Re: [PATCH 4/5] HWPOISON: report sticky EIO for poisoned file

[Wu Fengguang <fengguang.wu@intel.com>]

On Fri, Jun 12, 2009 at 06:07:16PM +0800, Andi Kleen wrote:
> On Thu, Jun 11, 2009 at 10:22:43PM +0800, Wu Fengguang wrote:
> > This makes the EIO reports on write(), fsync(), or the NFS close()
> > sticky enough. The only way to get rid of it may be
> > 
> > 	echo 3 > /proc/sys/vm/drop_caches
> > 
> > Note that the impacted process will only be killed if it mapped the page.
> > XXX
> > via read()/write()/fsync() instead of memory mapped reads/writes, simply
> > because it's very hard to find them.
> 
> I don't like the special case bit. Conceptually we shouldn't need
> to handle hwpoison specially here; it's just like a standard error. 
> 
> It makes hwpoison look more intrusive than it really is :)

This is already far more less intrusive than Nick's SIGKILL idea ;)

> I think it would be better to simply make
> the standard EIO sticky; that would fix a lot of other issues too (e.g.
> better reporting of metadata errors) But that's something for post .31.

Sure, fixing standard EIO is not the task of this patchset.

> For .31 I think hwpoison can live fine with non sticky errors; it was
> more a problem of the test suite anyways which we worked around.
> 
> So better drop this patch for now.

OK, if people in this list agree it to be intrusive ;)

Thanks,
Fengguang