cvs commit: src/sys/netinet ip_id.c

[Andre Oppermann]

Robert Watson wrote:
> 
> On Thu, 7 Feb 2008, Bruce M. Simpson wrote:
> 
>> Mike Silbersack wrote:
>>>    Note that this commit does not yet enable random IP ID
>>>   generation by default.  There are still some concerns
>>>   that doing so will adversely affect performance.
>>
>> Random ip_id generation also breaks MANET multicast as it's currently 
>> specified -- the ip_id field gets used for histogram based duplicate 
>> packet detection. However, after my inveigling, Fred Templin at Boeing 
>> brought encapsulation back into the proposal on the table.
>>
>> So some way to turn this stuff off on a per-interface basis is 
>> probably going to be needed later on for mobile stuff...
> 
> We also have much more general problems with our ip_id code -- using a 
> global IP ID counter or random generator breaks down as the packet rate 
> goes up, causing problems for NFS over UDP and other UDP applications 
> with large packet sizes.  We need to adopt some sort of host-based 
> approach, presumably with a fallback to global when we don't want to 
> commit state storage for a host (i.e., for RST's/SYN cookies, or even 
> TCP generally).  There's a comment in the Linux IP ID code regarding 
> PPTP also expecting IP IPs to behave in a more historical way as well.

A sufficiently good hash algorithm should be able to replace storage
allocation and tracking per host.  This way we have an IPID sequence
space per hash bucket where we can guarantee a full cycle.

-- 
Andre