this.
* lib/webrick/server.rb (start_thread): properly fix non-local return
introduced in r60208 and r60210
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@61239 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
webrick: compile RE correctly for beginning and end match
Using ^ and $ in regexps means we can accidentally get fooled
by "%0a" in HTTP request paths being decoded to newline
characters. Use \A and \z to match beginning and end-of-string
respectively, instead.
Thanks to mame and hsbt for reporting.
* lib/webrick/httpserver.rb (MountTable#compile):
use \A and \z instead of ^ and $
* lib/webrick/httpserver.rb (MountTable#normalize): use \z instead of $
* test/webrick/test_httpserver.rb (test_cntrl_in_path): new test
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@61238 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
lib/webrick/log.rb: sanitize any type of logs
It had failed to sanitize some type of exception messages. Reported and
patched by Yusuke Endoh (mame) at https://hackerone.com/reports/223363
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/branches/ruby_2_4@59898 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
There is no need to call Utils.getservername when the :ServerName option is
specified, so delay Utils.getservername until needed to avoid unnecessary
DNS lookups. [ruby-core:78492] [Bug #13007]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@57014 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
Symbol proc is less code and avoids confusion from variable
naming.
* lib/webrick/server.rb (shutdown): use symbol proc
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@56889 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* IOError does not happen even if another thread closes io
* Use symbol proc
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@56874 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
Process.daemon exists since Ruby 1.9.1 and does most of what we need.
* lib/webrick/server.rb (Daemon.start): simplify
[Misc #12937]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@56869 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
The monotonic clock is immune to system time changes and
a better option for implementing timing comparisons.
* lib/webrick/utils.rb (TimeoutHandler): use monotonic clock
(watch): ditto
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@56417 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
disable reverse lookup by default. [ruby-core:45514] [Feature #6559]
Socket.do_not_reverse_lookup is true by default but WEBrick
overwrote it.
patch by Eric Hodel [ruby-core:45527]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@55881 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* lib/webrick/httprequest.rb (setup_forwarded_info): Use the first
value in X-Forwarded-Proto, if header contains multiple comma
separated values. Some middlewares may add these values to the
list, not replacing. [Fix GH-1386]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@55484 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
to invoke ssl server with command line.
[fix GH-1329] Patch by @kerlin
* test/webrick/test_ssl_server.rb: Added test for GH-1329
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@54941 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
override the body more easily. [fix GH-1307]
* test/webrick/test_httpresponse.rb: ditto.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@54346 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* lib/webrick/utils.rb (WEBrick::Utils::TimeoutHandler#initialize):
use WEBrick::Utils::TimeoutHandler::Thread, which is ignored by
LeakChecker#find_threads, instead of ::Thread to get rid of
thread leak checker. since this TimeoutHandler is resident
during tests because of Singleton, it waits for the next timeout
if it has any schedules. in the case of nested timeouts, inner
timeout does not cancel outer timeouts and then those schedules
still remain.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@53201 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
potential deadlocks, Queue is used to tell update of @timeout_info
instead of sleep and wakeup. [Bug #11742] [ruby-dev:49387]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@53192 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
When you change this to true, you may need to add more tests.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@53141 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
TimeoutMutex only when accessing @timeout_info for avoiding
potential deadlock. [Bug #11742] [ruby-dev:49387]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@53134 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
TimeoutMutex should be acquired when accessing @timeout_info.
To avoid deadlock, interrupt() calls are delayed.
Due to the mutex, it is safe to treat ary without ary.dup.
[Bug #11742] [ruby-dev:49387]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@53130 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* lib/webrick/utils.rb (WEBrick::Utils::TimeoutHandler#register):
notify the handler thread of new timeout registration.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@51235 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* lib/webrick/utils.rb (WEBrick::Utils::TimeoutHandler#initialize):
make sleep intervals adaptive than fixed period intervals.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@51234 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
* lib/webrick/server.rb (WEBrick::GenericServer#start): flush
shutdown pipe.
* lib/webrick/server.rb (WEBrick::GenericServer#stop): request the
server to stop immediately by sending data via shutdown pipe.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@51231 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
IO#nonblock= and IO#close_on_exec= methods are simpler-to-use
and potentially more portable to for future OSes.
IO#nonblock= and IO#close_on_exec= are also smart enough to avoid
redundantly setting flags so a syscall may be avoided.
These methods could probably be removed entirely and inlined, but
it's unclear if there is 3rd-party code which relies on them.
* lib/webrick/utils.rb (set_non_blocking): use IO#nonblock=
* (set_close_on_exec): use IO#close_on_exec=
[Feature #11136]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@50523 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
Sockets are close-on-exec by default since Ruby 2.0, so it
is redundant to set it again.
* lib/webrick/server.rb (accept_client): avoid redundant fcntl call
[Feature #11137]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@50519 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
In case a process encounters high-numbered FDs, this allows
consistent performance on systems with ppoll support.
[ruby-core:35572]
* ext/socket/lib/socket.rb (connect_nonblock): use IO#wait_writable
* lib/drb/drb.rb (DRB::DRbTCPSocket#alive?): use IO#wait_readable
* lib/webrick/httpserver.rb (run): ditto
* lib/resolv.rb (request): ditto for single socket case
[ruby-core:68943] [Feature #11081]
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@50432 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
notify readability on the read side of the pipe.
write_nonblock() is not usable for pipe on Windows.
(cleanup_shutdown_pipe): Rescue IOError for @shutdown_pipe_w.close.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@48354 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
to avoid race condition.
(cleanup_shutdown_pipe): New private method.
(cleanup_listener): Extracted from shutdown method.
Call this method from start method to avoid race condition.
git-svn-id: svn+ssh://ci.ruby-lang.org/ruby/trunk@48353 b2dd03c8-39d4-4d8f-98ff-823fe69b080e
