Re: tcp_sendpage and page allocation lifetime vs. iscsi

[Avi Kivity <avi@argo.co.il>]

David S. Miller wrote:

>On Mon, 25 Apr 2005 22:11:16 +0300
>Avi Kivity <avi@argo.co.il> wrote:
>
>  
>
>>you need a completion to tell you when your buffer has been sent. you 
>>can use the kiocb parameter to tcp_sendmsg, as it has a completion. 
>>however, tcp_sendmsg does not appear to use it.
>>
>>in effect, you need tcp aio, but the mainline kernel does not support it 
>>yet.
>>    
>>
>
>Or, he could simply not try to reuse the private buffer he is
>giving to TCP.
>  
>
you are describing a memory leak. at some point he must free (or 
otherwise reuse) these pages.

theoretically he could peek at the tcp sequence number, but an 
event-driven, protocol-agnostic completion seems better to me.

* light goes on *

yes, if he frees the pages immediately after tcp_sendpage, then the 
reference count would remain elevated until tcp completes sending these 
pages. so the sequence

  allocate pages
  fill with data
  tcp_sendpage()
  free pages

should be safe?

(I am still wishing for tcp aio, though)

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.