[ltt-dev] [Patch] ltt-relay-alloc mmap support (due to NFS lack of splice support)

[Trond Myklebust]

On Mon, 2009-03-23 at 13:58 -0400, Mathieu Desnoyers wrote:
> * Masahiro Tamori (masahiro.tamori at gmail.com) wrote:
> > 2009/3/21 Mathieu Desnoyers <mathieu.desnoyers at polymtl.ca>:
> > > * Masahiro Tamori (masahiro.tamori at gmail.com) wrote:
> [...]
> > >> Our customer will use NFS to store trace data not storage device for
> > >> embedded devices.
> > >> Since newer lttng use splice() even if NFS can not support splice(),
> > >> I create a patch to support to mmap of ltt-relay-alloc.
> > >>
> [...]
> > >> ltt-relay-alloc mmap support
> > >>
> > >> Splice syscall does not support NFS. We can not save trace data to
> > >> NFS directory. If this feature is enabled, you can use mmap()
> > >> instead of splice().
> > >>
> > >
> > >
> > > Hi Masahiro,
> > >
> > > Maybe we should consider implementing splice() support in NFS instead ?
> > >
> > > I removed the mmap support from the ltt-relay-alloc files because splice
> > > is more efficient and does not require to vmap the pages.  Unless there
> > > is a strong argument telling what in NFS makes it impossible to
> > > implement splice(), I don't really see the gain in putting back the
> > > old mmap() mechanism we had.
> > >
> > > Maybe the NFS people will have some information about this ?
> > >
> > > Thanks,
> > >
> > > Mathieu
> > >
> > 
> > Hello Mathieu,
> > 
> > I think that the best solution is NFS can support splice() even if
> > it cannot zero copy.
> > 
> > The splice() will be used by any other tools, hence NFS should support
> > it sooner or later.
> > 
> > If technical issue is remained to support splice() in NFS,
> > we should support mmap in LTTng until the problem is resolved.
> > Embedded people will want to use LTTng with NFS,
> > though this is a bad choice.
> > 
> > Thank you,
> > Masahiro Tamori
> > 
> 
> There was a LKML thread on NFS splice support back in 2006 :
> 
> http://lkml.indiana.edu/hypermail/linux/kernel/0603.3/2102.html
> 
> I don't know what happened with this ? I have tested NFS v2 and v3 and
> have seen they do not support splice, but haven't tested NFSv4.
> 
> Mathieu

We do support splice reads and apparently don't support splice writes.
However I don't see what has been stopping anybody from implementing the
latter.

In fact, it looks to me as if we should just be able to use
generic_file_splice_write() as is. There is no O_APPEND support or
anything that would require us to revalidate file lengths; it's just a
perfectly ordinary write into the page cache...

Cheers
   Trond