[dm-devel] [PATCH 5/5] dax: always use _copy_mc_to_iter in dax_copy_to_iter

[Dan Williams]

On Mon, Dec 13, 2021 at 12:20 AM Christoph Hellwig <hch at lst.de> wrote:
>
> On Sun, Dec 12, 2021 at 06:48:05AM -0800, Dan Williams wrote:
> > On Fri, Dec 10, 2021 at 6:05 AM Vivek Goyal <vgoyal at redhat.com> wrote:
> > >
> > > On Thu, Dec 09, 2021 at 07:38:28AM +0100, Christoph Hellwig wrote:
> > > > While using the MC-safe copy routines is rather pointless on a virtual device
> > > > like virtiofs,
> > >
> > > I was wondering about that. Is it completely pointless.
> > >
> > > Typically we are just mapping host page cache into qemu address space.
> > > That shows as virtiofs device pfn in guest and that pfn is mapped into
> > > guest application address space in mmap() call.
> > >
> > > Given on host its DRAM, so I would not expect machine check on load side
> > > so there was no need to use machine check safe variant.
> >
> > That's a broken assumption, DRAM experiences multi-bit ECC errors.
> > Machine checks, data aborts, etc existed before PMEM.
>
> So the conclusion here is that we should always use the mc safe variant?

The proposal is one of the following:

1/ any paths not currently using the mc safe variant should continue
not using it to avoid the performance regression on older platforms,
i.e. drop this patch.

2/ add plumbing to switch to mcsafe variant, but only on newer
platforms, incremental new patch

3/ always use the mc safe variant, keep this patch

We could go with 3/ and see who screams, because 3/ is smallest
ongoing maintenance burden. However, I feel like 1/ is the path of
least resistance until the platforms with the need to do 'careful'
copying age out of use.