Re: [PLEASE-TESTME] Zerocopy networking patch, 2.4.0-1

[Manfred Spraul]

sct wrote:
> We've already got measurements showing how insane this is. Raw IO 
> requests, plus internal pagebuf contiguous requests from XFS, have to 
> get broken down into page-sized chunks by the current ll_rw_block() 
> API, only to get reassembled by the make_request code. It's 
> *enormous* overhead, and the kiobuf-based disk IO code demonstrates 
> this clearly. 

Stephen, I see one big difference between ll_rw_block and the proposed
tcp_sendpage():
You must allocate and initialize a complete buffer head for each page
you want to read, and then you pass the array of buffer heads to
ll_rw_block with one function call.
I'm certain the overhead is the allocation/initialization/freeing of the
buffer heads, not the function call.

AFAICS the proposed tcp_sendpage interface is the other way around:
you need one function call for each page, but no memory
allocation/setup. The memory is allocated internally by the tcp_sendpage
implementation, and it merges requests when possible, thus for a 9000
byte jumbopacket you'd need 3 function calls to tcp_sendpage(MSG_MORE),
but only one skb is allocated and set up.

Ingo is that correct?

--
	Manfred