Hello,

On Thu, 1 Sep 2011, Ward, David - 0663 - MITLL wrote:

quoted

	Not sure if adding size as parameter to flow_hash_code
is better approach. May be flow_cache_lookup needs to
determine size from family that can be used for flow_hash_code,
flow_key_compare and the memcpy(&fle->key, key, sizeof(*key))
after fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC).

Makes sense to me.  However should we just replace flow_key_compare with
memcmp then, since the assumptions about constant size and alignment will no
longer apply?  Or should there be a separate flow_key_compare function for
each family, and have all of the flowi* structures become
__attribute__((__aligned__(BITS_PER_LONG/8))) ?

	I don't know this code well but I guess memcmp is
not preferred. IMHO, as the callers provide per-family
structures and we do not want to change that, these
structures must be aligned to long type as required by
flow_compare_t and jhash2 (at least u32) usage. The second
option is to use memcmp and jhash instead of jhash2 to
avoid such alignment but I guess other developers will
oppose it. More opinions are needed here.

Regards

--
Julian Anastasov [off-list ref]