On Tue, 2012-10-16 at 09:37 +0800, Chen Gang wrote:
> 于 2012年10月15日 20:32, Myklebust, Trond 写道:
> > RPC is not ordered. The fact that we get one RPC reply before another
> > does not mean that the server sent them in that order.
> >
> > This is doubly true when you use UDP as the transport protocol.
>
> 1) is it means: nfs_inode_attrs_need_update need not consider async
> read_done situation ?
I don't understand what you mean. This is mainly about the asynchronous
write situation...
> 2) for correctness, I do not think "nfs_size_to_loff_t(fattr->size) >
> i_size_read(inode)" in nfs_size_need_update is enough. (at least need
> use "!=" instead of '>'), do you think so ?
No... If I did, I would have changed this 15 years ago when I was
writing that code. Nothing here is new... 2.6.27-rc9 has the exact same
heuristics.
It boils down to the rule that if you want to ensure that data is not
_lost_, then you have to ensure that the cached file size is not less
than the true file size.
> 3) another reference:
>
> A) for an old kernel version (such as 2.6.27-rc9), no such issue
> (because it did not have nfs_size_need_update).
>
> B) the test tools which I use is from the LTP (Linux Test Project),
> they use both udp and tcp to test both the nfsv2 and nfsv3.
So what combinations are failing?
> C) truly LTP has its limitations: "for stress test, LTP let nfs client
> and server under the same machine, which will cause kernel stable
> issue", but for net test, LTP use different machine (I got our issue
> from LTP net test).
Running the client and server on the same machine is likely to deadlock
due to memory pressure issues. The client needs to be able to _increase_
memory pressure on the server in order to reduce its own pressure. That
doesn't work well when client == server.
--
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
