buggy_openssl_with_fullduplex/README.md

45 lines
1.7 KiB
Markdown
Raw Normal View History

2013-09-09 13:17:17 +01:00
buggy_openssl_with_fullduplex
=============================
Toy code which shows problems with non-blocking, fullduplex I/O & renegotiation in OpenSSL
2013-09-18 13:11:46 +01:00
2013-10-08 23:21:47 +01:00
What is it:
This code shows that it is not possible to use SSL_write() and SSL_read() functions in the same
way as ::send() and ::recv().
The main difference between SSL_write/read functions and send/recv system calls is that
SSL_write function may in fact read data from the socket in some situations (similary
SSL_read may need to write data to the socket).
2013-09-18 13:11:46 +01:00
How it works:
Client & Server:
- it has two threads - sender & receiver
- writes and reads are mutexed
Client:
- I/O is blocking (but can be non-blocking)
Server:
- I/O is non-blocking
- each thread runs it's own select()
1. After client & server are connected (and SSL handshake done) client sender
thread starts sending first message (in a loop).
2. When server receives first query it starts sending string EXCHANGE_STRING for
SEND_ITERATIONS number of times. So now we have 4 threads that are sending
and receiving traffic at the same time ( 2 send/receive threads on each
server and client side )
3. When client receives RENEG_INIT_LEN number of characters it starts
renegotiation ( if other one is not pending ). Bug starts to occure here
BUG:
Client side: client starts to report SSL_ERROR_SYSCALL
Server side: server reports SSL_ERROR_WANT_READ when receive function is called
TCP:
In TCP exchange we can see that transfer between client & server is OK until
client sends "Client Hello" packet. This packet is sent when SSL_renegotiate
2013-10-08 23:21:47 +01:00
is called