On Wed, Mar 23, 2011 at 12:01:17AM -0700, Andi Kleen wrote:
> Dave Chinner <david@xxxxxxxxxxxxx> writes:
> > Of course, the major patch in this series is the conversion of the
> > buffer cache to using kmalloc and get_free_page() directly rather
> > than using the page cache. This has many benefits and I haven't
> > found any regressions due to making that change yet.
> One (somewhat obscure) regression will be that you won't be able to
> recover from uncorrected memory errors in the buffer cache anymore.
We can't do that right now, anyway.
> Previously memory_failure() could just drop it transparently when that
> happens and the page is currently not used.
If the page is not in use, we don't care about it after this patch
set is applied - the page is either active in a buffer or it has been
freed. If it is in use, then we'll shut the filesystem down if we
detect the memory corruption just like we currently do. Hence I
don't see any regression here.
As it is, there is no way for the filesytem to be notified about
such failures on active pages in buffers, so in reality we can't
reliably detect them so there is little point in trying to recover
from such errors.