[PATCH 0/2] Toeplitz hashing

[PATCH 0/2] Toeplitz hashing

[Tom Herbert]

These patches introduce software implementation of Toeplitz hashing.
The first use case is in inet_ehashfn and inet6_ehashfn using Toeplitz
for just IPv4 and IPv6 TCP connection lookup.

In some follow up patches we can move to using the HW hash itself to
do connection lookup (possibly eliminating SW hash computation on RX.
These will do:
  1) Convert skb_get_rxhash (used for RPS steering) to use Toeplitz. 
  2) Allow TCP connection lookup to use rxhash if the hash if the
     hash is marked as being proper Toeplitz hash.
  3) Allow drivers to mark a hash value as being proper Toeplitz
     of a TCP packet

There's also the possibility of calculating the Toeplitz hash for a
connection in order to deduce which RSS queue the connection maps
to, and thus allow a poor man's aRFS (idea from Jesse Brandeburg).

Re: [PATCH 0/2] Toeplitz hashing

[Rick Jones]

On 09/23/2013 03:41 PM, Tom Herbert wrote:
> There's also the possibility of calculating the Toeplitz hash for a
> connection in order to deduce which RSS queue the connection maps
> to, and thus allow a poor man's aRFS (idea from Jesse Brandeburg).

Perhaps the answer is in the meaning of the leading 'a' in aRFS but 
doesn't one still depend on the thread of execution handling only one 
flow for that to work effectively?  Or would flows be assigned to 
threads of execution based on the Toeplitz hash of the connections?

rick jones