[openssl-dev] DTLS retransmission api

[Alfred E. Heggestad]

On 01/06/16 13:58, Matt Caswell wrote:
>
>
> On 01/06/16 11:15, Alfred E. Heggestad wrote:
>> hi,
>>
>> we are using DTLS from OpenSSL to implement DTLS-SRTP in our
>> product (Wire.com) .. The code and implementation works really well
>> and is very robust. We are using OpenSSL version 1.0.2g
>>
>>
>> since our product is deployed globally on mobile data networks,
>> we have quite variable latency and packetloss. The patch below
>> shows my working code, it has an initial retransmit timeout
>> of 400 ms which is incrementing by 10% for every re-trans.
>>
>>
>> obviously this patch cannot make it into the official tree.
>>
>>
>> but I would like to discuss with you guys the option to
>> add some kind of API for:
>>
>> - Setting the initial RTO for DTLS (in milliseconds).
>> - Setting the retransmit policy for DTLS, i.e. should it
>>    double or increment by X for every re-trans.
>
> I think an API for that would be a great idea. Perhaps a callback could
> be used so that you can set exactly the policy you want?
>

Thank you, Matt


I can work on a patch for this, if you guys can help me to define
the API.


I think we only need one CTRL api to set the next re-transmit
interval. then in the application code that calls this:

- DTLSv1_handle_timeout
- DTLSv1_get_timeout


can also call DTLS_set_retrans_interval(400)


>>
>>
>> in addition we have seen the code hit this assert
>> in production:
>>
>>
>>    /*OPENSSL_assert(0);*/ /* XDTLS: want to see if we ever get here */
>>
>>
>> so I would say it should be safe to remove it.
>
> Hmmmmm....the question is why does it get there? It shouldn't.
>

I can try to reproduce this. We have seen that this assert was
executed, when the code was under quite heavy load and lots of traffic.




/alfred

>
> Matt
>
>
>>
>>
>>
>>
>> Best Regards,
>>
>> Alfred E. Heggestad
>> Berlin
>>
>>
>>
>> --
>>
>> diff -Naur openssl-1.0.2g/ssl/d1_lib.c openssl/ssl/d1_lib.c
>> --- openssl-1.0.2g/ssl/d1_lib.c    2016-03-01 14:35:53.000000000 +0100
>> +++ openssl/ssl/d1_lib.c    2016-06-01 10:45:27.000000000 +0200
>> @@ -359,6 +359,8 @@
>>
>>   void dtls1_start_timer(SSL *s)
>>   {
>> +    struct timeval diff;
>> +
>>   #ifndef OPENSSL_NO_SCTP
>>       /* Disable timer for SCTP */
>>       if (BIO_dgram_is_sctp(SSL_get_wbio(s))) {
>> @@ -369,14 +371,17 @@
>>
>>       /* If timer is not set, initialize duration with 1 second */
>>       if (s->d1->next_timeout.tv_sec == 0 && s->d1->next_timeout.tv_usec
>> == 0) {
>> -        s->d1->timeout_duration = 1;
>> +        s->d1->timeout_duration = 0.400;
>>       }
>>
>>       /* Set timeout to current time */
>>       get_current_time(&(s->d1->next_timeout));
>>
>>       /* Add duration to current time */
>> -    s->d1->next_timeout.tv_sec += s->d1->timeout_duration;
>> +    diff.tv_sec  = 0;
>> +    diff.tv_usec = 1000000*s->d1->timeout_duration;
>> +    timeradd(&s->d1->next_timeout, &diff, &s->d1->next_timeout);
>> +
>>       BIO_ctrl(SSL_get_rbio(s), BIO_CTRL_DGRAM_SET_NEXT_TIMEOUT, 0,
>>                &(s->d1->next_timeout));
>>   }
>> @@ -441,7 +446,7 @@
>>
>>   void dtls1_double_timeout(SSL *s)
>>   {
>> -    s->d1->timeout_duration *= 2;
>> +    s->d1->timeout_duration *= 1.10;
>>       if (s->d1->timeout_duration > 60)
>>           s->d1->timeout_duration = 60;
>>       dtls1_start_timer(s);
>> diff -Naur openssl-1.0.2g/ssl/d1_pkt.c openssl/ssl/d1_pkt.c
>> --- openssl-1.0.2g/ssl/d1_pkt.c    2016-03-01 14:35:53.000000000 +0100
>> +++ openssl/ssl/d1_pkt.c    2016-03-08 14:39:44.000000000 +0100
>> @@ -1502,7 +1502,7 @@
>>        * will happen with non blocking IO
>>        */
>>       if (s->s3->wbuf.left != 0) {
>> -        OPENSSL_assert(0);      /* XDTLS: want to see if we ever get
>> here */
>> +        /*OPENSSL_assert(0);*/      /* XDTLS: want to see if we ever
>> get here */
>>           return (ssl3_write_pending(s, type, buf, len));
>>       }
>>
>> diff -Naur openssl-1.0.2g/ssl/dtls1.h openssl/ssl/dtls1.h
>> --- openssl-1.0.2g/ssl/dtls1.h    2016-03-01 14:35:53.000000000 +0100
>> +++ openssl/ssl/dtls1.h    2016-03-08 14:39:44.000000000 +0100
>> @@ -225,8 +225,8 @@
>>        * Indicates when the last handshake msg or heartbeat sent will
>> timeout
>>        */
>>       struct timeval next_timeout;
>> -    /* Timeout duration */
>> -    unsigned short timeout_duration;
>> +    /* Timeout duration in Seconds */
>> +    double timeout_duration;
>>       /*
>>        * storage for Alert/Handshake protocol data received but not yet
>>        * processed by ssl3_read_bytes:
>>
>>