On Mon, 2012-09-10 at 19:52 +0000, Adamson, Andy wrote:
> On Sep 10, 2012, at 2:57 PM, Jeff Layton wrote:
> > My only concern is this behavior described in the last patch:
> >
> > "If the buffered WRITE is using a credential key that will expire
> > within low watermark seconds, fail the WRITE in nfs_write_begin
> > _before_ the WRITE is buffered and return -EACCES to the application."
> >
> > Shouldn't rejecting the write attempt be the purview of the server?
>
> The TGT lifetime is shared by the client and the server - it's the same, so I don't think we'd gain anything by doing that.
>
> >
> > It seems to me that we'd be best off just switching to synchronous
> > writes when we start approaching credential expiration, and letting the
> > server handle the case where the credential expires.
>
> You mean switching to direct IO? I believe that means taking a different path in the VFS layer which I found difficult to do from nfs_write_begin. But maybe I just didn't see a good way to do this.
Synchronous I/O is a completely different concept than direct I/O.
For ordinary writes, making them synchronous basically involves calling
vfs_fsync() before returning from the write() syscall (see
nfs_file_write()).
For direct I/O, making it synchronous means ensuring that we do wait in
nfs_direct_wait().
For mmap(), things are a bit more tricky since the dirtying process
happens via a callback from the page fault. I'm not sure how you could
force that to be synchronous.
--
Trond Myklebust
Linux NFS client maintainer
NetApp
Trond.Myklebust@xxxxxxxxxx
www.netapp.com
��.n��������+%������w��{.n�����{��w��� jg��� �����ݢj����G���� ���j:+v���w�m������w�������h�����٥
