On Mon, 2008-09-22 at 10:04 -0700, Aaron Straus wrote:
> Here is the crux. It was possible previously but unlikely e.g. our app
> never saw this behavior. The new writeout semantics causes visible
> holes in files often.
>
> Anyway, I agree the new writeout semantics are allowed and possibly
> saner than the previous writeout path. The problem is that it is
> __annoying__ for this use case (log files).
There is always the option of using syslog.
> I'm not sure if there is an easy solution. We want the VM to writeout
> the address space in order. Maybe we can start the scan for dirty
> pages at the last page we wrote out i.e. page 0 in the example above?
You can never guarantee that in a multi-threaded environment.
Two threads may, for instance, force 2 competing fsync() calls: that
again may cause out-of-order writes.
...and even if the client doesn't reorder the writes, the _server_ may
do it, since multiple nfsd threads may race when processing writes to
the same file.
Anyway, the patch to force a single threaded nfs client to write out the
data in order is trivial. See attachment...
Trond
Hi,
On Sep 22 01:29 PM, Trond Myklebust wrote:
> > Anyway, I agree the new writeout semantics are allowed and possibly
> > saner than the previous writeout path. The problem is that it is
> > __annoying__ for this use case (log files).
>
> There is always the option of using syslog.
Definitely. Everything in our control we can work around.... there are
a few applications we cannot easily change... see the follow-up in
another e-mail.
> > I'm not sure if there is an easy solution. We want the VM to writeout
> > the address space in order. Maybe we can start the scan for dirty
> > pages at the last page we wrote out i.e. page 0 in the example above?
>
> You can never guarantee that in a multi-threaded environment.
Definitely. This is a single writer, single reader case though.
> Two threads may, for instance, force 2 competing fsync() calls: that
> again may cause out-of-order writes.
Yup.
> ...and even if the client doesn't reorder the writes, the _server_ may
> do it, since multiple nfsd threads may race when processing writes to
> the same file.
Yup. We're definitely not asking for anything like that.
> Anyway, the patch to force a single threaded nfs client to write out the
> data in order is trivial. See attachment...
>
> diff --git a/fs/nfs/write.c b/fs/nfs/write.c
> index 3229e21..eb6b211 100644
> --- a/fs/nfs/write.c
> +++ b/fs/nfs/write.c
> @@ -1428,7 +1428,8 @@ static int nfs_write_mapping(struct address_space *mapping, int how)
> .sync_mode = WB_SYNC_NONE,
> .nr_to_write = LONG_MAX,
> .for_writepages = 1,
> - .range_cyclic = 1,
> + .range_start = 0,
> + .range_end = LLONG_MAX,
> };
> int ret;
>
Yeah I was looking at that while debugging. Would that change have
chance to make it into mainline? I assume it makes the normal writeout
path more expensive, by forcing a scan of the entire address space.
Also, I should test this, but I thought the VM was calling
nfs_writepages directly i.e. not going through nfs_write_mapping. Let
me test with this patch.
Thanks,
=a=
--
===================
Aaron Straus
aaron-bYFJunmd+ZV8UrSeD/[email protected]