Re: [patch 1/2]MD: raid5 trim support

[Shaohua Li <shli@kernel.org>]

On Thu, Sep 20, 2012 at 01:59:14PM +1000, NeilBrown wrote:
> On Thu, 20 Sep 2012 10:27:17 +0800 Shaohua Li <shli@kernel.org> wrote:
> 
> in which wrong sectors were trimmed....
> > 
> > Ok, just confirmed, delete raid5_compute_sector is ok if I adjust
> > logical_sector calculation. Here is the new patch.
> > 
> 
> Thanks.  That looks better.  I've applied it with some minor formatting
> changes.
> 
> I then went to look at the follow-up page and .....
> I count 11 separate places where you test the new flag and possibly memset a
> page to zero.  This doesn't seem like an improvement to me.

We do the zero page just before the stripe is hit in cache, which is rare case.
 
> Why don't we just mark the page as not up-to-date when we discard it?  That
> would avoid storing inconsistent data, and would avoid needing to zero pages.

We need re-read the strip if it's hit in cache, but it's rare case, we don't
care. So when we clear the up-to-date flag? I saw a lot of places checking
up-to-date flag in the write path. Need close look to check if there is race.

Thanks,
Shaohua