[Top] [All Lists]

Re: file data not flushed to disk by umount?

To: James Chapman <jchapman@xxxxxxxxxxx>
Subject: Re: file data not flushed to disk by umount?
From: Nathan Scott <nathans@xxxxxxx>
Date: Fri, 1 Apr 2005 10:25:29 +1000
Cc: linux-xfs@xxxxxxxxxxx
In-reply-to: <424C41E5.2070601@xxxxxxxxxxx>; from jchapman@xxxxxxxxxxx on Thu, Mar 31, 2005 at 07:31:01PM +0100
References: <4241471D.3090103@xxxxxxxxxxx> <20050324004850.GC704@frodo> <424288A1.7000406@xxxxxxxxxxx> <20050329031253.GA10203@frodo> <42496A94.5080708@xxxxxxxxxxx> <20050329225718.GC790@frodo> <424AF364.3030803@xxxxxxxxxxx> <20050331090123.D3536736@xxxxxxxxxxxxxxxxxxxxxxxx> <424C41E5.2070601@xxxxxxxxxxx>
Sender: linux-xfs-bounce@xxxxxxxxxxx
User-agent: Mutt/1.2.5i
Hi James,

On Thu, Mar 31, 2005 at 07:31:01PM +0100, James Chapman wrote:
> Nathan Scott wrote:
> > On Wed, Mar 30, 2005 at 07:43:48PM +0100, James Chapman wrote:
> > 
> >>Do you know of any differences with which ext2/ext3 and xfs use the VFS 
> >>interfaces? I've been using a h/w debugger all day on this and I see 
> > 
> > 
> > Yes, the main one is XFS's use of delayed allocation (i.e. a buffer_head
> > with the BH_delay flag set on it -- ext2/3 don't use that, XFS does).
> Thanks for the info.

No problem.

> >>writepage() write data as expected, but readpage() _never_ reads the 
> >>disk for file data. block_read_full_page() never gets a mapped buffer 
> >>(nr==0).
> > 
> > If its cached (BH_uptodate) it doesn't need to be read from disk,
> > thats probably the behaviour your seeing.
> But I umount the disk, remount and then do the read in this test.

Try using xfs_db to look at the inode after unmounting, perhaps
the inode extent map wasn't flushed or something like that?  In
that case, there'd be no read from disk too.

> btw, I finally got hold of an old x86 box today. I installed RH9 and 
> rebuilt the exact vendor 2.4.25 kernel (with config changes needed for 
> x86). I also installed the latest xfsprogs. Sure enough, XFS just works 
> so we now know the problem is platform specific.
> I'm getting the platform vendor involved now. If I find that any xfs 
> changes are needed to support my arch (unlikely) then I'll let you know.

Ah, OK.  CC'ing linux-xfs, for the person who sent me private mail
saying he was worried about this.  (you know who you are :)

> One more question. Is there an open source disk stress test that you 
> know of? I have LTP but it is too bloated to run on an embedded system.

You can use the tools from xfs-cmds/xfstests -- some of those are in
LTP too, some aren't.



<Prev in Thread] Current Thread [Next in Thread>