During writing some code I discovered some behaviour of httplib. When we
connect to host, which doesn’t respond with status line, but it just
sending data, httplib may consume more and more memory, becouce when we
execute
h = httplib.HTTPConnection(‘host’)
h.conect()
h.request(‘GET’, ‘/’)
r = h.getresponse()
httplib tries to read one line from host. If host doesn’t send new line
character (‘\n’), httplib reads more and more data. On my tests httplib
could consume all of 4GB of memory and the python process was killed by
oom_killer.
The resolution is to limit maximum amount of data read on getting
response. I have performed some test:
I received 3438293 from hosts located in the network. The longest valid
response line is
HTTP/1.1 500 ( The specified Secure Sockets Layer (SSL) port is not
allowed. ISA Server is not configured to allow SSL requests from this
port. Most Web browsers use port 443 for SSL requests. )\r\n
and it has 197 characters.
In RFC2616 in section 6.1 we have:
“The first line of a Response message is the Status-Line, consisting of
the protocol version followed by a numeric status code and its
associated textual phrase, with each element separated by SP characters.
No CR or LF is allowed except in the final CRLF sequence.
Status-Line = HTTP-Version SP Status-Code SP Reason-Phrase CRLF
(..)The Reason-Phrase is intended to give a short textual description of
the Status-Code.”
So limiting maximum status line length to 256 characters is a solution
of this problem. It doesn’t break compatibility withc RFC 2616.

My patch was written originally on python2.4, but I’ve tested it on
python2.6:
[ms@host python2.6]$ patch --dry-run -i /home/ms/httplib.patch
patching file httplib.py
Hunk #1 succeeded at 209 (offset 54 lines).

I've also check patch against code in svn tree:
wget http://svn.python.org/projects/python/trunk/Lib/httplib.py
patch -p0 -i httplib.patch --dry-run
patching file httplib.py
Hunk #1 succeeded at 209 (offset 54 lines).
Hunk #2 succeeded at 303 (offset 10 lines).

Sumar, to get this moved forward could you please provide a unit test.

Attached is a unit test which tests the issue.
Unfortunately, since it uses the resource module to limit memory to a workable size, it will only work on Unix.

The given patch appears to fix the issue well.

I think this should be taken as a security issue (even if a rather odd one) since a malicious http server could be set up in place of the normal one and crash any http python clients that connect to it.

Eg:
Run: dd if=/dev/zero bs=10M count=1000 | nc -l 8888
And then:

import httplib
h = httplib.HTTPConnection('localhost', 8888)
h.connect()
h.request('GET', '/')
r = h.getresponse()

This should cause python to use up all the memory available.

A py3k patch against revision 87228.

First, I don't think the resource module needs to be used here. Second, I don't see why getcode() would return 200. If no valid response was received then some kind of error should certainly be raised, shouldn't it?

By the way, looking at the code, readline() without any parameter is used all over http.client, so fixing only this one use case doesn't really make sense.

That's true. Near the bottom of the code, it says:

# The status-line parsing code calls readline(), which normally
# get the HTTP status line. For a 0.9 response, however, this is
# actually the first line of the body!

Limiting the length of the status line would break 0.9 responses so maybe this issue should be closed?

That's true. Near the bottom of the code, it says:

The status-line parsing code calls readline(), which normally

get the HTTP status line. For a 0.9 response, however, this is

actually the first line of the body!

Limiting the length of the status line would break 0.9 responses so maybe this issue should be closed?

Well, the HTTP 1.0 RFC was filed in 1996 and HTTP 1.1 is most commonly
used today. I don't think we need to support 0.9 anymore. I'll open a
separate issue for ripping off 0.9 support, though.

I just read the whole discussion and it seems that code was in place so that client can tolerant of a BAD HTTP 0.9 Server response.
http://www.w3.org/Protocols/HTTP/OldServers.html

Given that bpo-10711 talks about removing HTTP/0.9 support (+1 to that), this issue will become obsolete. I too support removing HTTP/0.9. There are hardly any advantages in keeping it around.

Well, removing 0.9 support doesn't make this obsolete, does it?

On Thu, Dec 16, 2010 at 01:18:30PM +0000, Antoine Pitrou wrote:

Well, removing 0.9 support doesn't make this obsolete, does it?

It does. Doesn't it? Because I saw in your patch that you fall back on
HTTP 1.0 behaviour when the server does not return a status line and
in which case a Exception will be raise and this issue won't be
observed.

It does. Doesn't it? Because I saw in your patch that you fall back on
HTTP 1.0 behaviour when the server does not return a status line and
in which case a Exception will be raise and this issue won't be
observed.

I don't think you understood the issue here. Calling readline() without
a maximum length means the process memory potentially explodes, if the
server sends gigabytes of data without a single "\n".

On Thu, Dec 16, 2010 at 02:02:10PM +0000, Antoine Pitrou wrote:

I don't think you understood the issue here. Calling readline() without
a maximum length means the process memory potentially explodes, if the
server sends gigabytes of data without a single "\n".

Yeah, I seem to have misunderstood the issue. Even if the response wa
s an *invalid* one but it was huge data without \n, the readline call
would just explode.

• reading chunked response is doing a readline call too.

Both this need to be addressed by having a limit on reading.

I thought readline() is being called only when parsing headers which
should almost always have CRLF (or at least LF) and thought valid
responses always start with headers.

Now that 0.9 client support has been removed, this can proceed (at least for 3.2).

Here is a patch limiting line length everywhere in http.client, + tests (it also affects http.server since the header parsing routine is shared).

In the morning, I had a comment on the patch wondering why read _MAXLENGH + 1 and then check for len of header > _MAXLENGH. Instead of just reading _MAXLENGH (and if the length matched rejecting). ( Looks like it did not go through).

I think that either way is okay. I am taking the privilege of committing the patch. Fixed for py3k in 87373. So it is be available in the next beta.

Shall merge the changes to other codelines.

Partially backported in r87382 (3.1) and r87383 (2.7). Not everything could be merged in because of HTTP 0.9 support and (in 2.7) a slightly different architecture. Thank you.