Skip to contentSkip to conversation

proto_register() - Justification for requesting slab allocation

DISCUSSION Eus 1 participant 1 message Apr 8, 2008

lore
Share & tools
lore

Eus

author
#0 lore raw

Hi Ho!

Currently I am trying to implement a new type of socket in Linux kernel 2.6.21.5. I am really curious about this function:
int proto_register(struct proto *prot, int alloc_slab)

I have investigated the source code and knew that, if alloc_slab is set to a non-zero integer, kmem_cache_create() will create a memory slab for prot->slab. At the end, when a socket needs to be created and sk_alloc() is invoked to create the socket object, if prot->slab has been initialized with kmem_cache_create(), sk_alloc() will simply create the socket object in the slab with kmem_cache_alloc. Otherwise, sk_alloc() will create the socket object in the ordinary way with kmalloc().

IMO, kmem_cache_alloc() should be less expensive than kmalloc() and, therefore, it is a good thing to request slab allocation when invoking proto_register(). But, from all networking protocols that invoke proto_register(), 50% of them, most of them are data link protocols, does not request slab allocation. The rest that request slab allocation mainly is network layer protocols. That is why I wonder whether or not there is an advantage of using kmalloc() over using kmem_cache_alloc().

A friend of mine said that those that do not request slab allocation do so because they are rarely used. But, I disagree because, although they are rarely used, once they are used, they are used heavily, for example AF_PACKET, so that it is a good idea to request slab allocation.

Therefore, what is the justification for requesting slab allocation or not?

Thank you very much.

Best regards,
Eus

      ____________________________________________________________________________________

You rock. That's why Blockbuster's offering you one month of Blockbuster Total Access, No Cost.
http://tc.deals.yahoo.com/tc/blockbuster/text5.com