Re: [RFC PATCH 27/28] sunrpc: Use sendmsg(MSG_SPLICE_PAGES) rather then sendpage

[Trond Myklebust]

> On Mar 16, 2023, at 14:06, David Howells <dhowells@xxxxxxxxxx> wrote:
> 
> Trond Myklebust <trondmy@xxxxxxxxxxxxxxx> wrote:
> 
>> 1) This is code that is common to the client and the server. Why are we
>> adding unused 3 bvec slots to every client RPC call?
> 
> Fair point, but I'm trying to avoid making four+ sendmsg calls in nfsd rather
> than one.

Add an enum iter_type for ITER_ITER ? :-)

Otherwise, please just split these functions into one for knfsd and a separate one for the client.

> 
>> 2) It obfuscates the existence of these bvec slots.
> 
> True, it'd be nice to find a better way to do it.  Question is, can the client
> make use of MSG_SPLICE_PAGES also?

The requirement for O_DIRECT support means we get the stable write issues with added extra spicy sauce.

> 
>> 3) knfsd may use splice_direct_to_actor() in order to avoid copying the page
>> cache data into private buffers (it just takes a reference to the
>> pages). Using MSG_SPLICE_PAGES will presumably require it to protect those
>> pages against further writes while the socket is referencing them.
> 
> Upstream sunrpc is using sendpage with TCP.  It already has that issue.
> MSG_SPLICE_PAGES is a way of doing sendpage through sendmsg.

Fair enough. I do seem to remember a schism with the knfsd developers over that issue.

_________________________________
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx