Re: [PATCH 0/6] use memcpy_mcsafe() for copy_to_iter()

[Andy Lutomirski]

On Tue, May 1, 2018 at 4:02 PM Dan Williams <dan.j.williams@xxxxxxxxx>
wrote:

> On Tue, May 1, 2018 at 2:05 PM, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> > On Tue, May 1, 2018 at 1:55 PM Dan Williams <dan.j.williams@xxxxxxxxx>
> > wrote:
> >
> >> The result of the bypass is that the kernel treats machine checks
during
> >> read as system fatal (reboot) when they could simply be flagged as an
> >> I/O error, similar to performing reads through the pmem driver. Prevent
> >> this fatal condition by deploying memcpy_mcsafe() in the fsdax read
> >> path.
> >
> > How about just changing the rules, and go the old "Don't do that then"
way?
> >
> > IOW, get rid of the whole idea that MCS errors should be fatal. It's
wrong
> > and pointless anyway.
> >
> > The while approach seems fundamentally buggered, if you ever want to
mmap
> > one of these things. And don't you want that?
> >
> > So why continue down a fundamentally broken path?

> I'm confused. Are you talking about getting rid of the block-layer
> bypass or changing how MCS errors are handled? If it's the former I've
> gotten push back in the past trying to remove the bypass, but I feel
> better about my chances to slay that beast wielding the +5 Hammer of
> Linus. If it's the latter, MCS error handling, I don't see how get
> around something like copy_to_iter_mcsafe().

> You mention mmap. Yes, we want the predominant access model to be
> dax-mmap for Persistent Memory, but there's still the question about
> what to do with media errors. To date we are trying to mirror the
> error handling model for System Memory, i.e. SIGBUS to the process
> that consumed the error. Is that error handling model also problematic
> in your view?

I'm not sure exactly what you mean here, but my understanding of the status
quo is that memory errors in user code are non-fatal but that memory errors
in kernel code are fatal unless there's an appropriate extable entry.  The
old iov_iter code assumes that memcpy() on kernel addresses can't fail.
I'm not sure how else it could work.