Re: [PATCH 2/6] net: pad skb data and shinfo as a whole rather than individually

[David Miller <davem@davemloft.net>]

From: Ian Campbell <Ian.Campbell@citrix.com>
Date: Thu, 5 Jan 2012 20:00:58 +0000

> Although this scheme works fine for Xen's netback I don't think it works
> for other use cases like the NFS one (or basically any kernel_sendpage
> usage). In those cases you don't want to wait for the last core mm ref
> on the page to get dropped, you just want to wait for the last ref due
> to the particular sendpage. If you use the core page_count() reference
> count then you end up waiting for the process to exit (and drop the core
> reference count) before the destructor fires and you can complete the
> write, which is obviously not what is desired!
> 
> There's also issues with things like two threads simultaneously doing
> I/O from the same page. If one lot of I/O is to NFS and the other to
> iSCSI (assuming they both use this facility in the future) then they
> will clash over the use of the struct page field. In fact even if they
> were both to NFS I bet nothing good would happen...

Ok, struct page doesn't seem the best place for this stuff then.

Thanks.